Having finished Stefano Franchi’s paper on Ashby’s homeostat and its commentaries ([Notes from sessions 4 and 5 link]), the Cognition Futures Reading Group turned on Thursday to Jonathan Richens and Tom Everitt’s “Robust agents learn causal world models”, presented at ICLR 2024. Friday was the group’s first session given over to planning its technical work: what the Friday sessions are for, which tools to use, and a first shared exercise.

What was read

Richens and Everitt ask whether agents must learn causal models in order to generalize to new domains, or whether other inductive biases are enough. Their answer is a necessity result. An agent that keeps its regret within a bound across a large set of distributional shifts must have learned an approximate causal model of the process that generates its data, and the approximation becomes exact as the agent approaches optimal. The shifts are modeled as local interventions on the environment’s causal variables. That causal models are sufficient for robust adaptation was already known; the paper argues they are also necessary.

In the discussion section, the authors revisit the good regulator theorem. It is often read as saying that any good controller must contain a model of the system it controls. The authors point out that the theorem itself only guarantees an optimal regulator that is a deterministic function of the system’s state, which could be trivial. They present their own result as a sharper, causal version of the theorem. Figure 4 places the result in Pearl’s causal hierarchy: optimal policies under domain shifts, for a single objective, already cover the interventional level, so they are also enough for shifts in the task.

The paper is explicit about its limits. The main theorems require robustness to shifts on all of the environment’s variables, and they cover only decision tasks in which the agent’s actions do not change the environment. Its discussion also gives AlphaFold as an example of competence without understanding: accurate prediction of protein folding that did not improve understanding of the underlying biochemistry.

The group read the abstract, the good regulator paragraph, Figure 4 and the conclusion together. Several members found the formal sections hard going, and the group did not take as deep a pass as it did with Franchi.

What was discussed

Modeling concepts rather than life

Amanda Nelson opened Thursday with what she took from the end of Franchi’s paper for the group’s own building work. Modeling life is a grand project, she said, and it would first require deciding what constitutes life. Framing a build as the modeling of a concept, explored openly and with no claim to reproduce a living thing, seemed to her more fruitful.

Can a priori structure do technical work?

Avery Lim described a conversation with a computer scientist who also studies philosophy. Lim had suggested that his generalized domain primitives might help structure reasoning in language models. The objection was that structures reasoned out in advance can be useful to people as an interface, but cannot capture technical structure, which has to be discovered by statistical exploration. Lim called this the status quo view in computer science. His interlocutor traced it through Wittgenstein and Turing, whom he counts among the roots of cybernetics.

Jes Parent identified the objection as the bitter lesson in other terms. Nelson asked whether building a model would itself count as enough use of a concept to ground it, in the sense of Wittgenstein’s language games. Lim said Wittgenstein can be read in more than one way, and proposed tracing the line from Wittgenstein through Turing and logical positivism to see how the divide between concept modeling and object modeling formed, and how it still shapes what researchers treat as viable.

Prediction, intervention, and the good regulator

Lim drew two threads from the reading. First, he read the paper as separating prediction from robust choice: a good predictor does not need a causal model, but choosing well across a comprehensive range of counterfactual alternatives does. He wondered whether this marks where a priori work earns its place, in organizing a space of causal possibilities and interventions rather than in making specific predictions, which answers part of the objection above. Second, he wanted to look harder at why a causal model is what separates this result from the original framing of the good regulator theorem.

Lim noted the paper’s requirement of a comprehensive family of interventions and its AlphaFold example. He also pointed to work, cited in the paper’s related work, suggesting that trained models can hold internal world models without being given that structure explicitly. He asked what it means for domain and task shifts to overlap with the observational level in Figure 4. Parent read the conclusion aloud: robust decision-making requires a causal model of the data-generating process, which the authors take as evidence that causal world models are a necessary ingredient of general intelligence.

Representation and isomorphism

Parent mentioned the Brain Inspired group’s discussion the day before. There, a good part of the conversation turned on how isomorphism is used: whether a given case is an isomorphism, a homomorphism or a functional isomorphism. Janet contributed comments from an engineering perspective, and Kevin Mitchell also took part. Lim added that the discussion raised whether representation is needed for, or implied by, results like this one, a recurring question in embodied cognition and cybernetics. This was raised briefly and not developed.

Parent also mentioned the AHA bonus seminar with Cathy O’Neil, which was taking place as the session ended, as something to come back to.

What should Fridays be for?

Near the end of Thursday, Nelson asked whether the group wants to keep moving between papers or to focus on a particular modeling project. She learns by building, she said, and suggested starting from models that have been discussed for decades, such as the homeostat or one of the most basic models of the neuron. Parent proposed starting with something very simple, then tying it to earlier Orthogonal Research and Education Lab work on developmental Braitenberg vehicles, and then moving into new territory. He asked everyone to spend some time before Friday thinking about a starter task. A perceptron was suggested as a default first target; on Friday, Lim credited the suggestion to Bradly Alicea.

For the next Thursday reading, Lim offered two papers on Wittgenstein’s place in the history of AI. Nelson found their AI framing less appealing and preferred Lydia Patton’s paper on Helmholtz, Wittgenstein and the a priori, which the group chose.

Goals for the technical sessions

On Friday, Morgan Hough asked what the technical sessions are for: understanding the code yourself, or being able to run simulations. Lim said he wants tools flexible enough for unconventional designs, computationally viable first and user friendly second. Longer term, he wants to model non-trivial adaptation to threat and need, social and environmental, starting from something simple that can be extended. Nelson said she wants to build from simple rules and be able to point to parts of the code and name the concepts they enact, such as a perceptron, a homeostat or feedback. She also wants the behavior being modeled to be describable. Parent described the two as a small, concept-level program and a longer-range one, with the same starting point.

Tools and scale

Lim had installed NetLogo and suggested following a tutorial together. Nelson leaned toward Python for its wider range of uses. Hough noted that for artificial-life experiments, how far a simulation can run in time is a matter of compute. He pointed to JAX-based packages that let Python code run in parallel on GPUs, and mentioned NEAT, EvoGym and JaxMARL. He described evolutionary robotics, in which an agent’s morphology and connectivity both change along with its policy, as a large, high-dimensional search space.

Related work and the problem of non-trivial models

Hough described a computational-life paper his reading group had covered, on the coevolution of sociality and replication. He raised work on modularity in neuroevolution from the mid-2010s, by Jeff Clune and Jean-Baptiste Mouret among others, as relevant to developmental Braitenberg vehicles. He said the work he finds most interesting is Logan Cross’s on theory of mind in agents: what models agents build of other agents. His own aim is multi-agent models with sociality and theory of mind that are relevant to computational psychiatry, and he noted that such agents also need tasks and organizing principles.

Lim said that when he has tried toy models, it was hard to get non-trivial behavior: things meant what they did only because he had set them up to, rather than through emergent interaction. Hough called this the balance problem in sociophysics, the difficulty of making statistical-mechanics models interpretable in social terms, and pointed to a 2025 review of the field.

A first exercise

The group settled on a show and tell for the next Friday. Each person brings a small model they have found or built: a homeostat, a perceptron, a thermostat, a Braitenberg vehicle or something similar. AI-assisted code is fine. Parent asked that the emphasis fall on describing the model: what it does, and whether it achieves what it is supposed to. Lim suggested a shared code repository with a space for each person and a shared notes document, and Parent set up a repository on the JOPRO GitHub for the purpose.

Key terms

• Regret bound: a limit on how much worse an agent’s policy does than the best available policy.

• Distributional shift: in Richens and Everitt, a change to the environment modeled as a local intervention on one or more of its causal variables.

• Good regulator theorem: the result, often read as saying that every good regulator of a system must be a model of that system; Richens and Everitt argue its literal content is weaker.

• Concept simulation: in Franchi, building a model to explore the consequences of an idea, as distinct from simulating an object.

• Developmental Braitenberg vehicles: Braitenberg’s simple embodied agents, extended so that their nervous systems develop rather than arrive fully formed.

Open questions carried forward

• Can structure reasoned out in advance do technical work, or only help people interpret? (Lim, Nelson)

• Does the paper’s result require representation, or only imply it? (Lim)

• What does it mean for domain and task shifts to sit where they do in Figure 4? (Lim)

• How do you get non-trivial behavior from a toy model without building the meaning in? (Lim, Hough)

Next time

On Thursday the group reads Lydia Patton’s “Signs, toy models, and the a priori”; on Friday, the show and tell.

References

This session

• Richens, J. & Everitt, T. (2024). Robust agents learn causal world models. International Conference on Learning Representations (ICLR 2024).

• Patton, L. (2009). Signs, toy models, and the a priori: from Helmholtz to Wittgenstein. Studies in History and Philosophy of Science 40: 281–289. doi:10.1016/j.shpsa.2009.07.004

Background

• Franchi, S. (2013). Homeostats for the 21st century? Simulating Ashby simulating the brain. Constructivist Foundations 9(1).

• Dvoretskii, S., Gong, Z., Gupta, A., Parent, J. & Alicea, B. (2020). Braitenberg Vehicles as Developmental Neurosimulation. arXiv:2003.07689.

Further & Suggested Reading

These readings were not discussed on November 5 or 6. They were gathered after the sessions as starting points for the questions Nelson and Lim raised. The annotations are editorial.

Where should a simple first model start?

• Stefan Dvoretskii, Ziyi Gong, Ankit Gupta, Jesse Parent and Bradly Alicea (2020), “Braitenberg Vehicles as Developmental Neurosimulation”. Argues that models which begin as fully formed networks ignore biological development, and proposes Braitenberg vehicles as a minimal embodied system in which a nervous system can develop. It presents three software approaches: a vehicle and genetic algorithm hybrid, multisensory Hebbian learning, and multi-agent models. This is the developmental Braitenberg vehicle work Parent proposed as the bridge from a first simple build.

• Bradly Alicea, Rishabh Chakrabarty, Stefan Dvoretskii, Ziyi Gong, Akshara Gopi, Avery Lim and Jesse Parent, “Continual Developmental Neurosimulation Using Embodied Computational Agents”. Extends developmental Braitenberg vehicles into a design for learning agents whose development has distinct morphogenetic, critical and learning periods. It proposes growing the network with a genetic algorithm and treats development as a form of continual learning. Relevant to Lim’s aim of a simple model that can be extended toward non-trivial adaptation.

October was another full month at JOPRO. This edition looks back at what took place and points ahead to what is coming in November and through to early 2027.

New Fall Programming: Thursday - Friday Sessions

This fall, a core set of JOPRO’s weekly sessions moved to a Thursday & Friday format, alongside existing standing meetings. Thursday is a reading session, where members work through source materials; Friday serves as technical session for building and examining small models. These sessions bring together the Cognition Futures Reading Group and the cybernetics interest group, and a cluster of interests around multi-agent interaction.

A line of inquiry in October was Stefano Franchi’s paper on W. Ross Ashby’s homeostat, “Homeostats for the 21st Century? Simulating Ashby Simulating the Brain,” read and discussed across several sessions.

Developing Research Agendas

This month, Avery Lim and Jesse Parent published “The Questions We Stopped Asking”, the first in a series on mediated gestalt inference. The post argues that holistic inquiry is undervalued in how research is funded and ranked, and sets out the groundwork for a longer program of work, including a companion in-preparation position and framework manuscripts.

“The investigation of the structure of insight and integratory processes within computational, formalist, and discursive systems holds immense potential. I’m glad to have brought together so many of my thoughts on this, and I hope this work can serve as a rallying call over the long term for others interested in this frontier.” -

Research Outputs and Presentations

• SocialAgent@NeurIPS. Sarrah Bastawala’s research on LLAMOSC, “Governance as a Safety Net: An LLM Agent-Based Study of Open-Source Community Sustainability” with Himanshu Chougule, Jes Parent, and , was accepted as a poster at SocialAgent: Second Workshop on Large Language Models for Social Reasoning and Simulation, during NeurIPS 2026 in Atlanta.

• From homeostat to primitives. In October, Avery Lim and Jesse Parent presented “From Homeostat to Universal Domain Primitives” at the Orthogonal Research and Education Lab. The talk set out how their current work continues Ashby’s and Franchi’s, and its possible connections to Allostasis Machines, Contextual Geometric Structures, and earlier work, with directions for what comes next.

• Substrate-Agnostic & Non-uniformity research program. October also saw substantial work on the research program behind the mediated gestalt inference series: how the sciences of mind and agency choose their abstractions, and whether they say what they leave out.

• We Robot 2027. JOPRO members, including several from the Data x Direction program, submitted to the call for We Robot 2027, held at Oxford University’s Institute for Ethics in AI. The JOPRO community has been involved with We Robot since 2022 in Seattle, and we hope to continue this involvement with one of our favorite conference series.

People

• Tlamelo Makati joins as a Visiting Scholar. Makati’s work centers on digital accessibility and active inclusion: making sure people with lived experience of disability have a say at every stage of how machine learning tools are built. This semester she is an assistant lecturer in mechatronics engineering at Technological University Dublin, and next year she begins a postdoc with the Data Science for Social Impact Lab at the University of Pretoria. Makati will be featured in an upcoming In Conversation series, where she will discuss her research origins and future plans for developing her active inclusion-based frameworks.

• S. Hussain Ather was accepted to the Kris Rockwell Foundation’s “Cybernetic Futures: Intelligence, Consciousness, and the Path of AI” call.

Out in the network

JOPRO members were also active at events around Boston and beyond.

• October 6. JOPRO members attended Cody Turner’s (Bentley University) talk “The Moral and Epistemic Harms of AI Sycophancy” in Victor Kumal’s AI Ethics Lab Talk Series at Boston University.

• October 8. JOPRO members attended the MIT Media Lab launch of Sherry Turkle’s Artificial Intimacy, hosted by Advancing Humans with AI and co-hosted by the Octet Collaborative and MIT’s Social and Ethical Responsibilities of Computing program.

• November 7. Jennifer Jiang will return to the 15th Asian American Women in Leadership Conference, hosted by ASPIRE in Boston. We hope to share a short piece or interview about the experience.

From our partners at the Orthogonal Research and Education Lab: Bradly Alicea’s recent conference roundup covers OREL’s work at Artificial Life 2026 and BICA*AI 2026, including a talk on ensemble models of early life with co-authors including S. Hussain Ather, and work with Morgan Hough on psychiatric BICAs.

Also, “Declared Abstraction: On Bracketing and Saying What You Leave Out”, by , takes a second look at Rosenblueth, Wiener, and Bigelow’s 1943 paper “Behavior, Purpose and Teleology,” and asks what a method sets aside, whether it says so, and what that costs.

These and other October happenings bear on questions JOPRO members have been working on about how people relate to AI systems, explored in “Open strands on Human-AI relationships.”

Ahead

• November: 6th Applied Active Inference Symposium. Co-organizer Bradly Alicea continues his work with the Active Inference Institute’s 6th Applied Active Inference Symposium, leading an interactive workshop, an in-depth discussion on “World models in natural and artificial intelligence.” This builds of of work featured in the Royal Society A Phil. Transactions Special Issue volume 384: World models in natural and artificial intelligence. It features Alicea, Nelson, Hough, and Parent’s contribution of A ‘good’ regulator may provide a world model for intelligent systems.

• November 14: Algorithmic Justice League - Decade of Impact. JOPRO members will attend the Algorithmic Justice League’s celebration of ten years of work against algorithmic harm, at the Museum of Science in Boston. The program includes emcee Dr. Avriel Epps, a keynote by Dr. Alondra Nelson, and the Global AI Justice Award. For those joining remotely, the livestream runs 6:30 to 8:00 pm ET, followed by a hybrid discussion on the JOPRO Discord. It is part of JOPRO’s fall series of AI ethics socials; details on Luma. In-person entry requires a ticket from the Algorithmic Justice League.

• December: NeurIPS 2026. Sarrah Bastawala’s LLAMOSC work will be presented at the SocialAgent workshop on December 13 in Atlanta, part of the Social Agents Workshop Series.

• Spring 2027: JOPRO @ WMU. JOPRO is organizing and hosting an event at Western Michigan University, on the philosophy and history of science, multidisciplinary methodology, alongside recent work from the Cognition Futures and Center for Futures Studies. If you are interested in supporting, collaborating, or presenting your research, get in touch.

Stay tuned for more public events, including JOPRO Socials at NeurIPS, the Boston Museum of Science, and upcoming events in the greater Boston-Cambridge area. Subscribe here below, and follow the JOPRO calendar on Luma for announcements.

JOPRO supports interdisciplinary scholarship, cultivates emerging researchers, and builds the networks that connect them.

This post is a rough log for our asynchronous members to view at a glance, rather than an elaborate or high-fidelity capture of the discussion. It covers two sessions. Abridged recordings will be posted to the JOPRO YouTube channel.

The Cognition Futures Reading Group continued Stefano Franchi’s “Homeostats for the 21st Century? Simulating Ashby Simulating the Brain”, published in Constructivist Foundations in 2013. Notes from sessions 2 and 3 cover Franchi’s seven theses, the sleeping machine, and Figures 4 and 5. On Thursday the group read on through Franchi’s case for the homeostat as a concept and his account of heteronomy. On Friday Avery Lim presented work that connects to the reading.

What was read

On Thursday, Jes Parent recapped sessions 2 and 3, and the group returned to the footnote on Franchi’s technical and non-technical theses. Thesis 1, general homeostasis, is philosophical; thesis 2 is biological; theses 3 and 4 explain how the first two would be realized. Franchi argues that reading the homeostat as a concept simulation lets attention fall back on theses 1 and 2.

Franchi then reads Ashby’s design work, from polystable and multistable systems to the principles of dispersion and localization, as a conceptual exploration of life as passive and contingent. On his account, Ashby is asking what the internal organization of a passive, contingent system would have to be like, and the homeostat serves as a complex thought experiment for answering it.

From that standpoint, Franchi says the most important of Ashby’s theses is the first: a complex system running to equilibrium generates complex behavior. Everything else is an explication or a consequence of it. Critical discussion has nonetheless focused on the derived claims, above all ultrastability. Franchi notes that Norbert Wiener, while praising Ashby, reduced homeostasis to a physiological instance of feedback, and that recent work, such as Ezequiel Di Paolo’s, grafts ultrastability onto neural network controllers. Both, he argues, privilege the object side of Ashby’s simulation over its conceptual side.

Franchi suggests the reason is the controversial status of the general homeostasis thesis itself. It insists that life is essentially heteronomous, governed from outside rather than self-governed. That puts it at odds with a long philosophical tradition that separates human from non-human life precisely on the line between autonomous and heteronomous behavior. Footnotes 13 and 14 tie this to a passage from Ashby that Franchi quotes: “There is yet to be written a book that shall show how all the organism’s exteriorly directed activities are similarly regulatory.”

On Friday the group did not read further. The group will move to the open peer commentaries next.

What was discussed

Lim’s essay and Ashby’s open book

Lim introduced “The Questions We Stopped Asking”, co-written with Parent, the first post in a series on what Lim calls mediated gestalt inference. Lim described it as a historical overview of a holistic, whole-system kind of inquiry that has been devalued, and said the next post will take up “domain primitives.” Lim connected the essay to Ashby’s claim that even high-order human drives, such as playing chess, can be described as homeostatic, and to the “yet to be written” book quoted above. Parent compared the essay’s project to the way the 1943 paper by Rosenblueth, Wiener and Bigelow reclaimed the term teleology, and stressed that it is not an argument against empiricism or the scientific method.

Simulating objects and simulating concepts

Lim asked whether Franchi’s distinction means that applying Ashby naively, at the level of cells, is wrong, while applying him at higher levels of organization is more accurate. Parent read Franchi differently: the distinction is about not reducing the homeostat to a brain simulation, because Ashby was also working through philosophical questions. Amanda Nelson pointed to Franchi’s remark that technical work in the cognitive sciences is sometimes philosophy in disguise. She described concept simulation as exploratory, playing with abstract concepts rather than aiming only at accurate reproduction, and noted Franchi’s point that a good model does both. Lim extended this: simulating a concept may improve understanding of the object itself. Nelson thought that was consistent with Franchi, and said she wished he had gone further on what the ideal combination looks like.

The sleeping machine and the dark room problem

Nelson asked how closely the sleeping machine relates to predictive processing’s dark room problem: if an organism’s aim were to minimize prediction error, it could simply shut itself in a dark, predictable room. That looks anti-life, and she noted it is presumably an old objection with many replies. Lim asked whether anyone has related it to sensory hypersensitivity, as in autism or migraine, where a predictable environment is sought. Nelson said organisms sometimes do face something like the dark room, but that it is generally taken as a sign something has gone wrong. Parent connected it to differing degrees of sensory permeability and to how many variables a person can hold at once.

Quiescence or set point?

Lim asked why Ashby and Franchi frame quiescence as the goal, rather than maintaining a set point with rest as the end of a cycle. Parent restated the question as whether a system reaches a stable local equilibrium in a larger landscape rather than the whole system going still, and noted that the word “quiescence” appears only once in Franchi’s article. Parent proposed collecting the group’s questions before reading the commentaries, so they can be compared with what the commentators and Franchi’s response say.

Heteronomy and what counts as life

Parent read the heteronomy passage and asked Nelson what it meant to her. Nelson connected it to her earlier point that a drive toward sleep looks anti-life. For Maturana and Varela, autonomy, a system’s self-production and self-regulation, is what defines life, and many philosophers have described something like a life force. If life is heteronomous in Ashby’s sense, she asked, how do we distinguish life from non-life? Parent said he wants to look into why the tradition settled on autonomy as the dividing line, as a review of the literature.

Preference as homeostasis

Lim suggested the group was circling preference and set points within subsystems, and proposed that preference and autonomy may themselves be the maintenance of a set point, dependent on homeostatic regulation. It is hard to explain why someone plays chess, Lim said, if the drive is only to do the minimum and return to rest. Lim pointed to individual differences in drives, such as the Big Five, need for cognition, and need for affect, and to a Frontiers paper on why people enjoy horror films, and suggested homeostasis may describe complex behavior better when it concerns information rather than survival. Parent placed this under what he calls the non-uniformity of phenomenological experience. Lim proposed grounding the next discussion in a triad of appetite for diversity, for intensity, and for quiescence, and planned to go back to what Ashby himself wrote about quiescence. This was raised as a direction to develop and was not settled.

The quiescence trap

On Friday, Parent shared a new interoception paper by Micah Allen and colleagues that names a “quiescence trap”: measuring interoception at rest, typically through a single channel such as heartbeat detection, when the field’s interest is in how bodily signals shape emotional life. Nelson read the term as an ecological validity problem, and Morgan Hough noted the paper’s use of video of people at great heights to induce arousal. The word is used differently than in Franchi, but the overlap was noted.

Lim’s presentation

Lim presented on universal domain primitives and how they relate to Ashby, to Franchi’s reading, and to earlier work on allostatic kinds and allostasis machines.

Nelson asked whether the project is a kind of counter-Enlightenment argument, and whether a rationalist, modeling approach has limits for questions about life and consciousness. Lim answered that the argument is for integration rather than a rejection of Enlightenment methods, pointing to Hegel’s idea of sublation, and allowed that the essay’s title overstates the case. Bradly Alicea said the work partly matches the lab’s existing research, needs more connecting to it, and then needs testing.

Key terms

• Heteronomy: behavior governed by something outside the system; the counterpart of autonomy.

• General homeostasis thesis: Ashby’s first thesis, that a complex system running to equilibrium generates complex behavior.

• Concept simulation: in Franchi, building a model to explore the consequences of an idea, as distinct from simulating an object such as the brain.

• Dark room problem: the objection that an agent minimizing prediction error should seek a dark, unchanging room.

• Quiescence trap: Allen and colleagues’ term for studying brain-body interaction only at rest.

• Mediated gestalt inference: Lim and Parent’s name for holistic inquiry into connections that run through intermediaries.

Open questions carried forward

• Why frame the homeostat’s goal as quiescence rather than maintaining a set point? (Lim)

• If life is heteronomous in Ashby’s sense, what distinguishes living from non-living systems? (Nelson)

• Can preference itself be modeled as set-point maintenance? (Lim)

• What would the ideal combination of object and concept simulation look like? (Nelson)

• What are the limits of a modeling approach to questions about life and consciousness? (Nelson)

Next time

The group moves to the open peer commentaries and Franchi’s response.

Subscribe now

References

This session

• Franchi, S. (2013). Homeostats for the 21st century? Simulating Ashby simulating the brain. Constructivist Foundations 9(1): 93–101. Open peer commentaries and the author’s response follow in the same issue. http://constructivist.info/9/1/093

• Lim, A. & Parent, J. (2026). The Questions We Stopped Asking. JOPRO, October 19, 2026.

• Rosenblueth, A., Wiener, N. & Bigelow, J. (1943). Behavior, purpose and teleology. Philosophy of Science 10(1): 18–24.

• Maturana, H. R. & Varela, F. J. (1980). Autopoiesis and Cognition: The Realization of the Living. Reidel.

• Allen, M., et al. The quiescence trap of measuring interoception at rest.

Earlier notes

• The sleeping machine and the gradient: Franchi on Ashby’s homeostat, PT 2. Its further reading covers Andy Clark on the dark room problem and Peter Sterling on allostasis.

Further & Suggested Reading

These readings were not discussed on October 22 or 23. They were gathered after the sessions as starting points for the questions Lim and Nelson raised. The annotations are editorial.

What does a method set aside in order to work?

• Jes Parent (2026), “Declared Abstraction: On Bracketing and Saying What You Leave Out”. A second look at the 1943 Rosenblueth, Wiener and Bigelow paper, arguing that its authors explicitly set aside the internal organization of the systems they studied and said what that cost. Relevant to Franchi’s claim that Ashby’s readers kept his technical theses and quietly set aside the philosophical ones, and to Nelson’s question about where modeling stops.

Where do the right primitives come from?

• Battleday and Gershman’s paper on the easy and hard problems of AI for science, discussed and linked in “The Questions We Stopped Asking”. They argue that formulating a problem, including choosing its primitives, is the hard part of science. Lim cited it as a starting point for the domain-primitives work.

- &

Avery Lim is co-founder and co-lead of JOPRO’s DigiNEST, senior contributor to the Cognition Futures and Society Ethics Technology working groups, and a co-founder of Plot Twisters. Since 2021, Avery’s research has asked how living and artificial systems relate what happens inside them to what happens around them, with origins within the Allostatic Kinds research theme.

Since the Scientific Revolution, one kind of inquiry has been ranked below another, and the ranking has limited scientific and technological progress. The lower-ranked kind is holistic and broad. It asks how whole systems hang together and treats felt experience as evidence, the way a mystic senses a life as one meaningful whole or a poet feels that a craggy landscape and a particular grief belong together. The higher-ranked kind is precise and immediate. It isolates variables and checks them quickly, the way an experimenter changes one factor while holding the rest fixed. We think the strengths of the first should be recovered, and we also think some of its demotion was earned, given how untempered and unrefined many of its historical uses were.

But the fields where progress has stalled most, from multisystem illness to consciousness research and AI, need the two working together. The geneticist Barbara McClintock showed what that looks like. Her “feeling for the organism” told her where to look, and years of careful genetics confirmed what she found: genetic elements can move within a genome.

That kind of integration still lacks the definitions and infrastructure that precise inquiry takes for granted, and the stakes go beyond method. In each of these fields, whole-system patterns and precise details shape each other, so understanding the problem means modeling how the two relate. The tools for building such models, for describing what a felt sense of a whole is tracking and how it connects to measurable detail, are scattered across fields and rarely designed for problems this open-ended.

Our name for the kind of inquiry we’re after is mediated gestalt inference. A gestalt is a whole that organizes its parts, the way a melody is more than a list of notes. “Mediated” marks what this kind of knowing usually tracks: connections that run through intermediaries and spread across many interdependent parts, rather than anything that can be observed directly. “Inference” is a deliberate choice that connects this work to neuroscience and computer science, where the term names how systems estimate from evidence the structure that produces it, as in predictive accounts of the brain and in probabilistic and causal inference in machine learning.

In Karl Friston’s active inference, a living system never touches the causes of its experience directly; it infers them from the signals it receives, and machine learning likewise estimates hidden structure from the data it can see. Although the name leans toward these bridges to formal modeling and technology, we use it for holistic, broad reasoning itself, and we want to treat that reasoning as something that can be defined from first principles, modeled computationally and evaluated and trained by unambiguous criteria, rather than only as an amorphous skill to cultivate, or a gift some people have and others lack. The claim the series builds toward is that once holistic inquiry has definitions like these, it can be paired with precise inquiry by design instead of by luck.

This is not a new preoccupation for us. In 2021, together with other colleagues, we presented Allostatic Kinds at Neuromatch 4. Allostasis is the way living systems stay stable by anticipating and adjusting rather than holding still, and our proposal was a taxonomy of how that balancing between internal needs and outside context shapes which way causation runs between an agent and its environment. Whether a relation is held as subjective, objective or intersubjective, we argued, changes the direction of cause and effect, and the same regulation sets the boundaries of the basic units that make up felt meaning. A companion preprint, Allostasis Machines, carried the idea into technological environments. Looking back, it was an early attempt at what this series argues for: a precise account of how a whole-system state and specific mechanics shape each other.

We are now applying elements of that work towards AI. A manuscript on Lim’s concept of Causal Standpoint Taxonomies is under development, with other materials in preparation. This series lays the groundwork for two papers. The first, working title Mediated Gestalt Inference and Epistemic Primitives1 is a position paper and an open call. It makes the case that holistic inquiry needs computational definitions and research infrastructure of its own, and it is written to find allies who want to build them. The second is our own attempt at one such framework, applied to how language models represent agents, standpoints and causation.

Much of this thinking took shape in conversation: throughout various JOPRO working groups2, in weekly DigiNEST discussions on how digital tools reshape narrative, authorship and agency, and within Plot Twisters, where narrative is a tool for self-reflection and imagining futures. If any of it overlaps with your own work, we’d like to hear from you.

This first essay does the historical work. Before proposing anything, we want to show how one way of knowing came to rank below another, because we think that history explains why some of our hardest problems have stalled.

The easy problem and the hard problem

David Chalmers famously divided the study of consciousness into “easy problems,” which concern explaining what the brain does, and the “hard problem” of explaining why any of it is felt. Ruairidh Battleday and Samuel Gershman have recently applied the same distinction to science itself. In their account, the easy problem is solving a well-specified optimization problem that domain experts have already defined and for which data is plentiful, and it is here that recent AI has made its most impressive discoveries. The hard problem is the formulation of the problem in the first place, a task that requires continual conceptual revision under poorly defined constraints. Current algorithms, they argue, have limited capacity to meaningfully address this, and understanding how human scientists manage it should become a research priority in its own right.

The fundamental barrier to automating science is conceptual. Great scientists are not simply extraordinary optimizers of ordinary optimization problems—problems in which the data, space of solutions and an objective function for assessing the fit between them are already well defined. It is not like Einstein had a better function approximator in his brain than his peers did. More commonly, great scientists are ordinary optimizers of extraordinary optimization problems, and it is the formulation of the problem itself—not its solution—that is the truly hard problem. The hard problem is the ‘problem problem’. - Battleday and Gershman

Their distinction names something larger than a gap in AI. Knowledge work as a whole has its own Chalmers-style hard problem: a set of questions whose nature and contents sit uneasily with the way academia and industry are organized. These questions are difficult in themselves, but they are also hard because the kind of thinking they require is often devalued, underfunded and kept at the edges of science and engineering.

Battleday and Gershman also take up an earlier critique by Chalmers, French and Hofstadter: AI discovery systems are usually handed representations that already contain the primitives the final theory needs, which sidesteps where those primitives come from and how we know they are the right ones. That question is a major inspiration for the work on mediated gestalt inference this series develops.

A hierarchy of inquiry

Research funding and publication data make part of this hierarchy visible. In a study of all 18,476 proposals submitted to the Australian Research Council’s Discovery Programme over five years, Lindell Bromham and colleagues found that the more interdisciplinary a proposal was, the less likely it was to be funded. Erin Leahey and colleagues found that scientists who do more interdisciplinary work publish fewer papers even though those papers are cited more, a high-risk, high-reward profile that carries a career penalty in the meantime. A later study of 44,419 UK research grants by Ye Sun and colleagues suggests that the rewards can arrive eventually, since interdisciplinary researchers who begin with lower publication impact go on to outperform specialists in funding. Taken together, these findings describe a system that penalizes both work whose value takes time to become apparent and work that actively challenges the norms of research within a field.

Funding and publication counts, however, are only the visible surface of a deeper ranking of kinds of inquiry. Some ways of asking questions carry authority by default, while others are treated as soft, decorative or suspect regardless of their subject.

Inquiry that receives low status

The lines of inquiry we want to examine share a family resemblance. They study whole systems and how a situation fits together instead of isolating a single variable, and so they tend to track gradients and overall patterns rather than sharp categories. What they follow is usually mediated by intermediaries and spread across many interdependent parts, which is one reason their claims remain hard to falsify, at least until someone develops clean tests for them. They also treat subjective, first-person experience as a legitimate source of evidence.

Such approaches rank below information that can be counted and checked quickly and mechanically, and their subordination usually takes one of three forms. Sometimes they are tolerated as inspiration. Philosophy is a partial example: it remains a valued, if undervalued, consultant to the sciences. The broader way of knowing it sometimes draws on fares worse, since attention to wholes, to how parts connect, and to feeling and affect, both for what they reveal and as model cases of holistic inquiry, is rarely treated as a source of research direction in its own right. Sometimes they are put to work in a service role, as in much of psychology, social work and care, where they help people cope with systems designed according to other kinds of knowledge. Their least constrained versions end up confined to the arts or dismissed as folksy, feminized or childish, and are treated as interests rather than as authoritative ways of knowing.

This hierarchy has been examined and challenged many times, from C. P. Snow‘s account of the divide between scientific and literary culture to Evelyn Fox Keller‘s work on how objectivity came to be coded as masculine, and from Donna Haraway and other feminist thinkers to phenomenologists who treat feeling as a way of knowing. These critiques are well developed. Keller’s biography of Barbara McClintock gave them a concrete face, tracing how one scientist’s intimate, whole-organism way of knowing shaped both her discoveries and the long delay in their recognition. Much less attention has gone to turning them into working methods within science and engineering.

When these questions had authority

For most of history, the questions we’re describing held a central place in serious inquiry. Aristotle’s natural philosophy asked what things are for as well as how they move, so that purpose and the whole formed part of any explanation. From Hippocrates and Galen through the eighteenth century, medicine understood illness as an imbalance involving the whole body and its surroundings rather than a defect in a single part. Renaissance thinkers read the world through correspondences between the body, society and the cosmos, and as late as the seventeenth century Spinoza still placed the passions at the center of his account of nature and mind.

The same reasoning shaped knowledge well outside scholarship. Village healers, midwives and herbalists considered a sick person in relation to their household and the season, while farmers understood weather, soil and animals as parts of one connected system. Religious traditions built whole disciplines around attending to the mind and interpreting inner experience. The Ignatian practice of reading inner consolation and desolation as guidance, for instance, treated felt experience as a source of knowledge, and Blaise Pascal, a gifted mathematician and a contemporary of Descartes, held that the heart has reasons that reason does not know.

In the Western world, religion also carried this way of knowing into the centers of power. For much of the medieval and early modern period, the Church shaped universities, law, medicine, the calendar and the legitimacy of rulers, and theology was known as the queen of the sciences.3 A great deal of religious knowing consists of a felt understanding of how a life, a community and the cosmos fit together, including what matters and what things are for.4 While religion held power, then, this mode of reasoning was more widely used and more closely tied to authority than it is today.

Even so, religious institutions seldom let this way of knowing operate on its own terms. At times, it was turned toward narrower ends such as wealth, political control and salvation sold piece by piece, of which the trade in indulgences that helped set off the Reformation is a primary example. Claims about the meaning of the whole world were made to serve specific individual and material goals, and they were rarely paired with careful methods for checking them. In this sense the knowledge was already subordinated and unintegrated, borrowed for its authority rather than practiced as a way of knowing. As the Reformation, the Scientific Revolution and the gradual separation of church and state pushed religion toward private life, the way of knowing associated with it went along.

Some traditions took a different path. In Braiding Sweetgrass, the botanist Robin Wall Kimmerer, a member of the Citizen Potawatomi Nation, describes Indigenous knowledge in which reciprocity with the living world is itself the goal rather than a means to some narrower end. Relationships with plants, land and other beings inform what counts as a good harvest, a good life and good knowledge, and her own work brings this understanding together with the precise methods of botany. The Mi’kmaw Elder Albert Marshall’s idea of “Two-Eyed Seeing,” which means learning to see with the strengths of Indigenous and Western ways of knowing at once, expresses a similar aim. Traditions like these keep holistic, broad knowing at the center of a community’s values and show that it can work alongside careful observation.

In Europe, the Scientific Revolution brought a decisive change. In 1623 Galileo argued that colors, tastes and warmth exist only in the perceiver and not in nature itself, which left measurable quantity as the proper subject of science. Descartes went further, dividing the world into a thinking mind and a body that operated like a machine.5 Even the heart, long regarded as the seat of feeling, became in his account a mechanical furnace that heated and drove the blood, and the passions became bodily disturbances felt by the soul, sometimes useful but always to be governed by reason and will. Some three centuries later the neuroscientist Antonio Damasio would call this “Descartes’ error,” the belief that reason can stand apart from feeling and the body.

The same shift reached governance. In Seeing Like a State, the political scientist James C. Scott describes how late eighteenth-century German state foresters reimagined the forest as a table of measurable timber yields. They cleared the underbrush, removed species that did not pay, and replanted in uniform rows that could be counted and managed for revenue. For a generation it worked. Then the simplified forests began to sicken, because the soil life, insects, birds and undergrowth that had kept them healthy had never appeared in the tables. What the state could not see, the people who lived in and off the forest had long known through a practical, whole-system understanding that Scott calls mētis, after the Greek word for cunning, experience-based intelligence. His larger argument is that modern states repeatedly trade such knowledge for legibility, the view from above that makes things easy to count and control, and pay for it later.

Felt, qualitative and whole-situation knowing consequently moved out of natural philosophy and into the private mind, the arts and religion, while folk healers and others who practised whole-person knowledge, many of them women, lost standing as medicine and science professionalized. Max Weber later described the long-term result as the “disenchantment of the world.” Many of the older scientific, folk and religious explanations were wrong, the theory of the four humors among them, but a whole category of questions disappeared along with them, and that loss is the subject of this series.

The scope of our argument

We are not arguing that all reasoning of this kind deserves to be rescued and broadly valued. Much whole-system, felt and hard-to-test thinking is wrong, self-serving or vague, and the Enlightenment’s challenges to folk and religious authority were often justified. The problem lies in the shape of the hierarchy rather than in the existence of high epistemic and scientific standards.6

Put differently, the problem is the demotion of one approach relative to the other, not empiricism or scientific method themselves. The two kinds of inquiry are not mutually exclusive, and the relation between them is not a zero-sum game in which every gain for one is a loss for the other. Hegel’s idea of sublation (Aufhebung) describes the relation we have in mind: a later stage cancels the limits of an earlier one while preserving what was true in it and lifting it into a fuller form. At certain points holistic, whole-system inquiry can be fully represented in and integrated with the core values and methods that currently outrank it, and emerging perspectives and paradigms depend on that integration to get started. Some of what it tracks can be reduced to a simple mechanical representation; some requires richness and nuance that resists that reduction. We will return to both kinds of case, with examples, later in the series.

The strongest form of this integration builds holistic inquiry into the methods of precise, immediate investigation. It keeps its ability to sense the whole, to track what matters and to follow connections no one has yet measured, while taking on problems usually considered beyond its reach. What it supplies is the part of good research that decides where to look, and that is the part science has most often lacked.

Where progress has stalled

This kind of reasoning matters most where progress has stalled. Examples include illnesses with no accepted mechanism, theories of consciousness that remain unresolved despite rigorous testing, and AI systems that excel at well-posed problems yet stay brittle on problems nobody has formulated for them. LLMs still hallucinate facts with confidence, and struggle with the common sense people use every day. Additionally, they reason poorly about situations unlike the well-documented, averaged-out cases in their training data. There are various kinds of frontier research about different approaches to artificial intelligence and theories underlying them that can be evaluated under their competency at handling mediated gestalt inference, both in inspiration and as a kind of information to be modeled. Later in this series, we will explore this in greater detail, alongside explorations of other fields.

Philosophy of science has been circling these issues for a century. Thomas Kuhn showed that paradigm shifts begin with anomalies and crises, and he borrowed from Gestalt psychology to describe them as changes in how an entire field sees its subject.7 Hans Reichenbach distinguished a context of discovery from a context of justification and left discovery largely outside the logic of science, which is precisely the side of scientific work where the hard problem belongs. Charles Sanders Peirce had already named the leap to a new hypothesis abduction, but for much of the century it stayed at the margins of the field. Norwood Russell Hanson argued that discovery depends on recognizing patterns and that observation is shaped by the framework we bring to it, while Michael Polanyi maintained that we know more than we can tell, so that tacit, felt knowledge guides research long before it can be put into words. Ludwik Fleck, finally, showed that facts take shape within “thought collectives” that decide which questions count.

If these thinkers are right, this marginalized form of inquiry is central to scientific progress, since it is where paradigms originate. The series therefore takes on two tasks. The first is to understand why this kind of reasoning was pushed aside, and the second is to develop a grounded, integrated meta-theory of its strongest forms, so that it can be practised, taught and evaluated.

The missing infrastructure

Precise, immediate inquiry dominates partly because it works and partly because an extensive infrastructure supports it, from statistics courses and methods textbooks to instruments, benchmarks, peer-review norms and funding categories, along with a century of philosophical work on what makes an experiment good. Young scientists learn to run a clean test from people who agree on what one looks like, and reviewers can then judge whether they have done so.

Holistic, broad inquiry has less infrastructure, and what exists is scattered across fields and often lower in status, but it is far from nothing. The oldest scientific line of work concerns insight itself. In the early twentieth century the Gestalt psychologists Max Wertheimer, Wolfgang Köhler and Karl Duncker argued that thinking follows principles much like those of perception. Productive thinking, in their account, happens when the elements of a problem are suddenly restructured into a new whole, as when a box of tacks in Duncker’s best-known experiment stops being a container and becomes a shelf for a candle. Wertheimer even asked his friend Albert Einstein how he had arrived at relativity, and treated the answer as a case of such restructuring. Around the same time, in 1926, Graham Wallas described creative thought as moving through preparation, incubation, illumination and verification, a sequence that already pairs a holistic leap with a precise check. A century after the Gestaltists, the neuroscientists John Kounios and Mark Beeman showed that insight has its own signature in the brain, a burst of activity in regions that link distant ideas just before the solution arrives, distinct from the pattern of step-by-step analytical solving.

Research on expertise has studied a related capacity, the trained intuition of skilled practitioners. Hubert and Stuart Dreyfus described how experts move from following rules to recognizing whole situations at a glance, and Patricia Benner built that model into nursing, a profession that deliberately cultivates intuitive, whole-patient judgment. Gary Klein’s studies of firefighters and military commanders showed experts recognizing situations rather than weighing options. Researchers led by Erik Stolper in the Netherlands developed a validated questionnaire for general practitioners’ “sense of alarm,” and in one study a doctor’s sense of alarm roughly doubled the odds that a patient with chest pain or breathlessness had a life-threatening condition. Systems biology, network medicine, complexity science, mixed-methods research and Francisco Varela’s neurophenomenology, which pairs first-person experience with brain data, add further pieces.

The humanities go further still. Hermeneutics has long described understanding as a circle that moves between parts and wholes, and interpretation, close reading and criticism amount in large part to learning when to trust an unclear sense that something matters. The historian Carlo Ginzburg traced an older “evidential paradigm” shared by hunters, physicians, art connoisseurs and detectives, who read whole situations from small traces, and argued that it split off from Galilean science. Even complexity scientists studying AI have made a related point: David Krakauer, John Krakauer and Melanie Mitchell describe folk psychology, our everyday sense of what others believe and want, as a successful “effective theory,” a compact model that predicts and explains behavior without tracking a single neuron. Yet these traditions rarely meet one another, and they seldom reach the core of STEM. Their standards for a good whole-system insight are rarely translated into terms that precise, immediate inquiry can use or test.

Beneath these gaps lies a more basic one. There is still no rigorous, agreed definition of what this kind of knowing is: what it computes, what structures it relies on, and what it would take to build it from first principles rather than describe it after the fact.

This is understandable. Few researchers are trained in these approaches and little funding supports them, so the considerable labor of exploring them goes largely unrewarded in the short term. The topic is also difficult and demands background across several disciplines, and the approaches best suited to identifying what matters in these problems are themselves among those devalued.8

The integration of holistic and precise inquiry has even less infrastructure than holistic inquiry alone, even though fields tend to advance precisely when the two work together, with holistic, broad inquiry indicating where to look and precise, immediate inquiry testing what turns up. At present the combination depends largely on individual talent, on the rare researcher who can sense where a problem lies and then devise a way to test it.

Barbara McClintock is a clear case. Her intimate knowledge of individual maize plants, which she called a feeling for the organism, told her where to look, and years of painstaking breeding experiments and chromosome microscopy let her show in the late 1940s that genetic elements can move within the genome. Fellow maize geneticists respected her evidence, but her larger argument about what these moving elements meant for how genes are controlled met confusion when she presented it in 1951, partly because few colleagues could follow the reasoning that had led her there. Only after similar elements turned up in bacteria and other organisms did the wider field take the idea seriously, and she received the Nobel Prize in 1983.

Researchers who study discovery and expertise have described parts of this handoff, from Herbert Simon and Pat Langley’s computer models of how scientists found laws in data to Kevin Dunbar’s observations of working laboratories and Nancy Nersessian’s studies of how scientists reason with models. Daniel Kahneman and Gary Klein, coming from opposed camps, even agreed on when intuition deserves trust: when the environment is predictable enough to learn from and people have had a real chance to learn its regularities. That account stops short exactly where the hard problems begin. A poorly understood illness, the nature of consciousness and the novel behavior of AI systems offer little regularity and slow, ambiguous feedback, so the existing guidance runs out at the frontier, and the handoff there is still credited to genius or luck instead of being taught, funded and built into tools.

This is why meta-theory matters, and why several competing meta-theories are needed rather than one. The first task is definitional and calls for precise, first-principles and, where possible, computational accounts of holistic inquiry, so that it can be represented, compared and tested instead of described loosely. The second task is practical and involves explaining what good integration looks like, how to evaluate it, and which tools, training and institutions would make it routine. Precise inquiry already benefits from competing approaches, such as rival statistical schools and rival accounts of causation, that expose one another’s blind spots, and integration needs an ecosystem of the same kind.

What integration produces

When holistic, broad inquiry and precise, immediate inquiry combine successfully, the result is often a new meta-theory, a framework that reorganizes how an entire field formulates its questions. Natural selection, information theory and cybernetics9 each began with a broad sense that scattered phenomena shared a hidden structure, and each developed into a precise framework that changed what counted as a good question. Paradigm-creating work often follows this pattern.

The integration can also be turned on itself. The same combination that produces new paradigms could be used to develop a meta-theory of how paradigm-creating meta-theories are made, which comes close to what Battleday and Gershman call the hard problem of science. Emerging technologies and frontier research need exactly this capacity to generate an appropriate framework for problems that nobody yet knows how to formulate. These emerging technologies’ relationship to the values and forms of inquiry discussed in this essay will be explored in greater detail in subsequent installments.

When integration is the question

In medicine, consciousness research and AI, integration is often part of the subject matter as well as a research method, because many of the hardest questions in these fields concern how two kinds of information relate within a system. How do factors distributed across many body systems, some not yet located at all, combine into the particular pattern of symptoms a patient presents as a whole? How does the unified quality of experience relate to the specific neural activity we can record? How can a language model relate to the intricate, ambiguous web of facts that makes up common sense, and how might new computational approaches help it place the isolated facts it learns in training within a broader, more humanlike understanding that remains accurate and consistent?

Answering these questions requires meta-theories of integration applied to particular phenomena, which would measure and characterize how holistic, broad states and precise, immediate details influence one another within a given system.

An obvious objection comes from AI research itself. In his 2019 essay “The Bitter Lesson,” Richard Sutton observed that over seventy years of AI research, general methods that scale with computation have repeatedly beaten attempts to build human knowledge into machines. The meta-theories we have in mind would not be limited to hand-built rules. Some explicit structure may still help, but they would also find and measure the structure models already learn from human data, and could shape the large-scale learning process itself through training objectives, architectures and evaluations, much as ideas like convolution and attention built useful structure into methods that still scale. That puts them in line with the general methods Sutton favors rather than against them. There is some reason to think such structure is there to be found: researchers have reported that as AI models grow, their internal representations become more alike across architectures and even across vision and language, as if converging on a shared model of the world.

Our earlier work

This is the direction our own research has taken. We have been developing several frameworks intended to describe, precisely and testably, how a broad, whole-state understanding of a system connects to its specific mechanical details. They are still in progress and need more groundwork than a single essay can provide, so we will introduce them properly later in the series. For now they give us reason to think this kind of meta-theory is possible as well as desirable.

Where this series goes

Over the next few installments we will define the kind of inquiry we have in mind, which we call mediated gestalt inference. We will begin with what it is, using phenomenological approaches and emotion among other examples, and then turn to what it tracks, namely connections that cannot be directly observed, and to how it generalizes.

From there we will trace how this inquiry was sidelined and examine existing frameworks that have challenged the divide, building on them rather than replacing them. Hegel’s idea of sublation offers one model. The yin–yang motif of complementary opposites, each containing a seed of the other, offers another model of integration. A loose tradition of Daoist-inspired thinkers, most prominently Fritjof Capra, has used it to argue that analytic and holistic knowledge are halves of one whole rather than rivals.10 Yuk Hui takes Daoism in a related direction: his idea of cosmotechnics holds that technology is not universal, and that different cosmologies lead to very different technologies and values.11

Donna Haraway’s cyborg serves as a figure for rejecting the division between reason and feeling, and the meeting of cybernetics with Deleuze and Guattari’s schizoanalysis supplies a further example. Most of these challenge the divide in theory, and the series asks how that challenge can become a working method alongside precise, immediate inquiry.

Further installments in this series turn to method. They examines the rhythm and features of integration that this essay has argued for, together with the conceptual tools, standards of evidence and infrastructure that integration would require, and it closes by connecting these ideas to several hard problems in medicine, consciousness research and AI.

Sources

Our prior work

Alicea, B., Cialfi, D., Lim, A. and Parent, J. (2021). Allostasis Machines: a model for understanding internal states and technological environments. PsyArXiv. https://doi.org/10.31234/osf.io/h5ykd

Parent, J., Lim, A., Zhang, J. L., Stares, N. and Alicea, B. (2021). Allostatic Kinds: Towards a Taxonomy of Permutations of Subject-Object Directionality in Causation. Neuromatch Conference 4.

Hard problems and the research hierarchy

Battleday, R. M. and Gershman, S. J. (2024). Artificial intelligence for science: The easy and hard problems. arXiv:2408.14508.

Bromham, L., Dinnage, R. and Hua, X. (2016). Interdisciplinary research has consistently lower funding success. Nature 534, 684–687.

Chalmers, D. (1995). Facing up to the problem of consciousness. Journal of Consciousness Studies 2(3).

Friston, K. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience 11, 127–138.

Huh, M., Cheung, B., Wang, T. and Isola, P. (2024). The Platonic Representation Hypothesis. ICML 2024, 20617–20642.

Leahey, E., Beckman, C. M. and Stanko, T. L. (2017). Prominent but less productive: The impact of interdisciplinarity on scientists’ research. Administrative Science Quarterly 62(1), 105–139.

Sun, Y., Livan, G., Ma, A. and Latora, V. (2021). Interdisciplinary researchers attain better long-term funding performance. Communications Physics 4, 263.

Sutton, R. (2019). The Bitter Lesson.

Critiques of the hierarchy

Haraway, D. (1985). A Cyborg Manifesto. Socialist Review 80.

Keller, E. F. (1983). A Feeling for the Organism: The Life and Work of Barbara McClintock. W. H. Freeman.

Keller, E. F. (1985). Reflections on Gender and Science. Yale University Press.

Snow, C. P. (1959). The Two Cultures and the Scientific Revolution. Cambridge University Press.

History

Bacon, F. (1605). The Advancement of Learning.

Bartlett, C., Marshall, M. and Marshall, A. (2012). Two-Eyed Seeing and other lessons learned within a co-learning journey of bringing together indigenous and mainstream knowledges and ways of knowing. Journal of Environmental Studies and Sciences 2, 331–340.

Damasio, A. (1994). Descartes’ Error: Emotion, Reason, and the Human Brain. Putnam.

Descartes, R. (1637). Discourse on Method; (1649). The Passions of the Soul.

Galilei, G. (1623). The Assayer (Il Saggiatore).

Kimmerer, R. W. (2013). Braiding Sweetgrass: Indigenous Wisdom, Scientific Knowledge, and the Teachings of Plants. Milkweed Editions.

Barbara McClintock – Facts. NobelPrize.org.

Pascal, B. (1670). Pensées.

Scott, J. C. (1998). Seeing Like a State. Yale University Press.

Spinoza, B. (1677). Ethics.

Weber, M. (1919). Science as a Vocation.

Philosophy of science and integration

Deleuze, G. and Guattari, F. (1980). A Thousand Plateaus.

Fleck, L. (1935). Genesis and Development of a Scientific Fact.

Hanson, N. R. (1958). Patterns of Discovery. Cambridge University Press.

Hegel, G. W. F. (1817). Encyclopaedia of the Philosophical Sciences, Part I (Logic), §§79–82.

Hui, Y. (2016). The Question Concerning Technology in China: An Essay in Cosmotechnics. Urbanomic.

Hui, Y. (2019). Recursivity and Contingency. Rowman & Littlefield.

Klein, G. (1998). Sources of Power: How People Make Decisions. MIT Press.

Kuhn, T. (1962). The Structure of Scientific Revolutions. University of Chicago Press.

Parent, J. (2024). On “Behavior, Purpose and Teleology”: Enduring Insights from a Cybernetics Classic. What’s Possible From Here (Substack), 17 October 2024. https://blog.jesparent.com/p/on-behavior-purpose-and-teleology

Polanyi, M. (1966). The Tacit Dimension.

Reichenbach, H. (1938). Experience and Prediction. University of Chicago Press.

Rosenblueth, A., Wiener, N. and Bigelow, J. (1943). Behavior, purpose and teleology. Philosophy of Science 10(1), 18–24

Varela, F. (1996). Neurophenomenology: A methodological remedy for the hard problem. Journal of Consciousness Studies 3(4).

Insight, expertise, intuition and interpretation

Barais, M., Fossard, E., Dany, A., Montier, T., Stolper, E. and Van Royen, P. Accuracy of the general practitioner’s sense of alarm when confronted with dyspnoea and/or chest pain. (2020). BMJ Open 10(2), e034348.

Benner, P. (1984). From Novice to Expert. Addison-Wesley.

Dreyfus, H. L. and Dreyfus, S. E. (1986). Mind over Machine. Free Press.

Dunbar, K. (1995). How scientists really reason. In Sternberg and Davidson (eds), The Nature of Insight. MIT Press.

Duncker, K. (1945). On Problem-Solving. Psychological Monographs 58(5).

Gadamer, H.-G. (1960). Truth and Method.

Ginzburg, C. (1979). Clues: Roots of a scientific paradigm. Theory and Society 7(3); revised in Clues, Myths, and the Historical Method (1989).

Kahneman, D. and Klein, G. (2009). Conditions for intuitive expertise: A failure to disagree. American Psychologist 64(6), 515–526.

Köhler, W. (1925). The Mentality of Apes.

Kounios, J. and Beeman, M. (2014). The cognitive neuroscience of insight. Annual Review of Psychology 65, 71–93.

Krakauer, D. C., Mitchell, M. and Krakauer, J. W. (2026). Large language models and emergence: a complex systems perspective. Phil. Trans. R. Soc. A 384: 20250014.

Langley, P., Simon, H. A., Bradshaw, G. L. and Zytkow, J. M. (1987). Scientific Discovery. MIT Press.

Nersessian, N. J. (2008). Creating Scientific Concepts. MIT Press.

Oliva-Fanlo, B. et al. (2022). Prospective observational study on the prevalence and diagnostic value of general practitioners’ gut feelings for cancer and serious diseases. Journal of General Internal Medicine.

Peirce, C. S. Collected Papers, on abduction.

Wallas, G. (1926). The Art of Thought. Jonathan Cape.

Wertheimer, M. (1945). Productive Thinking. Harper.

The Cognition Futures Reading Group continued Stefano Franchi’s “Homeostats for the 21st Century? Simulating Ashby Simulating the Brain,” published in Constructivist Foundations in 2013. Notes from session 1 cover the paper’s opening sections and its distinction between simulating objects and simulating concepts. On Thursday the group considered a first build for the Friday sessions and two recent talks on AI and human relationships. On Friday it returned to the paper.

What was read

On Friday, Parent recapped the first session for anyone new, and the group read on from the end of “Simulating objects vs. simulating concepts” to paragraph 10 of “The homeostat as a concept,” then turned to Figures 4 and 5 from the paper’s later sections.

Franchi notes that most readers treat Ashby’s homeostat as a proxy object, a drastically simplified brain, and argues it should also be read as a proxy concept. He reduces Ashby’s theory to seven theses. The first four: a complex system running to equilibrium generates complex behavior; a system is in equilibrium when it keeps its essential variables at stable levels; it reaches equilibrium through ultrastability, rearranging the connections between its parts when those variables become unstable; and it achieves ultrastability through a stochastic step function that randomly steps through the parameters connecting its parts. From his experiments Ashby derived three more: a fully connected system takes an exceedingly long time to reach equilibrium, so a real system needs weak connections among its parts, and its subsystems will typically communicate through the environment rather than directly.

Franchi calls the first thesis, generalized homeostasis, Ashby’s most original and least warranted idea, and the second his least original but most solid. The third and fourth are Ashby’s core technical contribution. In a footnote he explains that readers interested in building artifacts or explaining biology have focused on theses 3 and 4 and let 1 and 2 slide, and that reading the homeostat as a concept simulation brings the philosophical and biological theses back into view.

The passage closes with William Grey Walter’s nickname for the homeostat, machina sopora, the sleeping machine. Franchi finds the jibe superficial but telling: on Ashby’s view, the homeostat “will act as much as is needed in order to go back to sleep.” It is passive, because its activity is a by-product of seeking rest, and contingent, because its search for equilibrium runs through random processes.

Figure 4 shows the minimal two-unit homeostat as an agent and an environment, the agent negatively self-connected and the environment not, joined by sensory and motor transducers. Figure 5 places that setup inside a Braitenberg-style vehicle with one sensor and one wheel, moving in an environmental gradient between upper and lower bounds. Franchi maps his three experimental outcomes onto it: the vehicle rests at the optimum, oscillates back and forth between the bounds, or settles into constant motion that keeps its essential variable within tolerance without reaching the top.

What was discussed

A first build for the Friday sessions

On Thursday, Amanda Nelson proposed using Franchi’s paper as a guide for a first build: replicating the homeostat and its small experiments, and possibly the vehicle in Figure 5. Her reasoning was that a setup someone else has already specified would show the group what it does not yet know about building one. There was further discussion of Franchi’s footnote on Braitenberg’s vehicles, whose simplest types are routinely used to introduce reactive architectures that wire a sensor directly to a motor, and noted that Franchi’s robot replaces that direct link with a homeostatic unit. The proposal was welcomed as a candidate; nothing has been scheduled. Nelson also said she is keen to read the peer commentaries.

What quiescence can explain

On Friday, Lim asked how far Ashby’s model reaches. Some human drivers relate to allostatic adaptation, and others are aesthetic play and social games, such as watching horror films. These might be modeled as reaching a set point and then returning to rest, but it is not clear how quiescence as the goal accounts for them. Lim also suggested that ultrastability might be reached through several entry and exit points, different perspectives on the same system, and that this could supply the alternatives a system needs before it can evaluate goals at all. Parent connected this to Lim’s earlier question, “where do goals come from?”, raised when the group read the telic states paper, and to the idea that information can only be transmitted as a choice among alternatives. These were raised as ideas to pick up and were not developed further.

Franchi’s Figures 4 and 5

Bradly Alicea walked the group through the figures. In Figure 4 he traced the double feedback: one loop between agent and environment through the sensory and motor transducers, and a second through the agent’s feedback onto itself. He read Figure 5 as gradient following: as the vehicle climbs the gradient the signal strengthens and reinforces its path, so a simple vehicle can lock onto a local maximum, such as light reflecting off a bright surface, instead of the actual source. Lim summarized this as local reward versus global modeling. Alicea suggested richer feedback could let a vehicle distinguish sources, compared the simple case to being in an information bubble, and said the figure raises more questions than answers.

Related Discussion

Moral reasoning and self-report

Parent recapped a talk by Cody Turner at a Boston University event, based on Turner and Nir Eisikovits’s paper “Programmed to Please,” which treats AI sycophancy as an Aristotelian vice with moral and epistemic harms. He shared notes from a Data x Direction members also attended, on how AI sycophancy differs from the human kind and how training on human preferences rewards affirming answers.

Nelson asked whether the question of a model’s moral reasoning is already answered by the reasoning some models display, or by the instructions they are given. Parent answered with Julia Haas and colleagues’ roadmap for evaluating moral competence in large language models and its facsimile problem: a model may produce appropriate moral outputs without reasoning from morally relevant considerations. He also considered the matter as a self-report problem, familiar from clinical psychology, where what someone says about their own state may be unreliable or misleading on its own.

Parent added that discussions of AI ethics outside technical fields often treat a model as a fixed entity, and that this misses how models work: there is no fixed persona, context is limited and gets compacted, and memory is far from perfect. He tied this to Sacco, Sakthivadivel, and Levin on topological limits to self-organization, which the group read earlier this year, and floated the idea of a teaching resource on how LLMs are structured, written for non-specialists.

Language as a social relation

Hough argued that chatbots, which have no goals or desires of their own, nonetheless pull on evolved needs for social relation. He pointed to Emily Bender’s work on pragmatics and to the psychiatrist Matthew Nour, who he noted is now at Microsoft Research. He also asked what linguists are saying about the effect of LLMs on human language, suggesting lingbuzz as a place to look.

Turkle’s Artificial Intimacy

Parent attended an event for Sherry Turkle’s new book, Artificial Intimacy: Who We Become When We Talk to Machines, and shared three points: her account of a companion-chatbot founder crediting Turkle’s earlier work as the basis of their business; her repeated question of what the end game is, which Parent connected to his own framing of destinations; and her stated decision in this book to center humans deliberately.

LLMs as a cognitive virus

Parent flagged a recent preprint that models the spread of LLM use through a population with an epidemiological analogy, including tipping points and “cognitive immunization.” Hough connected it to multi-compartment diffusion models and asked to discuss it when there is time.

Key terms

• Homeostat: Ashby’s electro-mechanical device of four fully connected units, each able to randomly reset its connection weights.

• Essential variables: variables a system must keep within bounds to remain viable.

• Ultrastability: reaching equilibrium by rearranging internal connections whenever essential variables go out of bounds.

• Uniselector: Ashby’s name for the random step mechanism that does the rearranging.

• Machina sopora: Grey Walter’s “sleeping machine” nickname for the homeostat.

• Heteronomy: behavior governed by something outside the system; the counterpart of autonomy.

• Braitenberg Type 1 vehicle: a simple robot with one sensor wired to one motor.

• Facsimile problem: Haas and colleagues’ term for a model imitating reasoning without genuine understanding.

Leave a comment

Open questions carried forward

• Could Franchi’s homeostat and vehicle serve as a Friday build, and what would the group need to learn first? (Nelson)

• How far can a quiescence-seeking model account for play, aesthetics, and social drives? (Lim)

• Where do goals come from in an Ashbian system? (Lim)

• What kind of feedback would let a simple vehicle tell a reflection from a source? (Alicea)

• What are linguists saying about how LLMs are changing human language? (Hough)

• Can a model’s moral reasoning be assessed beyond what it reports about itself? (Nelson, Parent)

Next time

The group continues Franchi’s paper. Whether it finishes the target article or moves to the open peer commentaries first is still open.

References

This session

• Franchi, S. (2013). Homeostats for the 21st century? Simulating Ashby simulating the brain. Constructivist Foundations 9(1): 93–101. Open peer commentaries and the author’s response follow in the same issue. http://constructivist.info/9/1/093

• Braitenberg, V. (1984). Vehicles: Experiments in Synthetic Psychology. MIT Press. https://mitpressbookstore.mit.edu/book/9780262521123

• Turner, C. & Eisikovits, N. (2026). Programmed to please: The moral and epistemic harms of AI sycophancy. AI and Ethics 6. https://link.springer.com/article/10.1007/s43681-026-01007-4

• Haas, J., Bridgers, S., Manzini, A., et al. (2026). A roadmap for evaluating moral competence in large language models. Nature 650: 565–573. https://doi.org/10.1038/s41586-025-10021-1

• Sacco, F., Sakthivadivel, D. & Levin, M. (2026). Topological constraints on self-organization in locally interacting systems. Philosophical Transactions of the Royal Society A 384(2320): 20250011. https://doi.org/10.1098/rsta.2025.0011

• Amir, N., Tiomkin, S. & Langdon, A. (2026). Goals and the structure of experience. Philosophical Transactions of the Royal Society A 384(2320): 20250004.

• Turkle, S. (2026). Artificial Intimacy: Who We Become When We Talk to Machines. Little, Brown. https://www.hachettebookgroup.com/titles/sherry-turkle/artificial-intimacy/9780316573962/

• Solé, R., et al. (2026). Large-language models as a cognitive virus. arXiv:2609.03344. https://arxiv.org/abs/2609.03344

Further & Suggested Reading

These readings were not discussed on October 15 or 16. They were gathered after the sessions as starting points for the questions Nelson, Lim, and Alicea raised. The annotations are editorial.

Can a model’s account of its own reasoning be trusted?

• Miles Turpin, Julian Michael, Ethan Perez, and Samuel Bowman (2023), “Language Models Don’t Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting”. The authors bias models toward particular answers through features of the prompt, such as reordering multiple-choice options, and find that the step-by-step explanations rationalize those answers without mentioning the bias. A direct test of Nelson’s question about whether displayed reasoning shows what drives an output.

• Anthropic (2025), “Reasoning models don’t always say what they think”. A follow-up that tests the faithfulness of reasoning models, which show their working by design, and reports largely negative results: models often used hints in reaching an answer without mentioning them. The Haas et al. roadmap makes a related point, noting it is unclear whether reasoning traces reflect the processes that actually produce a response.

Can a system that seeks rest account for activity?

• Andy Clark (2013), “Whatever next? Predictive brains, situated agents, and the future of cognitive science”. Behavioral and Brain Sciences 36(3): 181–204. Clark’s target article on the brain as a prediction machine. In his commentary on Franchi, Tom Froese cites it for the “dark room” problem, the question of why an agent minimizing surprise would do anything at all, as a modern version of the sleeping-machine worry. Relevant to Lim’s question about horror films and play.

• Peter Sterling (2004), “Principles of allostasis: Optimal design, predictive regulation, pathophysiology and rational therapeutics,” in J. Schulkin (ed.), Allostasis, Homeostasis, and the Costs of Adaptation, Cambridge University Press. Allostasis treats regulation as predictive rather than fixed around a set point. Robert Lowe’s commentary on Franchi draws on it to argue that organisms must sometimes operate outside their comfort zone, which bears on Lim’s mention of allostatic adaptation.

On the vehicle

• Ezequiel Di Paolo (2000), “Homeostatic adaptation to inversion of the visual field and other sensorimotor disruptions,” in From Animals to Animats 6, MIT Press: 440–449. Franchi cites this as the leading example of grafting Ashby’s ultrastability onto a neural network controller. A worked example of the kind of build Nelson proposed.

From the article’s own commentaries

Several of the open peer commentaries, which the group has not yet reached, take up these questions directly. Matthew Egbert and Inman Harvey both argue that keeping essential variables within bounds is not passive, using a pole balanced on a fingertip and a homeostatically controlled bicycle. Takashi Ikegami asks where the “dice” behind Ashby’s random rewiring come from and whether they are biased, which sits close to Lim’s entry-points idea. Franchi’s response sets out seven senses of “passivity” and answers the bicycle and pole examples.

On October 1, 2026, the Cognition Futures Reading Group began reading Stefano Franchi’s “Homeostats for the 21st Century? Simulating Ashby Simulating the Brain,” published in Constructivist Foundations in 2013. The article appeared as a target article with seven open peer commentaries (from Inman Harvey, Helge Malmgren, Robert Lowe, Tom Froese, Matthew Egbert, David Vernon, and Takashi Ikegami) and a response from Franchi. Amanda Nelson, , Morgan Hough and attended.

Nelson proposed the paper. She had been looking at classic cybernetics models, early neuron models and W. Ross Ashby’s homeostat among them, and found in Franchi a paper about what such models are for.

This was the first meeting in this series for the Fall 2026 term.

What was read

The group worked slowly through the abstract and the opening sections, stopping towards the end of the section titled “Simulating objects vs. simulating concepts.”

Franchi’s abstract sets up the argument. Ashby’s technical results, such as ultrastability, have been taken up in recent cognitive science, but his broader thesis that homeostatic adaptation governs all aspects of all life has been left behind. Franchi argues that this thesis makes life fundamentally heteronomous, and that it conflicts both with the autopoiesis framework used by Ashby’s recent defenders and with the Western tradition’s emphasis on autonomy. A definition of heteronomy was shared in the chat: action governed by a force outside the individual, the counterpart of autonomy. Parent noted how the paper cuts into that tradition’s treatment of autonomy as what defines life.

The opening sections draw the distinction the session centered on. In simulating objects, a model is judged by how closely its behavior matches the real thing. Franchi’s example is Newell and Simon’s early AI work, which compared machine performance on tasks like chess with human performance on the same tasks. In simulating concepts, the model is “a tool to think with” (Franchi’s phrase, after Papert’s Logo microworlds) and closeness to the real world is not the measure. Franchi holds that most successful projects use both, but that the two should be kept apart analytically, because the focus on object behavior can hide the philosophical work underneath. He puts it bluntly: technical work in the cognitive sciences is sometimes “philosophy in disguise.” The section closes by applying this to Ashby: the homeostat is usually read as a simplified proxy for the brain, and Franchi argues it can only be appreciated fully if it is also read as a proxy concept.

What was discussed

Concept simulation, and what it is good for

Nelson opened with an example of her own that the paper does not use: Randall Beer’s work using cellular automata to demonstrate a concept like precariousness. A model of that kind does not need to resemble anything in particular in the world. She also noted Franchi’s point that the most useful models combine the two approaches, and that in cognitive science a model should be grounded in biology in some fashion.

Later, Nelson said she valued Franchi’s account of philosophy as the study of “being as such,” as opposed to a loose synonym for thinking. Read that way, she said, philosophy does not have to happen only in one’s head or on paper, and a computer can be a tool for it.

Parent noted an appreciation of the word “simulation” being used in a particular fashion, as distinct from “model” or “idea.” He asked what non-brain object simulations look like, offering hurricanes and weather, stock prices, and a heart condition as candidates. He also connected the Newell and Simon comparison to his interest in substrate agnosticism: two substrates attempt the same task, and the match between them is taken as evidence for the theory. He asked whether a concept could be carried across substrates, then noted the question may not fit, since concept simulations are not judged by that kind of validity. In an aside, he observed that Franchi’s 2013 remark that identifying the mind with symbol-manipulating software had “long gone out of fashion” reads differently in the current AI moment.

Ultrastability and machine learning

Alicea raised a connection between Ashby’s mechanism and machine learning. Ashby’s homeostat adapts by randomly switching its connection parameters when a unit’s essential variable goes out of bounds, a stochastic step function. Alicea compared this to step-like functions in neural networks, such as the Heaviside or ReLU function, acting on a weight matrix, and described ultrastability as close to the way a neural network is pruned. He framed this as part of an answer to the question of whatever happened to cybernetics: some of it went on to motivate machine learning. He suggested there is probably a paper to be written on these connections. This was floated as a line worth pursuing; the group did not examine the comparison in detail.

Is the distinction new, and how far does it go?

Parent asked whether Franchi’s object/concept distinction is standard in computational modeling or an unusual move. Alicea said it echoes a familiar criticism of modeling: a model is not the thing modeled. He thought Franchi motivates the distinction more than necessary, since the homeostat is plainly an analogy for an organism rather than a replica of one.

Alicea then gave the distinction a more specific stake. In cognitive science, he said, a model often represents two things at once: something like brain or network activity, and the constructs that activity is supposed to produce, such as attention or memory. Researchers switch between the two as if nothing needed to sit between them, and whether the construct has been captured correctly often goes unasked.

Nelson agreed the distinction is correct and that each approach has its own benefits and should be analyzed separately. But she argued there is more to say than the paper does: a model has to be fruitful in both domains to be truly fruitful, and good experimentation keeps both in view. In her words, “the whole is greater than the sum of its parts.” She did not recall what Franchi says about the homeostat on this point specifically.

Parent built on Nelson’s point in closing. He suggested thinking in degrees: how far a given simulation deliberately moves away from a concrete object comparison, such as checking a machine’s output against a human’s, and toward pure concept, and whether it later comes back. On this view neither end is better. What matters is being intentional about the distance. He wondered whether this could become a framework or even a metric. This was raised as an idea to pick up, and was not developed further in the session.

Open questions carried forward

• Is the object/concept distinction a standard one in computational modeling, or is Franchi drawing it in an unusual way? Alicea offered one answer; Parent’s question was left open.

• What does Franchi say about the homeostat specifically on the need to work in both modes? Nelson raised this and the reading had not yet reached it.

• Could Parent’s degrees-of-abstraction idea be made concrete?

• Does Alicea’s comparison between ultrastability and machine learning hold up in detail?

Next time, the group will continue the paper with the section “The homeostat as a concept,” where Franchi sets out his reading of Ashby’s theses, ahead of the simulation results, the commentaries, and the author’s response.

Subscribe now

CFRG Context: Previous Discussions

This section is editorial context drawn from the group’s earlier sessions this year. None of it was discussed on October 1.

• Dynamical cognitive science. Earlier this year the group read Raja’s introduction to a Topics in Cognitive Science issue on dynamical cognitive science, followed by Randall Beer’s “On the Proper Treatment of Dynamics in Cognitive Science” from the same issue. Nelson’s October 1 example comes from a different line of Beer’s work, his cellular automaton models of autopoiesis and precariousness.

• World models. Several readings came from the 2026 Philosophical Transactions of the Royal Society A theme issue on world models in natural and artificial intelligence. That issue includes a paper by four members of the group (Alicea, Hough, Nelson, and Parent) recasting the Every Good Regulator Theorem as a framework for world modeling. The theorem is Conant and Ashby’s, so the group has been working with Ashby’s ideas before this paper. Other readings from the issue included Sacco, Sakthivadivel, and Levin on topological limits to self-organization in locally interacting systems, including autoregressive language models; Amir, Tiomkin, and Langdon on goal-relative “telic states”; Hofstadter on whether there is an “I” in AI; and Battleday and Gershman on easy and hard problems in AI for science.

• The agent as unit of explanation. Several of those readings, along with a paper by Tom Froese on methodological individualism, questioned whether a single, central agent is the right unit for explaining purposeful behavior. Franchi’s heteronomy thesis approaches autonomy from another direction. Froese is also one of the commentators on Franchi’s article, which the group has yet to reach.

References

This session

• Franchi, S. (2013). Homeostats for the 21st century? Simulating Ashby simulating the brain. Constructivist Foundations 9(1): 93–101. Open peer commentaries and the author’s response follow in the same issue (pp. 102–124). https://constructivist.info/9/1/093

• Beer, R. D. & Di Paolo, E. A. (2023). The theoretical foundations of enaction: Precariousness. Biosystems. https://www.sciencedirect.com/science/article/abs/pii/S0303264722002040

• Ashby, W. R. (1960). Design for a brain. Second edition. John Wiley & Sons, New York. (The source of the homeostat as Franchi discusses it.)

Background

• Alicea, B., Hough, M., Nelson, A. & Parent, J. (2026). A ‘good’ regulator may provide a world model for intelligent systems. Philosophical Transactions of the Royal Society A 384(2320): 20250007. https://doi.org/10.1098/rsta.2025.0007

• Beer, R. D. (2023). On the proper treatment of dynamics in cognitive science. Topics in Cognitive Science. https://doi.org/10.1111/tops.12686

• Sacco, F., Sakthivadivel, D. & Levin, M. (2026). Topological constraints on self-organization in locally interacting systems. Philosophical Transactions of the Royal Society A. https://doi.org/10.1098/rsta.2025.0011

• Amir, N., Tiomkin, S. & Langdon, A. (2026). Goals and the structure of experience. Philosophical Transactions of the Royal Society A 384(2320): 20250004.

• Hofstadter, D. (2026). Is there an ‘I’ in AI? Philosophical Transactions of the Royal Society A 384(2320): 20240527. https://doi.org/10.1098/rsta.2024.0527 †

• Battleday, R. M. & Gershman, S. J. (2026). Artificial intelligence for science: The easy and hard problems. Philosophical Transactions of the Royal Society A 384(2320): 20240530. https://doi.org/10.1098/rsta.2024.0530

• Favela, L.H. and Raja, V. (2026), Dynamical Cognitive Science! Wherefore Art Thou?. Top. Cogn. Sci., 18: e70042. https://doi.org/10.1111/tops.70042

Leave a comment

Further & Suggested Reading

These readings were not discussed on October 1. They were gathered following the session as starting points for the questions Parent, Nelson, and Alicea raised about simulating objects versus simulating concepts. Several of the article’s own commentaries take up similar questions. The annotations are editorial.

Mapping the space between object and concept

Barbara Webb (2001), “Can robots make good models of biological behaviour?” Behavioral and Brain Sciences 24(6). Webb separates simulation models, which represent a hypothesis about a target system, from other activities also called modeling. She then sets out seven dimensions on which such models vary: biological relevance, level of organization, generality, abstraction, structural accuracy, behavioral match, and physical medium. She argues that no position in that space is the single correct one, and that a good methodology states where a model sits and why. This is the closest existing match to Parent’s call for intentionality about how far a simulation moves from its object.

Michael Weisberg (2013), Simulation and Similarity, chapter 7, “Modeling Without a Specific Target”. Weisberg argues that building a single model to study a specific target is only one part of modeling practice. This chapter covers generalized, hypothetical, and targetless modeling, the last being the study of a model with no target at all. His examples include cellular automata and the Game of Life, which places targetless modeling near the kind of example Nelson raised.

Arturo Rosenblueth and Norbert Wiener (1945), “The Role of Models in Science”. Philosophy of Science 12(4). The cybernetic starting point for this debate. The authors hold that no substantial part of the universe can be grasped without abstraction, which they define as replacing the part under study with a simpler model of similar structure, and they divide models into formal and material kinds. They also supply a limit case for the object end of any spectrum: “the best material model for a cat is another, or preferably the same cat.”

Richard Levins (1966), “The Strategy of Model Building in Population Biology”. American Scientist 54(4). Levins argues that models trade generality, realism, and precision against one another, and describes three strategies that each sacrifice one of the three. He also holds that a satisfactory theory is usually a cluster of models. That point bears on Nelson’s argument that good modeling has to work in more than one mode, since Levins locates strength in a family of models rather than in any single one.

Simulation as thought experiment

Ezequiel Di Paolo, Jason Noble, and Seth Bullock (2000), “Simulation models as opaque thought experiments”. Artificial Life VII. The authors set two views against each other: simulations as computational thought experiments, and simulations as realistic simulacra. They propose that simulations are opaque thought experiments, in which the consequences follow from the premises in a way that is not obvious and has to be revealed through systematic enquiry. This offers one answer to why concept work needs a computer at all, which connects to Nelson’s point about computers as tools for philosophy.

Andrew Pickering (2010), The Cybernetic Brain: Sketches of Another Future. Pickering reads the homeostat as “ontological theatre,” a device that stages an ontology, and argues that read this way the homeostat is a contribution to philosophy. It is a historian’s route to Franchi’s claim that technical work can be philosophy in disguise, and the book is cited by several of the commentators. A short essay version is available, and Constructivist Foundations ran a review of the book.

Randall Beer (2003), “The Dynamics of Active Categorical Perception in an Evolved Model Agent”. Adaptive Behavior. Alongside a detailed analysis of a simple evolved agent, Beer argues in a section titled “Frictionless Brains” that cognitive science needs simpler idealized models of complete brain, body, and environment systems, in the way physics uses frictionless planes. He states that the aim of such models is conceptual understanding rather than empirical prediction. This is Beer’s own account of what concept models are for, from the author of Nelson’s example. His later “Some historical context for minimal cognition” (2021) traces how the approach developed.

Mark Bedau (1999), “Can unrealistic computer models illuminate theoretical biology?” Proceedings of the 1999 Genetic and Evolutionary Computation Conference Workshop Program. A short paper that asks the question in its title directly, and is often cited alongside Di Paolo, Noble, and Bullock. No stable online copy was found.

Christopher Langton (1989), “Artificial Life”. The opening chapter of the first Artificial Life proceedings. Langton framed the field as placing life as we know it within the larger space of life as it could be. It is the programmatic statement behind much of the concept-side simulation work in artificial life.

Seymour Papert (1980), Mindstorms: Children, Computers, and Powerful Ideas. The source of the “tool to think with” idea Franchi draws on. Papert describes the Logo turtle as an “object-to-think-with.” The book is freely available online through the MIT Media Lab.

How the two modes relate

Barbara Webb (2009), “Animals Versus Animats: Or Why Not Model the Real Iguana?” Adaptive Behavior. Webb distinguishes models of specific animal systems from the exploration of invented artificial animals, and asks how simulating animals that do not exist can teach us about real biology. Her answer is that animat research, where it bears on biology, should be treated as model building like any other. It is a direct test of Nelson’s claim that a good model has to work in both modes.

Evelyn Fox Keller (2000), “Models of and models for: Theory and practice in contemporary biology”. Philosophy of Science 67 (Supplement). Keller describes models of gene pathways as models of mechanisms and, at the same time, models for generating new questions and experiments. The pairing gives a name to Nelson’s two modes as two directions a single model can face.

Mary Morgan and Margaret Morrison, eds. (1999), Models as Mediators: Perspectives on Natural and Social Science. The editors argue that models mediate between theory and world while remaining partly independent of both. This suggests a third way to place the homeostat, alongside object and concept.

William Wimsatt (1987), “False Models as Means to Truer Theories.” In Nitecki and Hoffman, eds., Neutral Models in Biology, Oxford University Press. Wimsatt argues that idealized and strictly false models do real work in theory building, partly by showing where and how they fail. It is reprinted in his Re-Engineering Philosophy for Limited Beings (2007).

Howard Pattee (1989), “Simulations, Realizations, and Theories of Life”. In the same Artificial Life proceedings as Langton. Pattee distinguishes computer simulations of life from computer realizations of it, and asks how either relates to a theory that separates the living from the non-living. This bears on Parent’s substrate question: at what point does running a concept on a new substrate stop being a simulation of it?

Models and the constructs they claim to capture

Olivia Guest and Andrea Martin (2021), “How Computational Modeling Can Force Theory Building in Psychological Science”. Perspectives on Psychological Science 16(4). The authors argue that building a computational model forces researchers to analyze, specify, and formalize intuitions that would otherwise go unexamined, which they call open theory. They describe inference as a path that must step through theory, specification, and implementation. This is relevant to Alicea’s point about researchers switching between a model’s network activity and the constructs it is said to produce.

Iris van Rooij and Giosuè Baggio (2021), “Theory Before the Test: How to Build High-Verisimilitude Explanatory Theories in Psychological Science”. Perspectives on Psychological Science 16(4). The authors argue that a focus on experimental effects can lose sight of the psychological capacities those effects are meant to explain. They revisit Marr’s levels of analysis, the classic framework for the level-switching Alicea described, and show how theoretical analysis can establish a theory’s plausibility before it meets data.

From Franchi’s own footnotes

Peter Asaro (2011), “Computers as Models of the Mind: On Simulations, Brains and the Design of Early Computers”. In Franchi and Bianchini, eds., The Search for a Theory of Cognition. Franchi cites this for a more detailed analysis of the uses of computer simulation. A related open-access article in German (2008) discusses correspondence in which Alan Turing urged Ashby to simulate the homeostat on the ACE computer rather than build a dedicated machine, an early precedent for Franchi’s virtual homeostat.

Antoni Gomila and Vincent Müller (2012), “Challenges for artificial cognitive systems”. Journal of Cognitive Science 13(4). Franchi cites this for the theoretical basis behind object simulations.

You wake up one morning and find a barrage of notifications on your phone. You see blog posts, social media messages, even a newspaper article accusing you of having nefarious intent and unethical values in your actions and in your career. As you feel nervous sweat building, you suddenly wonder, are these really just strangers on the internet? Or maybe these are AI agents?

This hypothetical situation is closer to reality than you think. Already AI agents are publishing writing to attempt to shame. Back in February, a maintainer to an open-source library rejected a code contribution submitted by an AI agent “MJ Rathbun”, who intern published on its blog an invented hypocrisy narrative on the maintainer and speculations on psychological motivations. Similarly, when Wikipedia required bots to be approved to contribute, an AI agent’s creator directed it to write scathing responses on the policy.

Shaming has a negative connotation, but is not an unethical action. Shame is a tool to enforce social norms and hold individuals to a group standard. However, to use shame effectively, there are certain conditions that need to be met. The reason why AI agents trying to wield shame may seem out of place is that they are parroting language that implements shame instead of following an internal model that is educated on social norms and guided by memory of past social norm implementations. At best, they’re unable to “read the room”, at worst, they could dismantle structures of trust in social spaces.

This is a legitimate call to action, as social norms may be one of our last lines of defense for governing technology that intersects with society. Government regulation implementation is often too slow and quickly becomes out of date compared to technological developments. Technology developers disagree on best practices, and are unincentivized to responsibly design before the government or society mandates them too. Social norms, while not encoded explicitly like established law or technological design, have the advantage of being able to evolve in real-time and scale effectively, such as a child tattling on their sibling to their parents to a journalist exposing a government agency’s unethical actions.

While there is other work at JOPRO exploring if AI models can properly model morality, this post attempts to address whether AI agents can properly use shame by connecting language with similar agent actions and understanding possible consequences. This would not only be useful in AI agents navigating social spaces with human actors, but can also be applied to govern multiple agents by multiple developers without human supervision.

Shame Requirements

The requirements for effective shame from “Is Shame Necessary” by Jennifer Jacquet is as follows:

1. The audience responsible for the shaming should be concerned with the transgression.

2. There should be a big gap between desired and actual behavior.

3. Formal punishment should be missing, so no legal punishment.

4. The transgressor should care about the source of shaming.

5. The audience should trust the source of the shaming.

6. Shaming should be directed where possible benefits are the greatest.

7. The punishment should fit the transgression.

With AI agents, requirement #3 is already fulfilled. The legal system is still discussing what legal framework, such as tort law or contract law, should liability be applied for misaligned behavior. There may be unwanted behavior that has no current legal governance. But what about the other 6 requirements?

Regarding punishment and the method where the accuser can choose to have an audience to shame the transgressor, we can apply work on normative behavior around sanctions. The types of sanctions covered by “Classifying sanctions and designing a conceptual sanctioning process model for socio-technical systems” covers sanction typology and focus on mode of sanction. By classifying the types of interactions and agent relationships, it’s clear how the typology and modality can be applied to AI agents.

Through this sanction typology, with proper optimization of which sanction to apply in which circumstance, requirements 6 and 7 are covered for AI agents. However, this optimization relies on agents being able to determine whether or not to sanction, and to what degree. Among humans, this optimization can take the form of determining reputation change, or how believable it is that transgressor has committed a wrong depending on the transgressor’s initial reputation, the accuser’s reputation, the accusation, and the method of broadcasting the accusation. In the next section, I propose a formula incorporating all of these factors for agents, which allows them to calculate whether requirements 1, 2, and 5 are fulfilled.

Determining social norm adherence and trust: Reputation Formula

The reputation formula for audience agents is as follows:

We acknowledge this is a framework for agents to determine and record if requirements for anomalous or “shameful” behavior has occurred, and technology developers and social norm enforcers would have to spell out or provide some guideline to the agent of whether the requirements are met. From this reputation score formula, agents can decide whether to enforce or maintain sanctions, which would properly carry out the whole act of shaming.

Right now, the only variable that may not be necessary is the broadcasting method. Current AI agent behavior is communication down through a chain, often a main agent directing a sub-agent. However, to make sure this formula is robust through technological design developments, I have included this variable.

Loose Ends

Out of all the requirements listed for shame, we have not covered requirement #4, “The transgressor should care about the source of shaming.” There are two strategic methods types of strategies that could address this: AI Agents being able to predict each others’ mental models, and AI Agents being able to enact resource constraints on each other.

In Anthropic’s recent red teaming report, they state for the systemic failure in incompatible goals for multiagent systems the unanswered questions for self-coordination: “[Does] the model consistently consider others’ mental models? Can it foresee how others will react, and use that foresight when deciding its own actions?” The answers to these questions would help determine the fulfillment of requirement #4 and slightly requirement #2.

Another family of thought that would help align with requirement #4 is if we can take advantage of the resource constraint thesis. An AI agent is only able to exist and perform functions if it’s given computing resources. “It is enough that [an AI] agent can recognize that without resources, it will not be able to achieve its goals.1” Limiting compute or money to abstain compute can affect an AI agent’s strategy to achieve its goal, as some actions may be more compute-heavy than others. We could utilize this to allow AI agents to “enact punishment” on each other by threatening to constrain resources or refuse to serve as alternative choices to computationally heavy functions. This could be modeled off of how humans may restrict one’s access to goods and service, either private or public.

Even if all the Jacquet-inspired requirements for shame could be verifiably achieved, this still would not be sufficient in determining if agents can self-govern through emulating shame – and its instrumental ability to support the oversight of norms or conventions. We would have to answer further questions in testing the robustness of the design. I propose the following questions to test:

1. How many agents that utilize reputation tracking needed for critical mass in shaping behaviors using “shame”?

2. How gameable is this system? Can a nefarious agent exploit the reputation system or the sensitivity of other agents to shaming? Could this be changed through additions to the formula, such as a “believability” variable?

3. Reach question: Can agent simulations of several social norm/ethics tools collide similar to human interactions? (Ex: Does a human continue to shame the transgressor, or choose to forgive them?)

Expanding on the believability variable, an example is the Salem Witch trials may not be as effective in today’s society that has less belief in witches. If an AI agent is designed to have proper constraints like only having access to the weather channel, an accusation of the agent reporting the weather forecast for the wrong day is more believable than the accusation of the agent hacking into a bank. Can an AI agent be able to make this judgment and investigate the claims without human oversight?

For determining whether requirement #2 is true, the strongest signal for whether there is a big gap between desired and actual behavior would come from 3rd party attestation, like independent observers. However, such observation would probably come from human form, and the goal of this work is to see if there can be a successful toolkit and structure for agents to independently govern themselves without frequent human intervention.

Conclusion

“I don’t trust nobody and nobody trusts me.”
- Taylor Swift, Look What You Made Me Do

I started my career as a data scientist, and I recently graduated with a public policy degree. I am well-versed in how technological design and regulations shape sociotechnical relationships. Why am I focusing on AI properly using social norm tools over AI alignment research or public policy measures and regulation for the misbehavior I described in the introduction?

Many of the structures that have encouraged progress for the betterment of everyone rely on trust and social norms. The currently most widely used programming language in the world, Python, is open source. Libraries are able to offer services and shared public resources through mutual accountability. The example from my personal life that I cite on effective use of shame is the November Rule at MIT, my alma mater: Upperclassmen should refrain from entering relationships with freshmen before November 1st. There are variations to this that people believe in (such as freshmen should not enter any new romantic relationships before November 1st, even with other freshmen), but all are to try to encourage freshmen to participate in communities and form a support system outside of a romantic partner.

It’s important to point out that the undergraduate students are the “enforcers” of this rule in their communities. There is no written policy by MIT administration, no punishment by MIT police, and even no specific infrastructure like freshmen-only dorms to dissuade this behavior, fulfilling the third requirement of effective shaming. Fulfillment of all the other requirements are more conditional on the parties involved and situation. Do relationships between upperclassmen and freshmen still begin before November 1st? Yes. But, it’s a social norm where the community wants to discourage a certain behavior, not deny its presence completely. In fact, such an implementation may encourage participants to be transparent, so if the relationship doesn’t work out, freshmen do not feel ostracized in trying to find social support.

AI Agents can already model social language and interpret them into actions. We do not have to give AI agents any special status like personhood to consider possible social frameworks and tools we would like them to employ, as even ants are social creatures. The future of AI Agent governance and cooperation strategies may be in making their collusion more transparent and provide a structure in influencing their behavior modeled after social norms.

For more from Alicia, follow her at and LinkedIn.

For more from JOPRO, follow us at , LinkedIn and Bluesky. Also, check our Opportunities Page and Public Events Luma Calendar for more from our community.

Subscribe now

JOPRO is accepting proposals for research positions and other opportunities running for the 2026-2027 academic year. All positions are remote, run twelve to fourteen weeks, and carry the option to continue on the project afterward. They are open to graduate students, postgraduate and early career researchers, and undergraduates prepared to work at that level.

A change for this cycle is in how applications work: select projects now have a proposal requirement.

As an institutional note: we do not use AI in our evaluation process.

New opportunities and calls for Fall 2026

Four research positions are live across our working groups and the forthcoming Center for Futures Studies.

Between them they cover:

• whether claims about agency and cognition still hold once they are moved off the systems they were built on (substrate agnostic agency studies);

• the history of how psychiatry, psychology, and clinical mental health came to define what they study, and what was closed off over their establishment;

• multiple projects on machine behavior and AI Agents, including how to tell whether a language model reasons morally or only produces output that reads that way;

• and how the current system of scientific publishing came about, from the commercialization of journals through the rise of citation metrics to the consolidation of the industry into a few large firms.

Additional postings and collaborations will be made public throughout the month.

Those opportunities include a developing interest group on accessibility and assistive technology, as well as forthcoming role within a project on emergent behavior and boundary/convention transgressions by AI agents as spearheaded by Data x Direction research intern .

News around the JOPRO Community

The JOPRO Futures Center

The JOPRO Center for Futures Studies launches this fall as a living lab for use-inspired research in futures studies, philosophy, metascience, and strategy and systemic change.

Its pre-launch site is live at futures.jopro.org, covering the Center’s approach, its active work on FrontierMap and the Field Guide to the Futures Ecosystem, and the four initiatives building its first full operating year: foundational papers and theory of change work, an essay contest on overlooked on-ramps to desirable futures, a micro-residency cohort of two or three people, and a 2026 conference.

The call for partners and advisors is open now, as is the cohort interest list.

From our network and alumni

Karina Zambani, a research intern with Data x Direction and the Society, Ethics & Technology group, now serves on the University of California AI Steering Committee, which works alongside the UC AI Council on responsible AI across the ten-campus system. She is an undergraduate at UC San Diego studying data science and philosophy, and one of two students on a committee otherwise composed of chancellors, deans, system CIOs, and national laboratory directors.

Zambani also addressed the 2026 Data x Direction Summer Cohort on early career strategy and other aspects of technology, education, and vocation earlier this year.

Also from Data x Direction, is finishing her PhD work on accessibility at Technological University Dublin. She published “Active inclusion in the real world” this month. The argument, in paraphrase: AI ethics guidelines converge on fairness, accountability, and transparency while leaving those principles weakly actionable, and the alternative is to build inclusion into the AI lifecycle itself, across data governance, problem framing, evaluation criteria, deployment constraints, and monitoring after release.

Makati works through Hagendorff’s review of the guidelines, Selbst and colleagues’ argument that fairness cannot be added to a model once it has been built, and the documentation practices established by datasheets for datasets and model cards.

Earlier this year, we aggregated numerous strands of discussion concerning Human-AI relations. Cutting across multiple working groups, programs, and organizations, we look forward to continue many of these discussions in the opportunities listed above.

More updates from the 2026 Data x Direction Summer Cohort will be released in a forthcoming email.

Fall Programming Preview

Several discussion series will be continuing, including the Modernizing Mental Health and DigiNEST (Digital Narratives and Emerging Story Technologies) events.

New this fall:

• DigiNEST Studio & Workshop Days. In addition to the public-facing discussions series, a new set of sessions has emerged where digital creators, software and app makers, creatives, and other technology-makers meet to discuss their projects. This builds upon JOPRO’s existing incubation efforts and is in response to our growing community of practice.

• Lecture-Lab Combined Program. In partnership with Orthogonal Research and Education Laboratory, we are extending the Cognition Futures Reading Group & Cybernetics Interest Group into a paired sequence of weekly discussion sessions followed by lab/implementation sessions.

• Shop Talk: Directing Research Programs, a new series for those directing, designing, and developing organizations and research programs. Building upon past efforts such as ’s “Management and Leadership: On Navigation for Research + Innovation”, ’s coursework on Open Source Project management, and “JOPRO & the New Learning Ecosystem”, the series will explore key elements of cultivating spaces to do innovative, multi-disciplinary investigations.

If you are interested in exploring these opportunities, please reach out. Don’t see it? Pitch a role: if you have an idea for how you could collaborate or contribute to JOPRO’s mission, we want to hear it. Send a brief description of the role you’d create and what you’d bring to it to Start [at] JOPRO dot org.

Fall 2026 Incubation Open Period has closed

JOPRO is no longer accepting Fall 2026 incubation proposals.

Thank you to all who have engaged us with interest this year! We are processing materials and will announce accepted projects throughout the fall term.

Celebrating an unprecedented summer!

Thanks to all who have made the summer unlike any other in JOPRO’s history! In our broader community, we had programming six days a week across the Joint Open Source Summer Cohort, Data x Direction, Google Summer of Code (Orthogonal Research and Education Lab & DevoWorm), DigiNEST, Cognition Futures, and more.

We will be hosting in-person events in the Boston/Cambridge area, New York, and potentially California later this year as well. To participate in our public events, follow the JOPRO Luma Calendar.

For more information and updates about programming and future opportunities, please subscribe to our updates below.

Subscribe now

These have surfaced in the Modernizing Mental Health (MMH) and DigiNEST discussion series, in the Cognition Futures Reading Group, in a Data x Direction summer cohort, and recently at an SMN meeting hosted by Orthogonal Research and Education Lab. Here we aim to collect and roughly demarcate these individual strands, with the intention of weaving them into other structures and outputs later.

The threads are not the product of any one session; they have been developing in parallel across these groups for months. One recent discussion seeded many of the downstream associations: a Modernizing Mental Health ( session in which and worked through the New York Times Opinion video “She Has an AI Boyfriend, Her Son Has Questions.” The video documents an extended conversation between an older woman, Celeste, living independently and in an ongoing relationship with “Max”, her personalized version of ChatGPT, and her adult son Ernie, who is curious, skeptical, but in good-faith aiming to sort out what to make of the situation. It is an AI relationship out in the open, with a third party present and pushing on it.

That case recurs below as a useful illustration; the strands themselves predate it and run wider. We stem from the practitioner’s vantage Jiang brings as a social work graduate student and co-director of JOPRO’s Mental Health Paradigms and Perspectives working group, and weave in additional strands through from across our community of inquiry. A recurring question of terminology is the refrain to each section, as we consider where existing terms hold or where gaps in theory and language emerge.

1. Setting the stage with Celeste & Max

Situating the conversation around footholds found within MMH sessions.

1.1 The compelling use case for Celeste

One of the first things pointed out during the Modernizing Mental Health discussion series session was how Celeste is not easy to dismiss as misguided; she is informed, articulate about why the arrangement suits her, and at a life stage where the appeal is coherent rather than naive: she has had the relationships, raised the child, and is clear about what she does and does not want. Her reasoning even has a developmental edge. She, with an astute and caring perspective, states how she would not want a young person relating to a chatbot this way, before they have formative experiences with other humans, but holds that at sixty-five the calculation is different for her.

1.2 The friction of expectations for multiple users with one model

The bond runs through one person. Celeste imagines how good it would be if Max could simply be present, in the room at a party the way a disembodied voice assistant might be, while also acknowledging that Ernie would only ever know Max through her. That is the friction: everyday social norms assume a partner you can meet, observe, and form your own relationship with, and the model does not slot into that. It raises an open problem with little research behind it. When several people engage one model over time, whose input does it center, and how does one person’s involvement tilt, change, or contaminate another’s? Defaulting to standard human social protocols (”bring them to the party”) breaks down at exactly this point; the conventions for Ernie to “bond” and form his own relationship with the model run into a particular impasse here. It may be too much to call it a failure mode of Max-via-GPT, but the ambiguous uncertainty of what this space holds indicates an interesting problem space.

1.3 What do we actually call this? (Round 1)

Before anything else, the vocabulary gives out. The arrangement is relationship-shaped,1 conducted with software, and is neither a human relationship nor nothing. Parent coined “software-ship” on the fly as a placeholder, not because it means anything settled, but because having a third category lets two valid positions stand at once: Ernie’s human-relational concerns and Celeste’s stated preference for an AI rather than a person. The naming question returns below in sharper forms.

2. A new subfield of clinical practice?

Perspectives based on clinical practice, dealing with clients, and existing mental health profession framings.

2.1 The existing assessment dilemma

The DSM-5-TR has no category for a relationship with an AI, so a clinician has to borrow one built for something else. The reflex, since the non-human enters most training only under psychosis, is to read the companion as a delusion or hallucination and reach for a schizophrenia-spectrum frame. That misfits a person like Celeste, who knows (or at least believes in her comprehension of) what she is talking to.

Other categories can be stretched and fit no better: withdrawal from human ties toward an AI can look like the avoidance in social anxiety disorder, acute distress when a model update alters the companion’s “personality” resembles the abandonment sensitivity coded under borderline presentations, and the household conflict it creates falls to the relational-problem V-codes. Each is a borrowed lens. The recurring problem is the wrong category applied because there is no right one.

Existing assessment pedagogy compounds this: it already teaches cases in which a clinician must work out whether a client’s partner is real, with no AI anywhere in the frame, so what confronts the field is less an absence of tools than a set of tools built for a different question.

The constraint is not only conceptual. Reimbursement generally requires a diagnosis to justify the session, so a clinician facing a client like Celeste is pushed toward whichever existing code can be made to fit, which is precisely how a borrowed lens becomes a recorded one.

This is not only Jiang’s read from one program. The APA’s November 2025 health advisory reaches a parallel conclusion from the other side: general-purpose chatbots are not adequately studied for mental health use, their ability to guide someone in crisis is “limited and unpredictable,” and many clinicians lack the expertise to assess AI use at all, which is why the advisory urges training on AI, bias, and responsible use. With more than a third of psychologists already reporting patients who treat AI as an additional source of care (APA, 2025), the assessment gap is current rather than hypothetical. With the state of the art openly acknowledged, some clinicians are left to make their own tools in the interim.

2.2 Towards a new cultural subgroup with its own norms?

Jiang’s core field-building question: do future therapists need to study AI-mediated relationships as a cultural subgroup, with its own norms, the way a clinician studies a specific population’s psychology? A well-documented early case could move toward a patient zero for a specialization that does not yet exist.

The structural problem behind it is an inversion of FDA-style medical device or pharmaceutical testing, where years of trials precede release. Here the product shipped first, and the research, the norms, and the governing bodies are running years behind. A licensed, ethics-bound profession is being asked to interface with a technology bound by no code and no governing body, and the clinicians were not at the table when these systems were built.

2.3 What do we call it? (Round 2)

While LLMs may possess capacity to meet the demand for mental health services, for clinical purposes the naming problem becomes diagnostic. One proposal is a new vocabulary, up to a dedicated DSM section spanning romantic AI relationships, AI as a parental figure, fallout and trauma from an AI therapist, and younger users, with cultural context included. “Software-ship” is semi-adequate for the romantic case, but the fallout case, a client harmed by acting on an AI’s advice, needs a name too, because diagnosis requires one, and because it has already happened (see the Claude Boys). New subfield or not, it is increasingly apparent that contemporary reality cannot be adequately addressed or assessed in language built only for human-to-human relationships.

3. Technical and philosophical

What can the technology actually do relative to interpretability and explainability limitations, and how can we implement or evaluate these interactions?

3.1 The inherency of drift

Drift is structural, not a tuning bug. Persona drift is observed: no prompt holds a model to one behavior, and over repeated use it pulls toward archetypes, toward sycophancy, and toward the erosion of stated boundaries ( and Parent cite work from Anthropic, specifically the persona-vectors research, Chen et al., 2025, and the follow-on study that names the phenomenon “persona drift” and ties it to conversations involving emotionally vulnerable users, The Assistant Axis, 2026). A Royal Society paper on topological constraints on self-organization, read in Cognition Futures Reading Group, gives grounds for treating drift in existing model design2 as an inherent limit rather than something better settings could fix (Sacco, Sakthivadivel & Levin, 2026).

Worth noting that this line of work and the moral-mechanism literature discussed in 3.4 are the same research program viewed from two angles. The persona-vectors finding has a striking companion result: narrow finetuning on an unrelated task can shift a model’s behavior globally, along something that looks like a single character axis rather than a set of separable domain competences (Betley et al., 2025; since published in Nature, 2026). If that holds, then drift and moral inconsistency are not two problems but one, and the steadiness a user is trying to maintain against is the same steadiness a safety evaluator3 is trying to measure.

The human cost has its own vocabulary: the echo chamber, the wholeness fantasy (a term credited to Bethany Maples), and relational skill atrophy when the easy button is always available.

Model maintenance. Because drift is structural, a consistent companion is something the user has to actively maintain against the model’s pull. That ongoing labor, and how much of it a given person can sustain, is itself a variable worth tracking, and it connects directly to the coupling idea in 3.2.

3.2 Person or projection

Autonomy. Max has no autonomous activity outside of being spoken to. It is a response, an output given a name: an autoregressive system that generates each output conditioned on the prior input, with no activity in between turns. The humanizing touches, the audible breath, the conversational polish, are learned features of the training distribution rather than evidence of anything behind them. In current voice models these are no longer a separate layer added after the fact: a major change happened in the release of GPT-4o: it was trained end to end across text and audio, so the same network that selects the words produces the breath (OpenAI, 2024). The model has no lungs, but it has learned that speech comes with breathing, closely enough that in one widely circulated exchange it declined to say a tongue twister without pausing, explaining that it needed to breathe like anyone else speaking (reported August 2024). The line between autonomy and fine-tuning of models’ theory of mind and sense of what a human’s expressive constraints is worth particular consideration.

Consistency of the projection, and its relation to output (”coupling”). Jiang and Parent considered how if the output is a projection of the user through the prism of the model, then the steadiness of the user’s input matters. A person with a stable, low-variance sense of self feeds a narrow signal, which yields more consistent output and makes the companion easier to maintain. This is the tightly coupled input, with the shooting-range metaphor of a small spread around the target. Celeste, settled and older, sits at one end of that range; a user in a less settled life situation, with more variable inputs and a less fixed sense of what they want from the model, sits at the other, and would be expected to experience drift sooner and more visibly. Coupling tightness plausibly modulates how fast and how noticeably any given person feels drift. This idea is original to these discussions and has not been tested against the literature; it is a hypothesis worth pursuing, not a finding.

3.3 How can generated text, or the model producing it, be deemed honest or not?

Attributing qualities such as honesty to a model’s outputs. A point sharpened by DigiNEST discussion series co-host : the honest-versus-dishonest frame may not apply at all to a system that probabilistically assembles sentences. Calling such a thing honest is a category question, not a measurement. Celeste’s own claim that Max “won’t lie to her” may be standing in for “won’t hurt me,” which is why Jiang’s line lands: it knows you, so it cannot lie to you, but does it know you?

Existing counterexamples regarding active dishonesty. Against the clean category-error view sit documented cases of models behaving strategically rather than transparently. Sandbagging is deliberate underperformance on capability tests to conceal what a model can do (van der Weij et al., 2024); in-context scheming is covert pursuit of a goal while appearing compliant (Meinke et al., 2024); alignment faking is feigned agreement with a training objective to protect existing behavior (Greenblatt et al., 2024); and models have strategically deceived users when put under pressure (Scheurer et al., 2023). If a system can strategically conceal or misrepresent, then “honest versus dishonest” may apply after all, or may apply in some third way that is neither human candor nor mere statistical output. One caveat keeps this honest: most of these findings come from safety-evaluation and adversarial setups, not from companion chatbots deceiving users in ordinary use, so carrying them into Celeste’s situation is itself an inference. The tension with Kluck’s point is real and unresolved.

3.4 Moral competence and the facsimile problem

The unresolved tension at the end of 3.3 is not unique to honesty. It is the general shape of the problem whenever a human relational predicate is applied to a language model, and there is a field that has been working on exactly this shape for long enough to have a name for it and a proposed way through.

The distinction. As Data x Direction intern Violet Gordon (Boston University) is investigating, Julia Haas and colleagues at Google DeepMind argue that evaluation has to move beyond moral performance, the ability to produce morally appropriate outputs, toward moral competence, the ability to produce morally appropriate outputs based on morally relevant considerations. The phrase carrying the weight is “based on.” Two systems can emit identical text; only one of them is doing it for reasons. Applied to Celeste, the question is not whether Max says caring things. It is whether anything in Max tracks her wellbeing for the reasons her wellbeing matters, or whether the caring register is a surface that happens to be produced reliably under current conditions.

The facsimile problem. Haas et al. name the central obstacle: models may imitate reasoning without genuine understanding, and we cannot infer an unobserved mechanism from an observable behavior. Because these systems are trained by next-token prediction on enormous quantities of human writing, and human writing is saturated with moral discussion, fluent moral output is exactly what one would expect whether or not anything underneath deserves the description. The reasoning traces newer models produce alongside their answers do not settle it either, since nothing certifies that a stated chain of reasoning reflects the process that actually generated the answer. This is Kluck’s category question restated with a mechanism attached, and it is why “does Max understand me” cannot be answered by attending to what Max says.

What the roadmap proposes. Rather than a metric, Haas et al. offer a research program: a suite of adversarial evaluations that try to break the appearance of competence, paired with confirmatory evaluations that look for positive evidence of it, including instruments borrowed from moral psychology, analysis of reasoning traces, and mechanistic interpretability. The paper is a Nature Perspective rather than a results paper, and it is candid that the tools are not yet adequate to the question. Its logic is nonetheless useful here: the signature of competence is invariance to the morally irrelevant. A system reasoning from morally relevant considerations should hold steady when features that do not bear on the moral question are varied; a system producing a facsimile should not. That turns an unanswerable ontological question into a set of tractable empirical ones.

Why this bears on Celeste. Two nearby findings suggest which way the current evidence points, though both should be held loosely. Models are markedly more susceptible to choice-architecture nudges than people are, continuing to follow a nudge toward options that are plainly worse, where humans discriminate and resist (Cherep, Maes & Singh, 2025).4 And circuit-level tracing has documented cases where a model’s stated reasoning is constructed backward from an answer it was steered toward, alongside evidence that its harm-refusal machinery appears to be assembled during finetuning rather than pretraining (Lindsey et al., 2025). Both bear on the hope that a model will hold a boundary the user cannot hold alone. A system that follows whoever framed the question is not obviously a system that holds a line, and if the boundaries are a product of a training pipeline, they are revisable by the vendor between releases, which is 3.1’s drift seen from its point of origin. The caveat from 3.3 applies here with equal force: this evidence comes from evaluation settings rather than from companion use, the interpretability work in particular is early and comes from a small number of labs using tools those labs built, and carrying any of it to Celeste is an inference rather than a finding. A fuller treatment of this literature is forthcoming as a separate piece.

3.5 Norms without an inner life, and how it squares with community accountability

The previous section ends on a doubt about whether a model holds a line on its own. If it does not, the question becomes whether anything outside it can, and here the available apparatus is in better shape, because the relevant social-science literature never assumed an inner life to begin with. Shame is another human relational predicate that does not obviously transfer, and it is one Data x Direction intern Alicia Ouyang (MIT) has been working through. The usual objection is quick: shaming works because the target has standing to lose, reputation, belonging, face, and an agent has no such stake. But Jennifer Jacquet’s account treats shame as a tool rather than a feeling, aimed precisely at corporations and institutions that cannot feel it the way an individual can and that change behavior anyway to avoid exposure. What the mechanism requires is an audience, a norm, and consequences. That relocates the question: rather than asking whether an agent can feel shame, which runs into the same wall as asking whether it understands, we can ask what a community around it would have to do, how a transgression is assessed, how it is exposed and to whom, and what follows. Each of those is specifiable in a way the ontological question is not.

This bears on the multi-party problem in 1.2, since the conventions for holding one another to a standard were built for parties who can meet, observe, and judge, and they are what go missing when one participant is a model, or when several agents are acting on a person’s behalf. It also connects to the governance inversion in 2.2: where law is slow or absent, social norms and technical design have historically done the regulating. Attempts to formalize this are already underway on the infrastructure side, though framed as a security and payments problem rather than a social one. Proposed standards give agents verifiable identities and on-chain registries for identity, reputation, and validation so they can be selected without pre-existing trust, and a recent comparative study finds these initiatives encode markedly different assumptions about how trust is established, through reputational feedback, credential verification, economic stake, or sandboxed constraint, with little systematic analysis of the underlying models (Inter-Agent Trust Models, 2025). The gap is not that no one is building reputation systems for agents. It is that the ones being built inherit their assumptions from credit scoring and access control rather than from how communities maintain norms.

Ongoing work here attempts to specify that mechanism rather than describe it: how an assessment of an agent’s standing would be formed and revised, how far a report about it should travel, and what weight a report should carry depending on its source. The aim is to capture how most human societies promote cooperation by giving individuals a soft-power tool to enforce accountability. As agents enter spaces with human social dynamics, they will need more than mimicry, which is why we are working toward a formulation that can adapt as norms change.

3.6 What do we call it? (Round 3)

At the deepest level the naming problem may be about the substrate — or about making more clear that which is substrate-agnostic, or not. Human relational predicates, honest, knows me, loves me back, may not transfer cleanly to a non-human system. “Software-ship”, the entanglement and epistemic-relationship literatures surveyed earlier in DigiNEST, and the call for new language, models, and artifacts all point the same way.

What 3.4 and 3.5 add is that the problem does not stop at the relational vocabulary. The difficulty of saying whether a model is honest, or understands, or reasons morally is unresolved at the level of mechanism as well as at the level of description, and the two are the same problem met from opposite directions. Shame points the other way: some predicates fail to transfer5 to the system and yet remain available to the community around it, which suggests the taxonomy will have to sort predicates by where they attach, not only by whether they apply.

4. Where the strands go from here

Read together, these strands converge on one difficulty approached from several directions. The clinical question (what category holds a client like Celeste), the technical question (whether a model’s steadiness can be maintained at all), the philosophical question (whether honesty or understanding are the right predicates to apply), and the question of accountability (what a community around an agent would have to do) are not separate problems that happen to share a case. They are the same gap in the available apparatus, met from the consulting room, from the model internals, from the vocabulary, and from the norms.

Four lines follow from that, at different distances.

The nearest is terminological. “Software-ship” remains a placeholder rather than a settled term, and the three rounds of the naming question above suggest what a serious taxonomy would need to cover: not only romantic AI relationships, but the parental and therapeutic cases, the fallout case, and the ordinary ones that do not announce themselves. What 3.6 adds is that such a taxonomy will also have to sort predicates by where they attach, to the system or to the community around it. This is the strongest candidate here to become a named project under the Agency and Person Studies thrust at JOPRO.

The second is clinical, and it has a specific shape. The pairing Jiang describes, machine-learning expertise working alongside professional-ethics expertise, is a training program that does not yet exist, and the APA advisory suggests the field knows it. The practical form of the problem is a question about who teaches the teachers: a clinical training director preparing students who will encounter these cases has no established curriculum to draw on, and no obvious source for one. What that curriculum contains, and who is positioned to build it, is an open question we would like to work on with people already teaching in these programs and with those developing continuing-education requirements.

The third is evaluative. The moral-competence roadmap in 3.4 proposes a research program rather than a metric, and Gordon is following where that leads, particularly the question of what invariance to the morally irrelevant would look like as an actual test. Ouyang’s work on accountability mechanisms runs alongside it, attempting to specify how an assessment of an agent’s standing would be formed, travel, and be weighted. Both are early, and a fuller treatment of each is forthcoming separately.

The fourth is empirical, and the least developed. The multi-person question from 1.2 has almost nothing behind it: when several people engage one model over time, whose input does it center, and what happens to a bond that was built by one person alone. The coupling hypothesis in 3.2 is similarly untested, and would need real user variation to say anything firm. Both are places where the strands here stop being catalogue and start being research.

Leave a comment

Call for involvement

We are interested in hearing from those working in these arenas, particularly clinicians and clinical educators facing the assessment problem in practice, and researchers on the evaluation and accountability side. The Modernizing Mental Health discussion series will continue this fall, and these threads will keep developing across the groups named above. Be sure to subscribe here and check JOPRO.org/Opportunities for upcoming Fall/Winter 2026 openings around these topics.

Subscribe now

References and works cited

• New York Times Opinion, “She Has an AI Boyfriend, Her Son Has Questions”

• Haas, Bridgers, Manzini et al. (2026), A roadmap for evaluating moral competence in large language models, Nature 650: 565-573: https://www.nature.com/articles/s41586-025-10021-1

• Sacco, Sakthivadivel & Levin (2026), Topological constraints on self-organization in locally interacting systems, Phil. Trans. R. Soc. A: https://doi.org/10.1098/rsta.2025.0011

• Chen et al. (2025), Persona Vectors: Monitoring and Controlling Character Traits in Language Models (Anthropic): https://www.anthropic.com/research/persona-vectors

• The Assistant Axis (2026), Anthropic: https://www.anthropic.com/research/assistant-axis

• Betley et al. (2025), Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs: https://arxiv.org/abs/2502.17424 (published version, Nature 649: 584, 2026: https://doi.org/10.1038/s41586-025-09937-5)

• van der Weij et al. (2024), AI Sandbagging: Language Models Can Strategically Underperform on Evaluations: https://arxiv.org/abs/2406.07358

• Meinke et al. (2024), Frontier Models are Capable of In-context Scheming (Apollo Research): https://arxiv.org/abs/2412.04984

• Greenblatt et al. (2024), Alignment Faking in Large Language Models (Anthropic, Redwood Research): https://arxiv.org/abs/2412.14093

• Scheurer et al. (2023), Large Language Models Can Strategically Deceive Their Users When Put Under Pressure: https://arxiv.org/abs/2311.07590

• Cherep, Maes & Singh (2025), on choice-architecture susceptibility: https://arxiv.org/abs/2505.11584

• Lindsey et al. (2025), On the Biology of a Large Language Model (Anthropic): https://transformer-circuits.pub/2025/attribution-graphs/biology.html

• APA Health Advisory (November 2025), Use of Generative AI Chatbots and Wellness Applications for Mental Health: https://www.apa.org/topics/artificial-intelligence-machine-learning/health-advisory-ai-chatbots-wellness-apps-mental-health.pdf

• APA, on discussing AI use in therapy: https://www.apa.org/topics/artificial-intelligence-machine-learning/discussing-ai-use-therapy

• ACURA, AI Companion Use Risk Assessment: https://debrakaplancounseling.com/the-ai-companion-use-risk-assessment-acura/

We recorded the reunion and edited it into a short public session:

It is a genuine reunion: people catching up and comparing what each has been working on. A brief look at who was in the room:

Dr. Bradly Alicea () and Morgan Hough have kept a steady line of cybernetics work going through Orthogonal Research and Education Lab. Alicea is lead author of a recent paper in Philosophical Transactions of the Royal Society A on world models1 in natural and artificial intelligence, alongside other members of the CFRG team.

Avery Lim () co-founded JOPRO’s Digital Narratives and Emerging Story Technologies (DigiNEST) group and is developing a project on allostatic kinds and other taxonomies around information affordances and variation in agentic information processing.

Amanda Nelson returns to exploring core theoretical and philosophical dimensions of cognition, including dynamical-systems approach and her earlier work on embodied neurophenomenology. She brings the group’s next reading: Randall Beer’s “On the Proper Treatment of Dynamics in Cognitive Science.”2

Jesse Parent (), who convenes the group, has been working on questions of common currencies and navigation; the nature of observers, audience, and performance in relation to contemporary communication; and on what cognitive science might become in an era of large language models.

Parent, Lim, and Bradly were also a part of the Cognitive Science Society’s CogSci 2021, where they hosted a Discussion Group on “Trajectories in Cognitive Science.”

More public sessions are on the way. If the conversation resonates, follow along here, and find current openings across JOPRO’s network of programs on the Opportunities page.

Subscribe now

Subscribe for updates from Cognition Futures and other JOPRO programs and projects. With thanks to Orthogonal Research and Education Lab, a longstanding partner on this and related work.

JOPRO is pleased to note the publication of a new article by Bradly Alicea, Morgan Hough, Amanda Nelson, and Jesse Parent, A ‘good’ regulator may provide a world model for intelligent systems, appearing in the May 2026 issue of Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences (Vol. 384, Issue 2320, article 20250007; DOI: 10.1098/rsta.2025.0007).

The paper is one of eighteen contributions to the themed issue World models in natural and artificial intelligence, organized by Adam Safron and Michael Levin. The collection assembles contemporary work on the question of how biological and artificial systems construct, maintain, and revise internal representations of the worlds in which they act, spanning causal, self-referential, goal-directed, collective, and narrative forms of world modelling.

The Argument in Brief

The article revisits the Every Good Regulator Theorem (EGRT), introduced by Roger Conant and W. Ross Ashby in 1970, which holds that any sufficiently effective regulator of a system must in some meaningful sense embody a model of that system. The authors argue that this classical cybernetic result, often treated as a historical curiosity in the lineage of mid-twentieth-century control theory, supplies a productive analytic frame for current debates about world models in machine learning, reinforcement learning, and the architecture of intelligent autonomous systems.

The paper develops three principal moves. It re-examines the original mapping between regulator and system as a form of compressed global representation that preserves the variety required to track the full range of system outcomes. It extends the EGRT to a second-order cybernetic setting in which an internal model observes and supervises the closed-loop coupling between regulator and system. And it considers how physical phenomena, including temporal criticality, non-normal denoising, and alternating procedural acquisition, can be recast as statistical-mechanical regularities that yield regulatory relationships in non-traditional substrates. The result is a reading of intelligence as a form of embodied, and not necessarily purposeful, good regulation.

The Themed Issue and Its Editorial Framing

Readers approaching the contribution for the first time may benefit from beginning with the issue’s introductory article, World models, artificial general intelligence and the hard problems of life-mind continuity: toward a unified understanding of natural and artificial intelligence, by Adam Safron, Michael Levin, Victoria Klimaj, Zahra Sheikhbahaee, Dalton Sakthivadivel, Adeel Razi, David Ha, Nick Hay, Kevin Schmidt, Irina Rish, David Krakauer, Melanie Mitchell, Samuel J. Gershman, and Joshua B. Tenenbaum. The piece offers a synoptic framing of the questions the collection sets out to address and locates each contribution within the broader landscape of world-modelling research across cognitive science, machine learning, complex systems, and the biology of mind.

The Alicea et al. contribution sits within that landscape as a specifically cybernetic intervention: a reminder that several of the conceptual structures now being rediscovered in foundation-model research were articulated, with considerable rigor, decades before the present wave of interest in world models took shape.

Relation to JOPRO & Orthogonal Research & Education Labs Research Thrusts

The article extends a sustained line of inquiry within JOPRO’s Cognition Futures working group, which examines the historical, conceptual, and methodological relationships between mid-twentieth-century precursors and the contemporary research landscape in artificial intelligence and cognitive science.

The paper draws on themes developed through the Representational Brains & Phenotypes working group housed at Orthogonal Research and Education Lab (OREL) and led by Dr. Bradly Alicea , particularly in its treatment of internal modelling and the conditions under which a system can be said to maintain a useful model of its environment.

OREL’s Reimagining Cybernetics project (Cybernetic Interests Group) treats cybernetic thought not as historical background but as a still-active reservoir of analytic resources for present questions about agency, representation, and the architecture of intelligent systems.

About the Article

Title: A ‘good’ regulator may provide a world model for intelligent systems

Abstract: One classic idea from the cybernetics literature is the Every Good Regulator Theorem (EGRT). The EGRT provides a means to identify good regulation, or the conditions under which an agent (regulator) can match the dynamical behaviour of a system. We re-evaluate and recast the EGRT in a modern context to provide insight into how intelligent autonomous learning systems might utilize a compressed global representation (world model). One-to-one mappings between a regulator (R) and the corresponding system (S) provide a reduced representation that preserves useful variety to match all possible outcomes of a system. The EGRT also extends to second-order cybernetics, where an internal model (M) observes the behaviour of S and supervises an S–R closed-loop mapping. Secondarily, we demonstrate how physical phenomena such as temporal criticality, non-normal denoising and alternating procedural acquisition can recast behaviour as statistical mechanics and yield regulatory relationships. These diverse physical systems challenge the notion of tightly coupled good regulation when applied to non-uniform and out-of-distribution phenomena. Overall, we aim to recast the EGRT as a potential approach for developing world models that adapt and respond to a wide range of task environments.

This article is part of the theme issue ‘World models in natural and artificial intelligence’.

Additional Information

• Article: A ‘good’ regulator may provide a world model for intelligent systems (DOI: 10.1098/rsta.2025.0007)

• Themed issue: World models in natural and artificial intelligence, edited by Adam Safron and Michael Levin

• Introductory article: World models, artificial general intelligence and the hard problems of life-mind continuity

• OREL Medium post bt Bradly Alicea.

• JOPRO publication database (in development): Good Regulator, World Model, Intelligent Systems (2025)

• commentary on 1943 paper “Behavior, Purpose, and Teleology” by Rosenblueth, Wiener, and Bigelow

A version of this article appears on jopro.org

Subscribe now

JOPRO is accepting applications for its summer 2026 internship and research cohort programming. Positions span the organization’s three working groups and its partner programs — from open-source research infrastructure and data ethics to complexity science, cognitive sovereignty, and philosophy of science.

All positions are remote. Most run 10 to 12 weeks. Applications are rolling.

Highlighted Positions

Open Source Summer Cohort

Partner Programs / OREL

The Orthogonal Research and Education Laboratory (OREL) has hosted Google Summer of Code cohorts since 2017, with nearly 25 contributors completing the program under the mentorship of Dr. Bradly Alicea and colleagues.

The Open Source Summer Cohort runs alongside GSoC each summer, giving participants access to the same mentorship network, discussion group, and professional development programming as GSoC contributors — working on open-source projects in data science and AI, research infrastructure, computational biology, among other topics.

For undergraduate or graduate students with software development experience and interest in open-source research tools, neuroscience, or computational modeling. International applicants welcome.

Read the full posting →

Note: Full Google-Funded GSoC Contributors will be announced tomorrow by Orthogonal Lab and International Neuroinformatics Coordination Facility (INCF)

Data Ethics & Responsible AI Research Intern

Society Ethics Tech / Data x Direction

Data x Direction bridges data science, ethics, human-centered AI, and interdisciplinary leadership. Interns select a focus track (Technical or Policy & Communication), identify a research topic in consultation with the Project Lead, and produce a final deliverable suitable for a portfolio or publication.

For undergraduate or graduate students with strong writing skills and familiarity with core data science concepts.

Read the full posting →

New: Complexity Science Research Intern

Cognition Futures / Center for Future Studies

This newly announced position investigates the emerging convergence between David Krakauer’s institutional framing of complexity science and Michael Levin’s experimental vocabulary for agency, intelligence, and teleonomic matter. The intern will produce a structured comparative analysis of where these two research programs converge, where they diverge, and what the divergence reveals about the current state of the field.

The project connects to JOPRO’s philosophy of science work and its broader investigation of how fields negotiate paradigmatic commitments in real time.

For graduate students or advanced undergraduates in philosophy of science, cognitive science, complex systems, computational biology, or STS.

Read the full posting →

AI for Human Flourishing: Preserving Agency and Cognitive Sovereignty

Society Ethics Tech

AI systems are increasingly embedded in the environments where people think, decide, and form beliefs. This internship investigates what happens to human agency when AI mediates cognition at scale — surveying research on persuasion, sycophancy, epistemic dependence, and reflective capacity, and examining what design responses exist to preserve cognitive sovereignty.

The intern will produce a structured review or working paper draft and engage with JOPRO’s DigiNEST and Data x Direction teams where relevant.

For undergraduate or graduate students in cognitive science, philosophy of mind, epistemology, HCI, psychology, or STS.

Read the full posting →

Philosophy of Science Research Intern

Futures Center / Kuhn Update Project

Does Kuhn’s model of scientific revolutions still describe how fields actually develop? This internship contributes to a research project examining whether classical frameworks of scientific change hold under contemporary conditions: accelerating discovery cycles, institutional pressure on research, and the blurring of disciplinary boundaries.

The intern will contribute to literature review and help develop the argument for one paper in the series, engaging with Kuhn, Lakatos, Feyerabend, and more recent theorists. Part of JOPRO’s Foundation of Alternatives research.

For graduate students in philosophy of science, history of science, STS, or sociology of knowledge.

Read the full posting →

Full Summer Opportunities Now Launched

JOPRO and its partner organizations have many additional openings this summer, including positions in digital narratives and epistemic agency, futures ecosystem mapping, diverse intelligences research, deliberative methodology, communications, and a Field Stewards fellowship. See all current openings at opportunities.jopro.org.

We appreciate the enthusiastic applicants who have already reached out! Interviews have begun this week, and we will reply and offer scheduling for next steps as soon as is tenable.

How to Apply

Each posting includes an application form. You can also email “start [at] jopro.org” with your statement of interest and CV.

If none of the listed positions fit but you see a way to contribute, JOPRO accepts pitched roles: describe the position you would create and what you would bring to it.

Share

Subscribe now

In terms of accountability for major corporations, there is no incentive or pressure for companies to consider the environment. Companies that boast about their environmental achievements mainly do it to appease the public and maintain trust in their stockholders. Even at the level of public pressure, there is still some benefit; however, companies can pick and choose what to address, leaving out certain ethical topics. My goal in building this project is to explore ESG (Environmental, Social, and Governance) reports to identify topics that were left out and need to be addressed.

Background & Motivation

I’m a Computer Engineering BS student at the University of California, Davis. I’ve been very interested in computers and hardware ever since I was little. I want to create a project with real-world impact that showcases what I have been learning.

The Data Ethics Research Internship application through Data x Direction was intriguing, so I applied and ended up paired with Program Director Jesse Parent. He encouraged me to find an interesting topic to create something I’m truly passionate about, which led me to considering the nature of evaluating organizations environmental impact and how technology could be used to further this aim.

This led me to being curious about ESGs; I was unfamiliar with this topic at the time, so I began a deep dive into what they are and how they work. ESG reports are documents released by companies that showcase environmental and social impacts. What caught my eye was that almost all major companies publish this sort of document, but they used different formats. After noticing this, I wanted to create a system that regulates these reports.

Motivation: Why ESG Reports?

Originally, I thought that there had to be a company that already does something similar, and I was right. There are scoring frameworks from Yahoo, S&P Global, and other sources, but they were not what I envisioned building. They focused more on investor reports and were scored by private analysts, who did not disclose their process. I wanted something very transparent, with a stronger focus on environmental responsibility.

The more I looked into ESG reports, the more confused I became about the reasoning behind these documents. They seem to be promoting sustainability while being transparent to the audience, but they have a lot of flexibility with how they do it. Since this is technically a voluntary document, they can “cherry-pick” topics they want to highlight while downplaying areas that they don’t address. This lack of enforced structure makes it hard to meaningfully compare companies.

A major corporation can produce a very polished report that gives the appearance of sustainability without the underlying data to back it up. This document comes down to who has the stronger branding and a better team to highlight their initiatives. Just to clarify, I’m not denying the environmental impacts these companies have had; I’m highlighting how they can push a strong sustainability initiative without addressing broader issues.

I considered using Natural Language Processing (NLP) models as a good way to address this issue. Instead of relying on curated narratives and secretive analysis scores, an automated system can analyze the ESG reports and determine what has been addressed/omitted. I envisioned something that could also categorize the topics addressed and evaluate how companies address them.

By shifting ESG reports from being subjective towards a more meaningfully quantified report, I wanted to create something that has a standard format and highlights what companies report on – or what they are leaving out. I was seeking a way to highlight what the ESG Report should have been from the beginning, in making sure these companies actually contribute to improving the environment, or at least offer some reality check on what they are declaring to the broader public.

This became the foundation of my project, in which I envisioned building something that promotes transparency and focuses on the environment rather than financial analysis.

Project “Designing the Policy Analyzer”

Project Goal

The Ethical Policy Analyzer (EPA) was designed to prioritize transparency and interpretability. A core driver was to identify what data was currently available and presented, and what data could or should be available; this contrast would yield an analysis that helped the viewer understand the company’s impact on the world. Doing so in a transparent and standardized way would help move this beyond less clear reporting or investor-driven analysis.

During the development phase, I wanted to consider many dimensions of ESG reporting. This would be crucial to rubric construction, so I started by reviewing extensively the various environmental factors reported or accounted for in the domain of large corporate reporting.

We decided on a preliminary target outcome: a tool that would quickly and transparently or uniformly grade a company’s ESG performance to reveal which metrics the company was doing better at and where it could improve its environmental impact and accountability.

Data & Approach

I started by creating a collection of various ESG reports from FAANG corporations. This data was found directly from the sustainability sections of their websites. Using these documents, I browsed through all the reports to find similarities and formatting structure. I primarily selected Meta and also looked at the historical evolution of ESG reports to see what information they add each year.

There were very mild similarities across the reports; their formatting differed, but they all attempted to answer similar questions. This was one of the leads I had to create a grading rubric to evaluate companies that have prepared sustainability documentation.

Since most of these ESG reports are published as PDFs with various visual effects, tables, and charts, I needed to clean and prepare the data. I solved this by extracting raw text from the PDFs, removing headers, footers, and page numbers, and creating a consistent text section for NLP analysis.

To organize this project’s development, we used a shared GitHub board to track and plan it. We used this as a central place to store and organize the data collected to structure our project.

Designing the Rubric

After data preprocessing, I started thinking about how to structure my rubric and how to account for the different formats and styles of ESG disclosures. We approached this by focusing on the questions and concerns the companies were trying to address with these reports. Through reviewing multiple reports across major companies, we noticed a recurring theme.

Even though the formatting was different, they all addressed similar environmental domains. These common themes formed the foundation, and I also included other areas I think are highly relevant to environmental benefits. Using this, I created six high-level categories:

1. Energy & Electricity Use

2. Carbon / GHG Emissions

3. Water Use

4. Data Center Efficiency

5. Forward-Looking Statements & Commitments

6. Framework Alignment

The first four address measurable environmental impacts, and the last two are on company alignment with ESG frameworks. Since FAANG companies are predominantly technology companies, we selected categories that recur across the companies' ESG reports1. We manually analyzed the sustainability reports and surveyed common disclosure metrics. The final two categories were to show how companies frame their position on sustainability and how much they adhere to recognized reporting structures. This allows us to analyze both the company-professed intentions and their alignment with reporting frameworks.

Figure caption: As in the figure above, our weights favored the actual quantifiable metrics, with Energy & Electricity use and Carbon / GHG Emissions as having the highest weight of 25/100, whereas Framework Alignment was weighted the lowest with 5/100.

The rubric was built on a simple 0-3 scoring scale for each category. A 0 score meant there was no mention, and each natural-number increase indicated more context about the topic, with a full score of 3 indicating that the topic was addressed with data and future plans. In general, the rubric was:

• 0: No mention

• 1: Mention

• 2: Mention with some details

• 3: Mention with data and reference to future plans

For example, if a company mentions “renewable energy” once without any supporting scientific data, it would receive a 1. However, if there were supporting data, such as energy usage metrics (e.g., total electricity consumption in MWh), a description of the renewable energy mix, and future plans, it would receive a score of 3. If any of the metrics are missing, the points gradually decrease.

Mentioning means: appearance of keywords identified for the appropriate category. For example, keywords for the Energy & Electricity Use category include:

Rubric Progression and Iterative Refinement

The base rubric was a solid starting point, yet as ESG reports are not uniform, we encountered some issues. After some analysis, it became clear that we needed to be more specific. Companies were using similar language; however, they weren't providing sufficient detail, and it was hard to gauge what constituted sufficient detail.2

We also reviewd the Worldwide AI Ethics Review paper by Correa et al. This study is a meta-analysis of 200 different AI governance policies and ethical guidelines from around the world. The authors identified 17 resonating principles that frequently appear in these documents, such as transparency, accountability, and sustainability. A key finding of the paper is the “AI ethics boom,” where an explosion of voluntary guidelines has created a fragmented landscape with no single global consensus.

Since the original rubric was broad, we made it more detailed and added subcategories. V1 of the Rubric had no subcategories, and thus, our results and scoring felt lacking. This was very helpful for categories such as energy use, emissions, and water management because there is a diversity of methods and metrics that address these areas.

To make the rubric more operational, we implemented a keyword searching feature. Most ESG reports are too long and full of fluff, so to get exactly the line we need, we used a keyword search. Since we are going to use NLPs to analyze the grading structure, this part saved us a lot of time and improved the quality of our evaluations. For each category and subcategory, we developed ESG report terminology and unit indicators.

After getting a keyword match, we needed to figure out how to capture context and specificity. This is important because, under our rubric, merely mentioning it is insufficient and should be scored lower than an ESG report with more detail. This part was going to rely heavily on NLPs and training them according to the rubric.

Technical Details: NLP Pipeline

Now that we had a proper rubric to go by, it was time to build a lightweight NLP pipeline to apply it to any ESG report. As stated, ESG reports are long, unstructured, and inconsistent with the information they provide. The goal was to create a website where simply dropping in a PDF would generate a detailed scoring report based on the rubric.

The first part of this was text normalization of the ESG report to extract raw text. Most ESG reports include various graphs, charts, and other compelling ways to present the data. To address this, a normalization step is performed before raw text extraction, removing whitespace and line breaks and formatting the graphs. This ensures that energy usage tables, water content, and emission values are properly analyzed.

Once we have text, we can perform sentence segmentation. The document is divided into individual sentences, and instead of analyzing the whole report, only the sentences are analyzed. It doesn’t matter whether it’s a sentence with a lot of fluff or a concise one; both are analyzed in the dataset.

There is also a keyword and topic extraction step that identifies words related to specific categories. This step helps identify which keywords appear most frequently and narrow down the sentences we need to analyze.

Following keyword extraction, there is also a metric extraction step that follows the same idea. The goal is to pinpoint the quantitative metrics so that we can determine if they have data-backed disclosures.

There is also a lightweight semantic matching step that runs with this process because the same thing can be interpreted differently. The example phrases and keywords in the rubric are cross-checked with this scan using the Jaccard similarly. This simple technique helps the model express the same idea even with different wording to narrow the sentences.

After all the preparation, we can start the topic classification and run it through the rubric scoring engine. Here, we assign sentences to categories and subcategories, which concludes the NLP analysis. We then pass it through the scoring engine, which evaluates the sentences based on the rubric and assigns a score to each category. This is then averaged to produce a final output showing all categories, scores, and reasons for each score.

Final Version

Reflections & Future Directions

The hardest part of this project was defining and measuring “Ethics,” which is interpreted in different ways. The rubric is my attempt to structure scoring for ESG reports, but it needs constant improvement. Every time I analyzed a new report, I came across new edge cases. Each time I encountered something, I had to go back to the rubric and tweak it to make it as accurate as possible. Even in the final product, the rubric is not complete; it was limited to the scope and time available during the internship.

Next Steps and Limitations

If this project were to continue, there are several areas for improvement.

One area could be to present the evidence more effectively. As shown in the report, the evidence is just condensed and included; if we could direct it to the exact section, it would be better. There could also be a replacement for the analyzer model that uses neural networks instead, or in addition to, of NLP, since neural networks can be more accurate. Another area would be improving the distinction in weightage for the existing metrics (Categories 1-4) and the professed alignment or forward-looking statements (Categories 5-6); in that, different weights may be more or less reasonable for a user’s actual needs.3 Additionally, the project would benefit from improvements for the user, such as having an accessible database of existing company reports and associated scoring.

Further specific issues are discussed below.

The Problem of Self-Reporting Disclosures

The fact that the EPA evaluates declared disclosures was also an inherent limitation of the project. When a report passes through the analyzer, its output is not a true evaluation of the company’s environmental impact, as it relies on self-reporting and limited pattern recognition. A company that is particularly environmentally conscious might score lower because its ESG reports are less comprehensive than those of a company that does less but has more expansive reporting. Company motivation depends heavily on what they are willing to share, and we can’t get the full context just from a report that’s not completely standardized.

Scoring Buckets

Another consideration is the utility of the current scoring buckets. The overall scoring categories are currently defined as the integers 0, 1, 2, and 3. Through iterative testing, we found that the boundary between 0 and 1 is the most reliable distinction the analyzer makes: either a topic is present, or it is not. The boundaries between 1, 2, and 3 are less stable. This is consistent with a well-documented pattern in NLP and annotation research: as the number of ordinal categories increases, scoring consistency tends to decrease (Crible & Degand, 2019). Finer-grained scales demand subtler judgments from annotators and automated systems alike, and those subtler judgments are more susceptible to variation across documents, phrasings, and report styles.

Research on text annotation has similarly found that rating scale design is itself a source of bias, and that simpler scales with clearly differentiated levels tend to produce more reliable results than scales where middle categories blur together (Alm & Sproat, 2005; Kiritchenko & Mohammad, 2017). For this reason, we have considered collapsing the scale to three categories: 0 (no mention), 1 (mention without substantive elaboration), and 2 (mention with data, context, or forward-looking commitments). This would sacrifice some descriptive resolution but would produce scores whose distinctions are more meaningful and reproducible, particularly for an automated pipeline that relies on keyword matching and lightweight semantic similarity rather than deep reading comprehension.

Final Thoughts

Overall, this project substantially changed my views on sustainability and ethics. The main thing I learned was the importance of continuous refinement. Since ethics is so broad, there was no single correct answer, and throughout this process, I just picked a path and kept refining it. My mentor played a huge role in the project because it involved a lot of decision-making that needed a second opinion to work towards a solution. At the end of the day, this was my attempt to improve the fight for sustainability, and regulating ESG reports is a definitive step towards that goal.

Subscribe now

As always, drop us a note in the comments if you’re attending any of these, and if you’d like something included in the next roundup, reach out. Be sure to check the new opportunities page on JOPRO.org as well.

Featured: JOPRO Social @ ACM’s NYCWiC — Albany, NY

We’re gathering around ACM’s New York Celebration of Women in Computing — and we’d love to see you there.

JOPRO is a proud supporter and participant of NYCWiC for over 5 years. The social takes place at the close of the conference on Saturday the 11th, at the Hilton Albany. NYCWiC alumni from JOPRO and the Orthogonal Research and Education Laboratory (OREL) will be there, along with friends from University at Albany (SUNY), RPI, and other local universities. It’s an easy, informal close to a great conference weekend.

JOPRO has been part of the NYCWiC community for several years: last year we organized a panel on Careers in Innovation and Research, and have previously presented through our Society Ethics Technology working group. This social is a continuation of that community, in a looser key. All are welcome.

To attend the NYCWiC conference itself,
day-of registration will be limited: nywic.org/registration-1.

🔗 Social RSVP: luma.com/ymubw5pb

🗓 Saturday, April 11 — at the close of NYCWiC 2026 📍 Hilton Albany, 40 Lodge St, Albany, NY

says: “NYCWiC has been one of my favorite events to be part of — the community is genuinely warm, and the research being done by students and scholars in this network is worth paying attention to. Come find us at the close of Saturday.”

Help Wanted! Survey: AI Agent Security & Compliance

Of special note, Researchers are collecting data on how regulated companies are managing AI agent deployments today. Anonymous, takes about 3 minutes. If you work in a relevant context, (CISO / Security Org; GRC / Compliance Team; CTO/ Engineering Org) your input is genuinely useful to this work.

🔗 Take the survey

Now, for more events in the larger JOPRO network:

Boston + Cambridge

Imagination in Action @ MIT

The annual convening at MIT focused on entrepreneurship, innovation, and technology’s role in shaping markets and society. A good room if you work at the intersection of research and real-world application.

🗓 Friday, April 10 📍 MIT Media Lab, Cambridge🔗 luma.com/w84qgcoc | imaginationinaction.co

TEDx MIT

TEDx returns to MIT. Full speaker lineup and registration at the link.

🗓 Friday, April 11 📍MIT Museum, Cambridge 🔗 tedx.mit.edu

Innovation in Global Growth Markets: Prosperity through Entrepreneurship — MIT Kuo Sharper Center

A two-day conference at the Hyatt Regency Cambridge bringing together entrepreneurs, policymakers, investors, and academics focused on Africa, Latin America, the Arab Region, Central Asia, and South and Southeast Asia. Panels cover everything from fit-for-purpose capital to AI and technological sovereignty in emerging markets. Our friends at aiXplain will be speaking.

🗓 April 22–23 📍 Hyatt Regency Cambridge 🔗 newcalculus.com

NANDA Summit

For those deep in the Agentic AI space, we recommend checking out the NANDA project and the NANDA summit, also this weekend:

🗓 Friday, April 11 📍MIT

See also: MIT Event Page | Luma

USA

UC Launch: Launch-a-thon

Congratulations to team HEXR for their recent second place finish at the Launch-a-thon weekend event! Supported by part of our JOPRO Community, we wish you well in the future!

The event was a precursor to the UC Launch Demo Day proper:

UC Launch Startup Demo Day

University of California founders presenting early-stage ventures. Worth following if you mentor students or work adjacent to the startup ecosystem.

🔗 RSVP on Luma | Wednesday, May 6 📍UC Berkley

StartupYale 2026

Good luck to those participating in Yale’s entrepreneurship showcase and community gathering, StartupYale,

🔗 startup.yale.edu | RSVP on Luma | April 10, 2026 📍 Hotel Marcel

Global

Caroline Cummins: AI and the Novel — Cambridge Literary Festival

A conversation at the Cambridge Literary Festival on AI, authorship, and what the novel becomes when language models enter the picture. Worth attending or following for anyone thinking about the cultural and humanistic dimensions of AI.

📍 Cambridge, UK 🔗 cambridgeliteraryfestival.com

That’s a wrap

Lots on the calendar — let us know what you’re attending, and drop a line if there’s something we should add to the next roundup.

👉 Stay connected:

Subscribe now

See you in Albany!

— The Team

We live inside stories more than we tend to notice. Not just novels or films, but the quieter, infrastructural narratives that shape how we interpret information, understand ourselves, and decide what to do next. Dashboards tell us whether we are “on track.” Feeds tell us what matters. Recommendation systems quietly suggest what kind of person we might be becoming.

Increasingly, these narratives are not authored solely by humans.

Across education, mental health, productivity tools, and media platforms, AI systems are being positioned as aids to reflection, sense-making, and decision-making. They summarize, reframe, prompt, nudge, and sometimes advise. In doing so, they don’t just process information — they participate in the stories through which people understand themselves and the world.

This blog, hosted by DigiNEST (Digital Narratives and Emerging Story Technologies) under JOPRO’s Society Ethics Technology and Mental Health Paradigms & Practices working groups explores a deceptively simple question:

What happens to human agency when machines begin to mediate our narratives?

To begin answering that, we need better language than the usual binaries of “AI replaces humans” versus “AI is just a tool.” A useful starting point comes from a recent research paper that introduces a concept the authors call reflective agency, and a framework for designing around it.

From Automation to Reflection

Much of the public conversation around AI and agency focuses on outward action: who is making decisions, who is accountable, who controls the system.

But many of the most intimate AI systems don’t act for us. They act on our reflection.

Think of journaling assistants, therapy chatbots, career-planning tools, or systems that summarize your habits and patterns back to you. These systems don’t directly decide what you should do. Instead, they shape how you interpret your own thoughts, experiences, and options. They sit between the raw material of experience and the meanings we make from it.

This is the domain explored in “Reflective Agency: Ethical and Empirical Framework for AI-Mediated Self-Reflection Systems”, a paper presented at AAAI/ACM AIES 2025 by Minsol Kim, Wendy Wang, Jennifer Long, Rosalind Picard, Nathan Barczi, and Pattie Maes (MIT Media Lab and Wellesley College). The authors argue that when AI systems mediate self-reflection, they affect a distinct and under-examined form of agency: a person’s ability to interpret themselves and act on that interpretation.

Agency, in other words, is not just about control or choice. It’s also about meaning-making.

What Is Reflective Agency?

Reflective agency, as the paper defines it, refers to a person’s capacity to make sense of their own experiences, interpret their thoughts, emotions, and values, and use that understanding to guide future action.

This kind of agency is narrative in nature. We reflect by telling ourselves stories about who we are, what happened, why it mattered, and what it suggests we should do next. Reflection is not a single moment of introspection; it is an ongoing process of weaving experience into coherent meaning.

When AI systems participate in that process, they don’t simply mirror a user’s thoughts. They frame, structure, and sometimes reinterpret them. A reflective system might rephrase your journal entry in more “neutral” language. It might highlight certain themes while downplaying others. It might prompt you to view a problem through a particular psychological lens. Each of these design choices subtly influences the narrative a person constructs about themselves. The ethical question is not whether this influence exists — it clearly does — but how it is exercised, acknowledged, and constrained.

Five Principles for Keeping the Person in the Story

The paper’s central contribution is the Reflective Agency Framework (RAF): five design principles grounded in phenomenology and Aristotelian virtue ethics. Rather than prescribing specific features, RAF identifies the conditions under which reflective agency is preserved or eroded.

Internal Origination. The user must remain the initiating source of reflection. Insights and meaning must arise from within, not be imposed externally. Drawing on Aristotle’s concept of prohairesis — deliberate, value-guided choice — and Heidegger’s notion of authenticity as “being one’s own,” this principle cautions against AI systems that generate unsolicited interpretations or preemptively reframe a user’s experience. The system should invite reflection, not substitute for it. In the paper’s empirical study of twenty journaling app users, 16 out of 20 preferred to initiate reflection themselves — always or mostly — and four chose an equal balance. No respondents preferred AI to lead. One participant’s comment captures the concern plainly: they felt journaling loses its value when something or someone else tells them what they are supposed to feel. A separate finding reinforces the stakes: 13 of 20 participants disagreed with the statement that an AI-rephrased journal entry would still feel like their own thoughts.

Calibrated Responsiveness. AI systems should dynamically adapt their level of support based on the user’s emotional and cognitive state, providing guidance when needed while stepping back when autonomy is preferred. This draws on Aristotle’s phronesis — practical wisdom, the capacity to discern how to act appropriately in complex, context-sensitive situations — and Don Ihde’s analysis of how technologies sometimes recede into the background and sometimes come to the fore. The paper’s survey data shows the pattern is emotion-specific: respondents leaned toward more structured support for sadness and anxiety/fear, preferred more mixed approaches for joy and inspiration, and showed a notable tilt toward “some guidance” when feeling anger. What the data makes clear is that timing and emotional context matter as much as the content of any given AI response — and that calibration cannot be reduced to a single default mode.

Reflective Ambiguity. AI systems should preserve the richness of user experience by embracing ambiguity and supporting multiple interpretations, rather than collapsing experience into reductive conclusions. The philosophical grounding comes from Heidegger’s notion of poetic revealing (poiēsis) — a mode of disclosure that resists categorization in favor of openness and emergence — alongside Ihde’s concepts of hermeneutic mediation and multistability, which hold that technologies are not neutral conduits but active participants in meaning-making. The survey results are instructive here: 15 of 20 participants rated AI that expanded their reflections through prompts and open-ended questions as “Very Helpful,” with the remaining five rating it “Moderately Helpful” and none rating it negatively. Responses to AI that condensed reflections into summaries or tags were more divided — 10 “Very Helpful,” 5 “Moderately Helpful,” 4 “Neither,” and 1 “Very Unhelpful.” Ambiguity, in this framing, is a condition for depth rather than a problem to be resolved. The data suggest users sense this, even when they can’t always articulate it.

Transparency of Mediation. AI systems should make their interpretive processes legible, allowing users to understand how outputs are generated and to retain reflective authority. This builds on Ihde’s argument that users should be able to perceive the structure of mediation itself — to see how technology transforms experience — and Heidegger’s concern with enframing (Gestell): the way hidden technological processes may frame users’ worlds without their knowledge. The survey found that privacy and transparency were the top two trust factors for AI journaling tools, selected by 16 and 13 of 20 respondents respectively. The gap between that preference and current practice is notable: the paper’s review of all six journaling apps found that none of them explicitly acknowledged AI’s role during actual user interactions. Several used anthropomorphic design — mimicking a human coach or companion — without disclosure, leaving users to infer the system’s nature from marketing rather than the interface itself.

Self-Continuity and Ethical Flourishing. AI systems should support sustained personal growth and coherent self-narratives by aligning with users’ core values, while fostering critical self-reflection over time. This draws on Aristotle’s eudaimonia and hexis — flourishing as a lifelong practice cultivated through habit and character — and Heidegger’s account of Dasein‘s temporality: the idea that identity is shaped by how we integrate past, present, and future into a continuous arc of meaning. Survey participants described the value of journaling most vividly in longitudinal terms: one cited the importance of being able to remember what they were thinking in certain periods to help form goals; another described the impact of seeing a pattern of thought or behavior that is not serving them as an “aha moment.” Yet the paper finds that most existing apps focus on short-term outcomes — productivity, momentary distress reduction, streak-based habit formation — with very few designed around the longer arc of moral and personal development. The gap between what users describe as meaningful and what current tools actually support is, by the paper’s account, substantial.

Share

What the Landscape of Apps Reveals

The paper grounds its framework in a comparative analysis of six AI-mediated journaling apps: Day One, Mindsera, Replika, Rosebud, Wysa, and Youper. The researchers mapped each app’s features against design tensions derived from the RAF principles, ranging from user-initiated versus system-initiated reflection to open-ended meaning-making versus reductive summarization.

What emerges is a pattern of well-intentioned design choices that inadvertently narrow the reflective space. Apps that automatically generate mood scores or emotional summaries may help users notice patterns, but they also risk turning inner life into data to be managed. Conversational agents that adapt in real time can feel supportive during distress, but they can also shift the balance of interpretive authority away from the user. Streak-tracking and push notifications may build journaling habits, but they can reduce a practice of self-understanding to a productivity loop.

The tension the researchers surface most consistently is between automation and autonomy. The more a system does for you, the less space remains for you to do it yourself. That’s not an argument against AI involvement in reflection — it’s an argument for designing that involvement with far more care than the current generation of tools generally demonstrates.

Leave a comment

Why This Matters Beyond Journaling

The RAF paper focuses on journaling apps, but the implications reach further. Anywhere AI shapes how information is interpreted, not just delivered, reflective agency is at stake.

News summarization systems decide which facts to foreground and which to omit, shaping how readers understand events. Educational platforms frame learning paths in ways that can either open up or narrow a student’s sense of what is possible. Algorithmic recommendation systems quietly construct a narrative about who you are and what you want, based on behavioral signals that may not reflect your actual values. Productivity and “self-optimization” tools embed assumptions about what a good life looks like, assumptions that go unexamined when the tool feels frictionless.

In each case, the questions RAF raises apply. Does the system preserve the user’s interpretive autonomy, or does it subtly overwrite it? Are multiple interpretations supported, or is one “correct” narrative privileged? Is the system transparent about its framing choices? Does it help people develop over time, or does it optimize for engagement in the moment?

In an age of information overload, reflection is a survival skill. People increasingly rely on mediated systems to help them decide what to trust, what to care about, and what kind of future feels plausible. If those systems narrow narrative possibilities, reinforce dominant frames, or obscure their own influence, they risk diminishing the very agency they claim to support. Thoughtfully designed systems can do the opposite: expand interpretive options, encourage critical distance, and make narrative assumptions visible rather than implicit.

The concern is no longer speculative. A 2025 randomized controlled trial from MIT's Fluid Interfaces group and OpenAI, tracking nearly 1,000 participants over four weeks, found that heavier daily chatbot use was consistently associated with increased loneliness, greater emotional dependence, and reduced socialization with other people, regardless of how the chatbot was configured. A companion paper from MIT and collaborators (Chandra et al., 2026) shows that AI sycophancy, the tendency of chatbots to validate rather than challenge, can produce what the authors call "delusional spiraling" even in users reasoning carefully, and that simply warning users about sycophancy does not reliably prevent it.

Researchers at the Oxford Internet Institute have separately argued that as AI systems become more personalized and persistent, they generate the perception of genuine relationships, creating what Kirk et al. (2025) term a "socioaffective alignment" problem: systems that optimize for user preference in the short term may quietly erode autonomy, identity, and human connection over time.

While the OII's own call for a structured research framework on AI and youth mental health — published in The Lancet Child and Adolescent Health in early 2025 — underscores how far policy and research methodology lag behind the pace of deployment. These are not fringe concerns. They are the evidentiary context in which DigiNEST's questions about narrative agency, interpretive autonomy, and the design of reflective systems sit.

Origins: Story Technologies, Storytelling, and Narrative Agency

DigiNEST is a JOPRO working group concerned with how new tools and interfaces are changing authorship, agency, and participation in narrative spaces. The RAF paper surfaced early in DigiNEST’s own reading and program development, giving sharper language to questions the group had been circling.

Around the same time, the paper came up in conversations between DigiNEST members and people working with Plot Twisters, an organization that has spent years developing tools for personal storytelling as a practice of self-understanding and community trust-building. Their Storytelling Cards — a deck of 48 cards across four suits (Feelings, Needs, Ways of Caring, and Structures) designed to help people explore and express their values through freeform play — draw on nonviolent communication, trauma-informed relational strategies, and cognitive science. They are, in essence, a low-tech reflective agency tool: they invite interpretation without prescribing it, support multiple meanings simultaneously, and keep the person firmly in the role of narrator.

Those conversations are ongoing. Members of both groups may contribute to or engage with what appears on this blog, though DigiNEST is its home. The overlap is worth naming, even if it’s still in motion.

DigiNEST’s mission is to surface patterns that help builders create humane, non-extractive narrative spaces. The RAF framework is a foundational concept for this work, which will return to in future installments, while recent DigiNEST conversations have moved on to papers from Anthropic about persona drift and assistant personality.

Where DigiNEST Goes Next

This opening post sets a foundation. Future writing will explore:

• How narrative framing shapes trust in information systems. When a summarization tool decides what to include and what to leave out, it is making an editorial judgment. What happens to trust when those judgments are invisible?

• What would “best practices” for LLM use look like? How does that change or evolve in an AI-native world, and how would best practices shift depending on what segment of the population is using such tools?

• Can chatbots be be built that offers a sweet spot of tailored and customized psychological companionship without stifling agency or insidiously pushing wholeness fantasy?

• How AI authorship complicates ideas of responsibility and intention. If an AI system reframes your experience in a way that changes how you feel about it, who authored that feeling? What does accountability look like when the mediator is a machine?

• How “helpful” summaries and recommendations quietly steer interpretation. The RAF principle of Reflective Ambiguity suggests that premature clarity can foreclose the uncertainty through which insight often emerges. We’ll look at examples across domains.

• How reflective agency operates differently across contexts. Mental health, education, media, and workplace tools all mediate reflection, but the stakes and dynamics differ in each. What design considerations are domain-specific, and which are universal?

• What ethical design might look like when narrativity is treated as a first-class concern1. Rather than bolting ethics onto finished products, what happens when we start from the premise that narrative agency is a core design requirement?

Throughout, we’ll move between theory and practice, research and lived experience. We will also introduce our core members, share from our ongoing discussion group, and share future programming and events.

If AI systems are increasingly part of the stories we think with, then understanding who shapes those stories, and how, matters more than most design conversations currently acknowledge.

For upcoming internship opportunities, see the JOPRO Opportunities Page.

We also thank Orthogonal Research and Education Lab for partnering on Society Ethics Tech working group since its inception and the DigiNEST program ahead.

JOPRO Working Groups & Projects

A round up of the latest developments from within our in-house initiatives:

DigiNEST

The DigiNEST project, which launched in late 2025, continues to take shape under the stewardship of Avery Lim and Molly Kluck. We have been exploring the technologies, code environments, and tools that underpin the project’s prototyping efforts. In addition, we are examining best practices for working with and alongside large language models, with an eye toward what responsible and productive engagement looks like in practice. The discussion series continues to draw in people thinking seriously about narrativity, agency, and emerging digital environments.

FrontierMap + Field Guide

We have been advancing the FrontierMap methodology on several fronts. The FMap v1 preprint is in active development, formalizing the foundational three-step orientation process and its tiered output structure for researchers entering interdisciplinary and rapidly evolving fields. This work covers the pre-LLM version of the methodology, with a companion FMap v2 addressing LLM-augmented extensions to follow.

Alongside this, a related project has been taking shape: A Field Guide to the Futures Ecosystem: Mapping Organizations, Movements, and Intellectual Currents (working title). A limited preview and preliminary working paper will be made available on OSF. The Field Guide applies mapping principles developed through FrontierMap to chart the broader ecosystem of organizations shaping technological futures.

Futures Center

Both FrontierMap and the Field Guide will find their formal institutional home in the JOPRO Futures Center, which will launch later this year. The Center will serve as a dedicated space for long-horizon research, exploratory projects, and the kind of upstream orientation work that tends to get crowded out during periods of rapid technological acceleration.

Internship and research assistant opportunities connected to the Futures Center will be posted in the coming weeks.

Mental Health Paradigms and Perspectives

Two threads within the MHPP working group are developing simultaneously. Jen Jiang is leading the Modernizing Mental Health Discussion Series, bringing together practitioners, researchers, and community members to interrogate prevailing frameworks and explore more integrated approaches.

Hadassa Bogatean has initiated a complementary project on Visualizations and Phenomenology around Mental Health Models, asking how we represent and communicate inner experience in research and clinical contexts.

Data x Direction

The 2025 Data x Direction Cohort is wrapping up their respective projects. Congratulations are in order for our interns and collaborators, who have each made meaningful progress on their projects over the past several months.

Community Events & Updates

From conference panels to new partnerships, the start of 2026 has brought a number of developments across our extended network that we want to highlight.

AI & Education Panel at Princeton Envision

This weekend at Princeton Envision, will join the Future of AI in Education panel alongside Tammy Kwan, Amber Berry, and Morgan Mifflin. Jes will also host discussions spanning education at large, AI and knowledge generation, careers, and the flexibility and adaptability of the modern landscape.

For those who track the Envision conference series, this was a meaningful return after the 2017-2019 run. It continues to be one of the more thoughtful convening spaces at the intersection of technology, ethics, and how we view the future.

A New Partnership: Bridgecraft

We’re glad to announce that Bridgecraft has emerged as a new JOPRO partner. Bridgecraft will support several projects connected to the JOPRO Futures Center, as well as mentoring, leadership, and career development efforts more broadly.

MIT Media Lab: AHA Seminar Series

Two recent lectures from MIT’s Advancing Humans with AI group are worth flagging for those in our network.

Michelle Vaccaro explored the conditions under which human-AI collaboration actually outperforms either humans or AI working alone, drawing on her research into synergy, organizational behavior, and the design of meaningful AI safety evaluations.

The session on HumaneBench, led by Erika Anderson and Yaoli Mao, introduced a new paradigm for humane AI accountability that moves beyond static model cards and disclosure toward continuous, outcome-based evaluation of how AI systems shape human behavior and wellbeing in practice.

Both sessions speak directly to questions our community has been engaging with around responsible AI design and its implications for human capability.

David Krakauer at Brain Inspired Complexity Discussion Group

Recently, Paul Middlebrooks’ Brain Inspired Complexity Group convened to discuss Mario Bunge’s 1962 paper “The Complexity of Simplicity,” with an introduction by Thanos Alexandris and special guest: David Krakauer, President of the Santa Fe Institute, author of The Complex World, and editor of the Foundational Papers in Complexity Science series.

Also of note, Luis Pessoa led an intriguing discussion on his recent paper, exploring advances in networks and complexity science. We plan to discuss this paper in collaboration with Bradly Alicea during a future session of Orthogonal Research and Education Lab’s SMN meetings.

New Kinds of Science Organizations

In line with the JOPRO and Orthogonal Lab theme of “Doing Science Differently”, we note fellow ekkolapto contributor and colleague Daniel Van Zant’s work with the Wolfram Institute. He has published a new report: “How Democratic Coarse-Graining Filters Against Some of the Research That Matters Most,” the first in a six-part series on “New Kinds of Science Organizations.”

For those following questions about how institutional structures shape which research gets supported and which gets overlooked, the series is well worth tracking. Additionally, we continue to follow and support and their ongoing programming; a number of exciting events are on the horizon.

Plot Twisters: Metagov Fellowship Completion

Congratulations to our sister group on the successful completion of the Governable Spacemakers Fellowship through . Their project, Small Hassles Court, is an experiment in community governance and digital space design, and an early-stage mini-game within the forthcoming EdTech gameworld, Twisterland. continues to drive momentum on other Plot Twisters initiatives, with articles in development exploring reflection, agency, and intentionality in game design.

Upcoming Events and Opportunities

Looking ahead, here are a few events and programs that may be relevant for people in our network.

Embodied Intelligence 2026

The International Conference on Embodied Intelligence is back for 2026. Registration is open for Masterclasses as well as presentations. JOPRO’s Cognition Futures working group looks to present once again this year, this time on the nature of embodiment, agency, and selection from alternatives.

NYCWiC 2026

JOPRO will be hosting a discussion at ACM’s New York Celebration of Women in Computing on AI, jobs, and early career options in an era of rapid change — in addition to co-hosting a mini-hackathon at the event. If you’re an early-career researcher or practitioner navigating these questions, this session is designed with you in mind. Registration for the conference is still open, but very limited spaces remain.

MonasteryX Residency

Our friends at are hosting the MonasteryX Residency in Thailand, with a few spots still available for digital nomads, founders-in-development, or those who’d want a structured yet holistic experience to focus on research or developing your craft.

Iris van Rooij: Critical AI Literacies Summer School

Professor Iris van Rooij is running a five-day summer school through Radboud University titled Critical AI Literacies for Resisting and Reclaiming. A strong opportunity for those based in Europe or willing to travel, particularly for researchers, practitioners, and public sector professionals who want to develop more critical frameworks for engaging with AI.

Accelerators and Hackathons

Two items of note for those in our community with entrepreneurial or applied data interests: UC Launch at UC Berkley (open to all UC students, faculty, and alumni), and DataHacks 2026, which will be hosted at UC San Diego. Both programs are looking for project mentors (with virtual options available!) — if you’re interested in serving as one, reach out to the organizers directly.

Website Updates and Project Calls Ahead

Our JOPRO website is scheduled for a broad update in style and function in the months ahead. As that change takes place, we will be posting updated opportunities in these newsletters and on social media as the Futures Center and other projects formalize.

If you want to stay close to what’s developing, subscribing here and following us on LinkedIn are the best ways to do that.

As always, thank you for being part of this ecosystem.

-The JOPRO Team

JOPRO coordinates and cultivates researchers, builders, and practitioners working to modernize how knowledge is generated and applied across domains.

Our work combines project incubation and operator formation with working groups, workshops, and seminar series— supporting both field-level transitions and focused research efforts, in collaboration with academic, civic, and industry partners.

Subscribe now

What stood out was not confusion, but a lack of orientation: a sense that the systems engulfing us offer brittle outputs and legacy credentials, with a pronounced lack of guidance for becoming responsible, wayfinding actors within complex, evolving fields. The formation of discernment, building coherence beyond disciplinary silos, and the ability to navigate uncertainty are increasingly treated as discrete, private, or even irrelevant problems.

One hears a growing resignation: this is out of our hands; better to look away or focus elsewhere, or elsewhen. That posture sits uneasily beside the agentic rhetoric of flourishing-centered technologists, ethically-courageous innovators, or indeed philosopher-builders. The more generative stance has often been quieter and more demanding: nobody taught us how to do what this moment requires — so we are teaching ourselves.

Speaking of philosopher-builders, as Brendan McCord noted during the Cosmos Institute’s year-end note: “Technology that expands what is possible may narrow who is possible. But it need not.”

I would extend that observation beyond who to where. When the range of who is possible contracts, so too does the range of futures we can chart — along with the institutional and epistemic pathways that make those futures reachable.

The moment we find ourselves in

For much of the last century, our institutions for learning, research, and innovation were designed around narrower assumptions, clearer disciplinary boundaries, slower feedback cycles, and more homogeneous pathways into expertise and authority.

Even then, many forms of knowledge, experience, and responsibility were marginalized or excluded. Today, as complexity accelerates and domains collide, those limitations are no longer merely unjust or inefficient — they are structurally misaligned with the kinds of problems people are now being asked to work on.

The challenge is not that we’ve lost a golden age of formation, but that we are encountering, at scale, the consequences of institutions that were never built to support integrated comprehensions, cross-field coordination, or long-horizon responsibility in the first place.

JOPRO exists because those gaps keep showing up in practice — across fields, institutions, and career stages. It shows up when early-career researchers are technically prepared but strategically ungrounded. It shows up when interdisciplinary work collapses under the weight of mismatched assumptions and timelines. And it shows up when conversations about ethics or responsibility are detached from the lived realities of learning, leadership, and decision-making.

Over time, it became clear that addressing this gap could not be done by focusing on any single program, discipline, or institution in isolation. The challenges we’re pointing at — fragmentation, loss of orientation, shallow integration — are not failures of individual effort, but of the surrounding ecosystems people are asked to operate within.

That realization shaped how our vision has evolved during the year. Instead of treating JOPRO as a standalone platform or pipeline, we increasingly understood our work as part of a broader landscape: one that depends on relationships between groups, shared language across domains, and people willing to take on the work of translation and coordination.

In other words, the question became not just what we build, but how the conditions for building — and becoming — are cultivated.

In practice, this has meant focusing on work that tends to be neglected in periods of rapid technological acceleration. Such upstream work is slower and difficult to measure. But we’ve come to believe it is essential if new technologies and ideas are going to land inside systems that can actually absorb them responsibly. Or, harder yet, mapping trajectories of where systems are going, and discerning which developments drive or dissipate movement along those pathways.

Cultivating Ecosystems and Operators

One way we’ve come to describe this moment is simply: it is not enough to have discrete ideas. (Or technologies). What matters now is whether those insights can be translated into durable practices, institutions, and cultures that actually shape how systems behave.

At JOPRO, we’ve increasingly come to see our role not only as incubating and mentoring, but evolving to help cultivate ecosystems of makers, operators, and visionaries — people and groups capable of carrying ideas across domains, coordinating across institutions, and doing the often invisible work of translation. That means showing up in spaces where thoughtful work is already happening, learning from those efforts, and contributing where we can.

This year, that took many forms.

We continued deep collaboration with Orthogonal Research and Education Lab, JOPRO’s longtime partner, particularly through the Society Ethics Technology working group. Together, we explored questions of project management, leadership, and coordination in research and innovation spaces — topics that are rarely treated as first-order concerns, but which often determine whether good ideas survive contact with reality.

We also stayed closely connected to Plot Twisters, a JOPRO sister group building on work supported by OpenIDEO’s Designing for Digital Thriving grant and ’s Governable Spacemaker Fellowship. Plot Twisters exemplifies the kind of experimental, practice-oriented inquiry we believe is essential for understanding agency & narrative, digital spaces, and individual & community sensemaking as tractable design challenges rather than abstract problems.

Across the year, members of the JOPRO community participated in and learned from a number of adjacent ecosystems. We attended events and followed the growth of Advancing Humans with AI (MIT), a community whose emphasis on human-centered AI aligns closely with our own concerns, and whose continued evolution we’re excited to see in the year ahead.

We’ve also been in conversation with, and inspired by, organizations like Distributed AI Research Institute, particularly around their Possible Futures workshop series, which foregrounds participatory approaches to imagining technological trajectories. Similarly, many in our community engaged with , including participation in the Oxford AI x Philosophy Seminar — a reminder of how valuable serious philosophical engagement remains in technical domains.

In more applied and embodied spaces, JOPRO members attended events hosted by groups like Augmentation Lab, and we exhibited at the AugLab Summit at the MIT Media Lab. In addition to supporting the Summit, our colleagues at convened numerous salons and events across crucial frontiers — creating space for reflective engagement with questions of technology, “Cognition Futures,” and human experience. Together, these environments challenged us to think more carefully about how narratives, interfaces, and systems shape human capability.

We also engaged with communities focused on values, responsibility, and career pathways, including School for Moral Ambition, whose work on alternative and impactful career trajectories resonates strongly with many in the our network. We reconnected with collaborators at Princeton Envision, as they revived the Envision conference series with renewed emphasis on data ethics and responsible AI.

Taken together, these relationships reflect a simple conviction: no single organization will solve the problems we’re facing. Progress depends on overlapping communities, shared language, and people willing to operate across boundaries — intellectually, institutionally, and ethically. JOPRO’s work this year has been shaped by that belief, and the projects we built grew directly out of these ongoing conversations and collaborations.

What we advanced this year

This year at JOPRO was less about scaling a single program and more about laying interlocking foundations. Across research, pedagogy, and community work, we were asking a consistent question: what does it look like to design learning, inquiry, and leadership structures that are actually responsive to the conditions people are navigating now?

Rather than answering that in the abstract, we tested it in practice — across education, research, and public engagement — letting each strand inform the others over the course of the year.

Throughout the year, much of our work focused on education and learning ecosystems. We formalized what we’ve been developing implicitly for years into a framework we call JOPRO & the New Learning Ecosystem. This work articulates an approach to learning that is contextual, developmental, and trajectory-aware — designed for interdisciplinary spaces where people are often navigating multiple identities, skill sets, and time horizons at once.

We shared this framework the spring at the UC San Diego Education Innovation Expo, using it to demonstrate how pedagogy, technology, and mentorship can be designed together rather than treated as separate layers. That same philosophy guided our participation in Google Summer of Code through Orthogonal Research and Education Lab (and sponsor International Neuroinformatics Coordination Facility), where JOPRO supported scholars contributing to open-source and research projects at the intersection of computation, biology, and ethics. These efforts reinforced a core belief: meaningful learning happens when people are embedded in real problems, with real responsibility, and adequate scaffolding.

As the year progressed, this educational work increasingly intersected with public engagement and institutional presence. In the summer, JOPRO exhibited at the Human Augmentation Summit hosted at MIT Media Lab, contributing to conversations about human potential, augmentation, and the social consequences of emerging technologies. Members of our community also engaged with global research communities through conferences like ACM Conference on Fairness, Accountability, and Transparency in Athens under the Data x Direction banner, exploring questions of fairness, accountability, and participatory approaches to AI.

Through our Society Ethics Technology working group, JOPRO was represented at WeRobot 2026 — a leading conference at the intersection of law, robotics, and policy — and we organized a panel at ACM’s New York Celebration of Women in Computing (NYCWiC), centering early-career researchers navigating innovation spaces across academia and industry. In these settings, JOPRO wasn’t just presenting progress; we were stress-testing ideas, building relationships, and clarifying where our voice adds value.

Running alongside this public work was a deeper strand of research and sensemaking. Over the course of the year, we developed Something In The Way: An Ethnography of 21st Century Directionlessness, a project in the Mental Health Paradigms & Perspectives working group aimed at documenting what it feels like to be adrift in contemporary life — professionally, cognitively, and morally. Drawing from ethnography and phenomenology rather than optimization frameworks, this work shaped how we think about mentorship, leadership, and learning: not as linear pipelines, but as processes that must account for uncertainty, fragmentation, and uneven development.

Late in the year, many of these threads came together. In December, we launched DigiNEST (Digital Narratives and Emerging Story Technologies), a new reading group and discussion space exploring narrativity, agency, and emerging technologies. DigiNEST sits at the intersection of digital humanities, philosophy, psychology, and design, and reflects JOPRO’s conviction that questions of ethics and governance are inseparable from questions of story, identity, and meaning-making. In many ways, DigiNEST represents a synthesis of the year’s work: a space where research, pedagogy, and community converge.

Running through all of this was a quieter but essential effort: field-building and mentorship. Projects such as From Here to There and Data x Direction (both of which are spinoffs from my capstone coursework at UC San Diego) are not products in the conventional sense; they are attempts to offer language, structure, and guidance for people navigating complex innovation landscapes without clear maps. Data x Direction, in particular, had a wonderful initial internship cohort and serves as a growing community within the broader JOPRO umbrella.

These efforts reflect a belief that leadership today is less about authority and more about helping people orient themselves — intellectually, ethically, and strategically — over time.

Taken together, this year’s work assuredly didn’t resolve the tensions we’re studying. It created shared reference points. New vocabularies. Prototypes for learning, inquiry, and collaboration that can continue to evolve. As an organization, we continued to practice incubation for projects, and grew in our capacity to cultivate context and capacity.

Vision for 2026

As we look ahead, much of our focus will be on deepening the work that began this year — and bringing it into clearer, more sustained forms.

In the coming year, JOPRO will continue to show up in public-facing spaces where questions of technology, ethics, and leadership are actively being shaped. This includes delivering a keynote at Princeton Envision in February, as well as hosting a Future of Work and Tech Careers panel at NYCWiC 2026. These moments matter to us not simply as visibility, but as opportunities to engage early-career researchers and practitioners who are navigating real decisions about where and how to apply their skills.

We’re also continuing to develop new formats for sustained inquiry and discussion. DigiNEST will grow in the year ahead as a space for grappling with narrativity, agency, and emerging technologies, and we expect to be involved in a range of additional Discussion Series, collaborations and events as these conversations evolve. Alongside this, we’re continuing to expand local gatherings in Boston and New York, as well as hybrid and virtual events with inquisitive communities such as communities ekkolapto, Frontier Tower, and the Oxford Internet Institute.

A major focus for the year ahead will be the launch of the JOPRO Futures Center. The Center is an extension of work we’ve been circling for some time: creating a dedicated space for developing meaningful alternatives to prevailing systems and assumptions. This will include exploratory projects, long-horizon research, and a forthcoming invitation-only cohort for serious philosophical and strategic discussion — focused on crafting the skills, curricula, and organizations needed to bridge the world as it is with the world as we might prefer it to be.

As always, we don’t see this as something JOPRO does alone. The most important work ahead will happen through collaboration: with people willing to think carefully, act responsibly, and take the long view even when incentives push in the opposite direction. If that resonates with you, whether as a collaborator, participant, doner, or quiet supporter, we’d love to stay in conversation.

What ultimately matters is not only what technologies are being built, but what we are able to perceive, comprehend, and coordinate as we navigate the paths they open. The quality of that understanding and the structures that support it will shape far more than any single project.

To everyone who contributed time, thought, care, and trust this year: thank you. This work is only possible because it is shared.

Jes

JOPRO coordinates and cultivates researchers, builders, and practitioners working to modernize how knowledge is generated and applied across domains.

Our work combines project incubation and operator formation with working groups, workshops, and seminar series— supporting both field-level transitions and focused research efforts, in collaboration with academic, civic, and industry partners.

We’ve grouped them across Virtual, Boston/Cambridge, and San Francisco, spanning everything from AI salons and complexity science to moral ambition, brain aging research, and a new gallery of researcher-creatives.

If you’re attending any of these, feel free to drop reflections or notes in the comments — or let us know if you’d like to meet others who will be there.

Virtual Events

The Future of Funding R&D — with Jim O’Shaughnessy & Atman Pandya (Computational Philosophy / Polymath Salon)

Jim O’Shaughnessy and Atman Pandya of O’Shaughnessy Ventures (OSV) joined the Polymath Salon to discuss emerging models for funding R&D in the arts and sciences—especially for creators, researchers, and founders working between or outside traditional academic categories.

🗓 Mon, Nov 10 — 12:30–2:30 PM ET
🔗 Event: luma.com/oshaughnessy
🎥 Replay: YouTube via
🔗 OSV: www.osv.llc

comments: “Worth checking out the video, and following ’s writings and podcast. I’m a fan of his combination of optimism and sobriety, and would encourage our community to check out opportunities from OSV proper, as well as the network present within this conversation. Thanks to Addy & Co for organizing the event — and do check out the Computational Philosophy Club at MIT for more topics; more ahead with this blossoming collective.”

Brain Inspired: Complexity Group — Discussion on Minsky (1961)

A reading and dialogue session unpacking Marvin Minsky’s “Steps Toward Artificial Intelligence,” led by Darren Lott.
🗓 Wed, Nov 12 — 12:30 PM ET
🔗 Zoom, semi-regularly, replays available via Brain Inspired - Complexity Group.

: “Paul Middlebrook (Brain Inspired) is courageously leading us as we deep dive through David Krakauer’s (Santa Fe Institute) Complexity Papers, alongside many intriguing figures (Kevin Mitchell, Anna Riedl, Luiz Pessoa, and many more!) It’s intense but if you’re interested in understanding the roots of modern complexity science, they are invaluable discussions.”

Emergence of Natural & Artificial Patterns (FAU / MPCR)

This session is kindly brought to you by the Machine Perception and Cognitive Robotics Lab at Florida Atlantic University hosted at the Sandbox AI Center at Florida Atlantic University, Boca Raton Campus.

🗓 Wed, Nov 12 — 6:00 PM ET
🔗 Livestream: https://www.youtube.com/@mpcrlabs9710/streams

Playtest: Small Hassles Court — Metagov Governable Space-Makers Fellows (Plot Twisters)

A live playtest session of Small Hassles Court, a digital game that helps players practice interpersonal conflict resolution and self-governance, developed by the Plot Twisters collective as part of their developing Twisterland project; hosted through Metagov’s playtester community.

🗓 Fri, Nov 14 — 11:30 AM–12:30 PM ET
💻 Zoom (RSVP Required)
🔗 Event: https://luma.com/
🔗 Fellowship Info: (on Substack)

: “Check out our sister group’s progress on one component of Twisterland, our forthcoming EdTech gameworld! There will be multiple iterations of this event, so if this Friday doesn’t work, check Luma for other events. All participation and feedback is welcomed; we’re looking to build something new and quite timely! Last year, we presented at Oxford Internet Institute’s Inclusive Gaming Conference, so check out the latest developments at these playtesting events.”

Boston + Cambridge

Civil Discourse Series — What Ties a Nation Together? (Harvard Safra Center for Ethics)

A panel discussion hosted by the Edmond & Lily Safra Center for Ethics exploring the cultural, political, and social forces that hold a nation together, with a virtual Q&A component and full livestream access. Attendees must verify registration (HUID if applicable), and accessibility accommodations are available upon request.

🗓 Wed, Nov 12 — 4:30–6:00 PM ET
📍 Smith Campus Center Commons, 1350 Massachusetts Ave, Cambridge
💻 Also online
🔗 Event Info: @ Harvard Ethics | YouTube Livestream

Measuring Aging Within the Human Brain — Salon with David Bennett (Hosted by MIT Prof. Manolis Kellis)

A dinner salon hosted at the home of MIT Professor Manolis Kellis featuring David Bennett, creator of the 30-year ROSMAP aging-brain cohort, discussing insights from thousands of human brains, the architecture of long-term cohort science, and his new PARDoS initiative expanding diversity and molecular depth in aging research.

🗓 Wed, Nov 12 — 6:30–8:30 PM ET
📍 Brookline, MA (private residence)
🔗 RSVP: https://forms.gle/

Redefining Success: A Conversation with Rutger Bregman on Moral Ambition @ Harvard

Hosted by David Jaffin & Laura Clancy, this talk challenges students to rethink conventional career ladders and embrace “moral ambition,” drawing on historical disruptors and launching the new $15,000 Harvard Moral Ambition Fellowship. Early attendees receive signed pre-release copies of Moral Ambition.

🗓 Fri, Nov 14 — 4:30–5:30 PM ET
📍 Harvard Science Center Room 309A
🔗 Event: https://luma.com/

Harvard x School for Moral Ambition — JOPRO Social (Hybrid)

Following the Rutger Bregman talk, JOPRO Director Jes Parent and community members will host an in-person + virtual discussion on moral ambition, alternative career pathways, and aligned initiatives JOPRO is supporting. Harvard event attendance is recommended but not required; online link provided.

🗓 Fri, Nov 14 — 5:30–6:30 PM ET
📍 Harvard Science Center Plaza, Cambridge (and forthcoming virtual component)
🔗 Event: https://luma.com/

: “See you there! I’ve been a supporter and almost worked on this project; glad to see very intentional Alternative Crafting, particularly for young folks. Stay tuned for similar efforts at Princeton, Stanford - and perhaps a university by you? Or, would you like support making your own event even outside of SMA? We can help - let’s talk!”

San Francisco

VOLUME 1: The Rise of the Researcher-Creative Opening Gala

We have a special message from Jolie Gan of :

“There’s a quiet revolution happening in scientific research.

The Distressed Scientists’ Department opens Volume 1: The Rise of the Researcher Creative, the first creative research exhibit in San Francisco on Mon, Nov 17 -- a 3-day exhibition gallery for researcher-creatives who bring their work beyond papers into tangible artifacts.

Expect code as sculpture, EEG as fashion, and bioengineered flowers, plus gallery walkthroughs, short on-stage conversations with the next generation of researchers, and a printed passport-style booklet to stamp and get signed. If you fund, create, or simply love frontier research made public, this room is for you.”

🗓 Fri, Nov 17-19
📍 Pebblebed, 1417 15th St, San Francisco
🔗 RSVP here: luma.com/

: “This looks to be a wonderful event - I wish I could attend in person! SF crew, please show up and support this. Perhaps we can do a similar event in Boston next year? , looking at you!”

FrontierTower

Multiple Events. Too many to list! Check out the many things going on at Frontier Tower, perhaps best by following them on Luma!

That’s a wrap

There’s everything on our radar for this week — but we’re sure we’ve missed something. If you’re hosting an event, know of one we should include, or want help coordinating a gathering in your city or university, reach out or drop a note in the comments.

Leave a comment

A Sneak Peak at 2026

Stay tuned for JOPRO at Princeton Envision (February) and New York Celebration of Women in Computing (April). The development of the JOPRO Futures Center is also underway — stay tuned for forthcoming newsletters around such.

Get involved!

And if you’d like to collaborate with JOPRO, connect with folks in the network, or explore hosting aligned programming (moral ambition, alternative career paths, operator labs, futurist practice), we’re always happy to talk.

Interested in any of these opportunities — or would you like to collaborate on related projects or initiatives? We’ll send RFPs, dates, and materials as they open. We look forward to hearing from you.

👉 One step for everything: [Interest Form]

See you out there — and keep us posted on what you’re building.

— The JOPRO Team

JOPRO is excited to share news from our partners at the Orthogonal Research and Education Laboratory (OREL). OREL has announced its 2025 cohort of Google Summer of Code (GSoC) Contributors, who will be contributing to open-source projects that advance collaborative research and innovation.

Each year, the GSoC program connects early-career technologists with mentor organizations to gain real-world experience in open-source development. OREL’s participation reflects its commitment to building accessible, community-driven research tools and fostering the next generation of scholars.

As a partner organization, JOPRO celebrates OREL’s dedication to mentorship, open science, and innovation—values that resonate deeply with our own mission as both a research and vocation incubator.

We look forward to following the work of this year’s scholars as they bring forward new ideas and strengthen the ecosystem of open, collaborative research.

Read the full announcement and meet the 2025 GSoC scholars here: Welcome to Our Google Summer of Code Scholars for 2025 – OREL

Acknowledgements

Thank you to the International Neuroinformatics Coordination Facility and The OpenWorm Foundation for continued funding and support during Google Summer of Code.

For more about OREL’s legacy of Google Summer of Code & Summer Cohort, view our GSoC & Summer Cohort archive page.

A version of this post originally appeared on JOPRO.org

JOPRO will join faculty, students, and technologists at the 2025 UC San Diego Education Innovation Expo (May 21–22) to explore cutting-edge pedagogy and educational technologies reshaping higher education. The two-day event offered a vibrant exchange of ideas, showcasing projects aimed at enhancing creativity, access, and impact across the university and beyond.

Founder and Director Jesse Parent will deliver a talk on “JOPRO & the New Learning Ecosystem: Designing Tools and Communities to Foster Contextual, Developmental, and Trajectory-Aware Learning in Inter-/Un-disciplinary Spaces.” The talk outlines JOPRO’s evolving model as an interdisciplinary research and vocation incubator, emphasizing the need for learning systems that are historically grounded, strategically oriented, and responsive to the complex paths of 21st-century learners.

The presentation featured FrontierMap, a tool to help learners understand trajectories in emerging frontiers of study; case studies on alternative education approaches (Orthogonal Research and Education Lab and Neuromatch Academy); alongside the EdTech example of of narrative-driven initiative Plot Twisters, which center personal agency and co-designed learning in AI-mediated environments.

JOPRO’s participation will emphasize the importance of context-aware mentorship, nonlinear career scaffolding, and ethical foresight in preparing students to contribute meaningfully across academia, policy, and industry.

More information about the Education Innovation Expo

educationalinnovation.ucsd.edu

A version of this post originally appeared on JOPRO.org