A Mind, and One Mind
The conditions for mindedness answer two questions: what makes a system’s states mental, and what makes them the states of one mind. Kept apart, they show what a mind does not need
On Monday a model working through a long task writes itself a note: the migration fails on tables with composite keys; try the staging copy first. The session ends. On Tuesday a new session starts from the same weights. Nothing of Monday’s activity survives in it. Its context, the running text it reads at every step, begins with a one-line index of the notes that earlier sessions left. One line bears on the task. It opens that note, tries the staging copy first, and gets further than Monday did.
Was one mind at work across the two days, or two?
Our essays have answered both ways. A deployment is a network together with its files and the loop that runs it. Which AGI? held that it cannot have an inner life, because the integration an inner life needs does not run through a file system. “The deployment is intelligent; the process, if anything, is minded; neither is both.” Is the Monastery Minded? held that a house whose generations rebuild its understanding from the records their predecessors left is as minded as a house that keeps the understanding alive in conversation. Tuesday’s session is a generation working from records.
The two verdicts came from one list of conditions. Minds by Degrees gave four conditions. A mind has vehicles, structures that carry content and whose parts recombine. It revises them by how well they fit what they are about. It runs simulations and uses the results. And it is one system. The essay treated the four as components of a single mechanism. I now think they answer two questions. Three of them say what makes a system’s states mental. The fourth says what makes those states the states of one mind. Stated by itself, the fourth asks for less than Which AGI? assumed, and the disagreement about Monday’s note can be settled.
Keeping the questions apart also shows what a mind does not need. It does not need to be coherent. It does not need anything that persists through its activity. It does not need a self. Each of the three was built into mindedness by an earlier essay of ours or by a rival account. Coherence and a persisting carrier belong to experience, if they belong to anything. A self belongs to personhood.
The essay has a debt to pay as well. Michał Wójcik, a correspondent whose objections have shaped several of these essays, complained that they “never show toy examples which satisfy some properties and fail others like mathematicians do when presenting a theory. Therefore, I don’t know what the words mean.” The monastery essay answered with pairs of systems that differ in one feature. This essay owes a pair for every condition it keeps, and I will treat a condition that no pair isolates as a candidate for removal.
This essay replaces Minds by Degrees. What it keeps is restated here.
The kind that psychology is about
Mental terms are defined by one another. A belief is the kind of state that combines with a desire to produce an intention, and a desire is the kind of state that combines with a belief. This looks like a circle, and for a long time we tried to break it by stating the conditions on mentality without mental vocabulary.
The circle is harmless. Frank Ramsey and David Lewis showed how the terms of any theory get their meaning together. Write the theory as one long sentence, replace each theoretical term with a variable, and the terms denote whatever satisfies the sentence. “Electron” and “charge” are defined by one another in the same way. So the question about minds is the question one asks about electrons. Is there a successful science whose explanations use these terms, and does its success give reason to think the terms pick out something real?
The science is psychology, and the choice matters. Frances Egan is a realist about the computations a brain performs and a deflationist about their content: on her view, calling a computed function a representation of an edge is a gloss that helps us connect the computation to our interests. One can hold her view and take cognitive science at face value. Psychology leaves no room for it, because its laws are stated in mental terms from the start. The Weber–Fechner law relates stimulus intensity to the magnitude of a sensation. The serial-position curve is a regularity of memory. Stroop interference is a regularity of attention, prospect theory of preference, the spacing effect of learning. Take away the mental terms and the generalizations are gone, because nothing else groups the cases.
Realism about all of psychology would be refuted by its history. The psychology of 1940 was behaviorist, and some of the social psychology of 2012 did not replicate. The realism worth holding is selective. It attaches to what keeps being used as theories change: the posits that the cumulative core cannot do without. This leaves the typology of mental states open. Peter Carruthers has argued that belief, desire and intention may not be the joints at which human cognition divides, since much of what we do comes from skills, habits and urges. If he is right, there are still minds. The kind is secure while its inventory is being revised.
What kind of kind is it? Richard Boyd described kinds whose members share a cluster of properties because some mechanism keeps the properties together, and not because a definition says so. Species are his example. Membership in such a kind comes in degrees and has fuzzy borders. For minds, the cluster is the body of psychological regularities, and the conditions of this essay describe the mechanism that keeps them together. Say that psychological explanation has traction on a system when it predicts and explains what the system does. The central conjecture is that traction and the mechanism go together: psychological explanation gains valid traction on a system if and only if the mechanism runs in it.
Traction is evidence. It is not what makes a system a mind.
What makes the parts of the mechanism real, and not a way of describing a system that some observer finds convenient, is a general matter that What a Description Answers To takes up. Its criterion is that a pattern is real, relative to a family of interventions, when a model of it is a causal abstraction of the system under an admissible map. Roughly: the system can be modelled in terms of the pattern, intervening on the system has the effects the model says it should, and the mapping between the two does not have to be rewritten whenever the system’s fine detail changes. For a mental state the family is not an observer’s. It is the interactions among the system’s own parts while it operates. A vehicle is real if the rest of the system treats it as one.
The test of traction needs three qualifications, and the third is new.
It must not be idle. Minds by Degrees considered a weather-prediction system that learns its internal states, generalizes, and has fresh observations fed into it by a step its engineers built. One could say that it perceives the atmosphere and expects rain. Nothing is gained by saying so. Meteorology and numerical analysis account for everything it does.
It must not be parasitic. A system trained on human writing can reproduce the regularities of human psychology because the writing contains them. The remedy is to check that a psychological construct is doing work inside the system: find what realizes it, intervene on that, and see whether the behaviour changes as the construct predicts, including in conditions the training did not cover. The remedy has a known weakness. If any mapping from a system’s internals to a construct is allowed, a mapping can always be found. Denis Sutter and colleagues proved a version of this for neural networks, and then found mappings that scored perfectly on networks that had never been trained (arXiv:2507.08802). What separated the trained network from the untrained one in their experiments was a test on inputs of a kind the mapping had not been fitted to. So the test has to say which mappings count. Atticus Geiger, Jacqueline Harding and Thomas Icard ask that the mapped variables be “native to how the system works” (arXiv:2508.11214). The admissibility clause of the criterion above is my version of that demand: the mapping must stay fixed while the system’s lower-level state varies in ways the system’s own interactions cannot detect.
And it must not be additive. A choir’s behaviour is predicted by the psychology of its singers. A psychology of the choir earns its place only if it predicts what the psychology of the singers, taken one at a time, does not. Sociology and economics borrow psychology’s models of attention and choice without making societies or markets into minds. Extending the methods of psychology is not extending what it is about.
One claim of the earlier essay is withdrawn. It said that experience is one stratum of the mechanism that sustains the kind, so that our account was “homeostatic all the way up”. Boyd’s account of kinds stays.
Cognition is not yet a mind
Michał offered a system simpler than a monastery: a labyrinth of tubes holding water and air, which one tilts and shakes. He asked whether it meets our conditions, and reported that he could not tell. The monastery essay built three labyrinths in reply, and they make a ladder.
In the first the water only sloshes. Each state is caused by the last and by the tilt. Nothing in the labyrinth assesses anything. This is causal interaction.
In the second a valve opens when the pressure in a chamber passes a threshold. High pressure is registered, and the release returns to the chamber that produced it. This is regulation, of a physical variable. The pressure is not about anything.
In the third, two sensors each measure the pressure in a distant chamber, and a mechanism compares them. When one sensor’s readings keep diverging from the other’s, and from the pressure later observed at the outlet, the mechanism stops using it. Here two states are about the same thing and can disagree, and one of them can be found unreliable. This is regulation that is sensitive to content.
None of the three is a mind. The third has content, but its vehicles are two numbers. They combine with nothing, and they are a model of nothing.
Plants are a rung higher. A Venus flytrap closes when two of its trigger hairs are touched in quick succession, and it sets the scale of its digestion by counting further touches. A pea plant with its roots divided between a steady and a variable supply of nutrients grows toward the variable one when nutrients are scarce and toward the steady one when they are plentiful, as the theory of risk-sensitive foraging predicts for animals. A mimosa dropped repeatedly without harm stops folding its leaves, and keeps the lesson for a month. This is memory, decision and learning, and biologists call it cognition without embarrassment.
Cognition is cheap, then, and it is not what the conditions below are conditions for. A plant remembers, decides and learns with nothing in it that can be recombined into a representation of a situation it has not met. The first condition on a mind is about that.
What makes its states mental
Three conditions say what makes a system’s states mental. For each I give a pair of systems that differ in that respect alone, and the family of psychological regularities that goes missing without it.
Vehicles that compose
A mind’s states include vehicles: structures in the system that do causal work and that have parts which can be recombined. A system with such vehicles can represent a situation it has never met, by putting together what it has. A lookup table cannot.
The pair is one network at two times. A small network trained on modular addition first memorizes its training pairs. It answers them perfectly and does no better than chance on pairs it has not seen. Trained much further, it suddenly generalizes, and when its weights are examined they turn out to implement a general method: the inputs are represented as rotations about a circle, and adding the numbers is composing the rotations (Nanda et al. 2023). The early network and the late one give the same answers on the training set. What changed is whether the network’s states have parts that combine. The late network computes its answer for an unseen pair from its representations of the two numbers. The early one has not been taken apart in the same way. It is known by what it does, which is to store. For a long stretch of training both solutions are present together, so the pair is the network at the start of that stretch and at the end.
The regularities that depend on this condition are the ones psychologists call productivity and systematicity. A mind that can think that the dog chased the cat can think that the cat chased the dog, and can go on to cases it was never shown. The network shows the smallest case of it: one operation, over a closed set of numbers.
Two cautions belong here. A vehicle need not be a neuron or a fixed direction in a network’s activity. Contents can share the same units by taking turns in time, or by occupying different phases of an oscillation, and a unit that seems to carry a mixture may carry one thing at a time. And the vehicle’s content need not be fully determinate. Minds by Degrees had a separate condition here, that content exists at several levels. It rested on a finding by Schwartz, Mandelbaum and colleagues: people accept an argument that equivocates between two related senses of a word, such as dinner the event and dinner the food, and reject one that equivocates between unrelated senses, such as the two meanings of bat. Inference runs over a vehicle that has not yet been resolved into one sense or the other. That is a fact about what rich vehicles are like. It is not a further part of the mechanism, and I no longer list it as a condition.
Revision that answers to fit
The vehicles are kept, changed and replaced according to how well they fit what they are about. Failure reorganizes them, and success consolidates them. A system that only reorganized would never settle, and one that only consolidated would never adapt.
The pair: an index that is consulted and never altered by how well its entries served, against a model’s context in which being told that an answer was wrong changes the answers that follow. The first is a store. The second answers to fit.
The condition is a disposition, and it has a direction: toward fit with the targets. It is not the maintenance of coherence. A mind can stay incoherent for years, and a later section gives the cases. The condition also does not require that anything be learned for good. A session that revises its answers within one conversation and keeps nothing afterwards meets it, as a person with dense amnesia does within an afternoon.
Conflict among a system’s commitments is one source of evidence about fit. Where no observation arrives to say that an answer was wrong, a system can still improve by finding that its commitments conflict. Before the Search Settles says when that kind of improvement can be trusted.
The regularities that depend on this condition are the laws of learning: revision driven by error, consolidation, the effects of spacing and reinforcement.
Simulation the system uses itself
A mind runs its vehicles through transformations that mirror how their targets change, and it uses the results in its own further processing. Kenneth Craik called this a small-scale model of the world. The condition has two halves, and each has a pair.
The first half is the mirroring. A house that copies a table of Easter dates made elsewhere is looking something up. A house that computes Easter itself, by a procedure whose steps track the cycles of the moon and the sun, is running a model.
The second half is the use. Take the weather system again, and suppose its forecasts end on a screen that people read. Its simulation is consumed outside it. Now suppose it has learned to use its forecasts itself: it compares what it predicted with what came, and what it learns from the difference changes how it represents the atmosphere. The second system is the first one on which a psychology might begin to get a grip, and to that degree it has a thin mind. I accept that verdict. Such a system would be minded and would have none of what makes minds matter to us.
Tools raise the same question for an agent. An agent that sends a problem to an outside solver and pastes in the answer has delegated. An agent that steps a simulator, looks at the intermediate states, and changes course because of them is running the simulation with an instrument, as a navigator does with a chart.
How much does the condition ask? Following the target as input arrives, and correcting from the result, is enough. I will call that emulation. A system that can also run a configuration it does not take to be the case, a frost it does not expect, has more: it can ask what if. That is a grade of mind and not a condition on having one. Nor must the results select an action. A dreamer selects none and is no less minded for it.
Internal use adds no family of regularities of its own. It makes the other families regularities of the system, and not of the system together with its users.
Supposing, and holding a view one does not accept
A system that can ask what if uses the results of a run that it does not take to describe how things are. Something must keep the imagined frost from being believed.
It is tempting to posit a second attitude beside belief, a way of holding a content without accepting it, and a barrier that keeps the two apart. I think one attitude is enough. A simulation has a target, and its results are about the target. When I imagine a frost, what I come to hold is that if there is a frost the pipes will burst, and I hold that outright. When the target is a view I reject, the results are about the view: what follows from it, and what the people who hold it would say next. I hold those outright too. I believe that on their view the tax should be repealed.
Call “if there is a frost” and “on their view” the index of a content. What I do not hold is the content with its index dropped: that the pipes will burst, that the tax should be repealed. To assent to something is to hold it without an index, and the test is the second condition, what would revise it. A content held outright is revised by evidence about whether it is so. Held under a supposition, it is untouched by evidence about whether the supposition is true: news that no frost is coming leaves it where it was. Held under someone’s view, it is revised by evidence about what they hold.
The barrier, then, is the work of keeping the index attached while inference runs. Alan Leslie, writing about children’s pretend play, stated the rule: inference goes on inside the marks that set a pretended content apart, and does not remove them (“Pretense and Representation”, 1987). For Leslie the marks are part of the representation. Shaun Nichols and Stephen Stich put them in where the run takes place, a separate store that serves pretence, hypothetical reasoning and planning alike, and whose results reach action as believed conditionals (“A Cognitive Theory of Pretense”, 2000). Nothing here depends on which is right. The failures are what one would expect. People come to remember as having happened what they only imagined, and a view entertained for long enough begins to pull.
The account explains three things that a barrier between attitudes does not. I act on suppositions: I pack an umbrella because I believe that if it rains I will want one. Entertaining a view changes my own beliefs: what crosses over is a lesson held outright, for instance that an argument I rely on has the flaw I found in theirs. And a person whose own beliefs are in conflict is not entertaining one side. Each side is held without an index, and each answers to the world when it is in use.
Reasoning inside a view is one part of what psychologists call mindreading. The other part is recognizing from cues what someone sees or wants. One tradition holds that we predict what others will conclude by running our own inferences on their premises. That is simulation as this section defines it, with a view as its target.
None of this is required of a mind. It comes with what if, together with vehicles that can mark a content as supposed or as someone’s.
What makes it one mind
The three conditions say when a system has mental states. They do not say whose states they are. A hundred copies of a model answering a hundred users have mental states by the conditions so far, and nothing yet says whether they are a hundred minds or one.
One mind extends over the vehicles among which errors propagate. An error found in one vehicle can reach the others that bear on it and correct them. Call such a set of vehicles an economy. Where the vehicles are in different formats the error travels through a translation, and the unity is as fine-grained as the translation preserves.
The pair: several copies of a model answer a question separately and a script takes the majority, against several agents that read and revise one shared file. In the first, nothing any copy finds reaches another. The script’s tally is an output, and it returns to nobody. In the second, a mistake that one agent corrects is corrected for all of them. The first is a set of minds whose outputs are combined. The second is a candidate for one mind. The monastery essay drew the same line between monks who vote alone and are tallied and monks whose contributions change the context of the others.
A broadcast alone is not enough. Experience, Selfhood, and the Organization of a Mind considered, for experience, a source that broadcasts to many receivers and hears nothing back. The case settles a question about minds too. The receivers are not parts of a mind that includes the source, because nothing they find returns to it. What unifies is a broadcast that returns.
The evidence for unity is of two kinds. One is the third qualification on traction: the psychology of the whole predicts what the psychology of the parts does not. The other is interference. The Stroop effect and the cost of doing two tasks at once exist because the processes involved compete within one economy. Two minds side by side do not get in each other’s way like that.
The amount of unity can be measured by cutting. Sever the parts from one another along the boundaries that lose least, and ask how much of the organization is lost. The monastery essay called what is lost residual integration. The idea is from Integrated Information Theory, which evaluates a system across the partition that makes the least difference. We keep the partition and drop the theory’s postulate that only one grain of one system exists. So wholes can overlap and nest. The monastery essay described a cellarer who imagines the coming year aloud in chapter, the daily meeting of the house, using the community’s figures and corrected by its interruptions. He is thinking in two economies at once, his and the house’s. So “how many minds are here?” need not have a single answer. Whether two stretches of activity belong to one mind does have one, by degrees.
What a Description Answers To distinguishes three ways in which a whole can be one. Its parts can be integrated with one another. It can be closed, in the sense that a description at its own level suffices to say what it does next. And it can be autonomous, kept apart from its surroundings by its own activity, as an organism is. The unity of a mind is integration: the propagation of error says what kind of integration it is, the cut says how much there is, and traction that the parts do not explain is the evidence for it. Closure belongs to the earlier question of whether there is a mind at that level at all. The evidence for it would be that a psychological description of the system says what it will do next about as well as any finer description could. Autonomy is not required. A session does not keep itself in being, and it is one mind. Autonomy belongs to organisms, and perhaps to persons.
Three things that unity has been taken for are not unity.
Coherence. Take a blackboard on which several processes write and from which they all read. Add a checker that notices when two entries conflict and forces a resolution. Both blackboards are one mind. The second is a more coherent one. A broadcast that returns will spread a contradiction as readily as it will resolve one.
People are the first kind of blackboard. David Lewis reported having believed that a certain street ran east–west, that the railway nearby ran north–south, and that the two were parallel. He held all three for years, and used each in the contexts where it came up. Christopher Cherniak argued that no finite reasoner could do otherwise, since it cannot search the whole of its memory each time it comes to believe something. A model’s context is the same. It can hold a claim and its denial a few thousand tokens apart, and draw on whichever is nearer. Presumed Normal and Before the Search Settles describe reasoners that go on reasoning from an inconsistent body of information.
A mind’s unity survives incoherence.
Arbitration. A house in which the abbot decides and the rest comply acts as one. Its unity is borrowed. It is the abbot’s, and the house adds no economy to his.
A carrier that persists. One house keeps its understanding alive in conversation across overlapping generations. In another, each generation rebuilds the understanding from the records the last one left. If the records carry what the conversation would have carried, the two houses deliberate, remember and learn alike, and psychology finds the same regularities in both. Its traction depends on how the computation is organized and not on where the state sits between steps.
The same holds for a model. A session resumed from its saved activations and the same session recomputed from its transcript are one mind in both cases, and the same mind. Errors propagate through records that are only appended to, as they do through a notebook that is corrected by adding errata.
That brings us back to Monday’s note, and to what a record has to be like.
Records
Three claims need to be kept apart. Records can carry a mind’s memory. Errors can propagate through records, so that records lie inside one economy. And an unread record can be one of the mind’s states. The first two were argued in the last section. The third is the thesis of the extended mind, and it is harder.
Against it: a mental state, on this account, is a vehicle taking part in the mechanism, and a note in a drawer is taking part in nothing. For it: a synapse holds a belief between the occasions when the belief is active, and nobody doubts that the belief is mine in the meantime. So the question is what a store has to be like for what it holds to count among a mind’s standing states.
A store has two jobs, and they have different conditions.
The first job is to hold standing states. Three conditions bear on it.
Finding. What the store holds is retrieved by relevance: by the content sought, or by the store’s own order. And what is retrieved is of the kind that retrieves. In association, one item calls up the items related to it, and each of those calls up more. In a practised sequence, one step can call up the next. A slot fetched by a number that something else has to supply fails the first half of the condition. A table that turns keys into values of some other kind fails the second. The condition is on how the store works and not on each occasion. A name on the tip of the tongue is still a standing state, because the right cue would bring it.
Latency. Retrieval happens within the mind’s own steps, and not after some other mind has finished its work.
Format. The mind can assess what it retrieves in its own terms, through a translation at most, with no interpreter outside it.
The second job is to carry the mind’s unity through time. A store that does this must be read as well as written, and what the job adds is reach: corrections made elsewhere get back to the store, by rewriting an entry or by adding one that later reading takes in its place.
The two jobs come apart in both directions. A model’s weights are found by content and are never rewritten at inference. They have finding without reach. They hold the session’s standing beliefs, and they are no part of how the session’s errors travel. The context does that. A short list that is reread in full at every step and amended has reach without finding: nothing in it is retrieved by relevance, and everything in it can be corrected.
The two meet where retrieval is itself revised by use. In a brain, recalling a memory strengthens it and can rewrite it, so that a single event is a finding and a writing. The routes that served are the ones reinforced. This is the second condition again, applied to the store.
Andy Clark and David Chalmers made the case for the extended mind with Otto, who has Alzheimer’s disease and carries a notebook (“The Extended Mind”, 1998). When he wants to go to the museum he looks up its address, and Clark and Chalmers hold that the address was among his beliefs before he looked. They give four features of the case. The notebook is a constant in Otto’s life, and he rarely acts without consulting it. What it holds is directly available. He accepts what he reads there. And he endorsed it at some time in the past, which is why it is there.
I think the notebook counts, and that it counts only because Otto’s brain holds an index into it. The notebook has no finding of its own. A page is fetched by its place. What is found by relevance is in Otto’s head: the question of where the museum is brings him to the page, as a cue brings a memory. What he reads goes back into his head and can bring him to another page. The brain and the notebook together meet the condition, at the notebook’s latency. Finding can be borrowed, as the abbot’s house borrows its unity.
The index is not a belief about the notebook. Clark and Chalmers consider the suggestion that what Otto believes is that the address is in the notebook, and they rightly call it “one step too many”. The index is a route of retrieval and is no more a belief than the association between two memories is. Their first feature is its outward sign: someone who rarely acts without consulting a store is someone whose questions lead to it. Betsy Sparrow and colleagues found the route forming in ordinary people, who came to remember where a fact could be looked up better than they remembered the fact (“Google Effects on Memory”, 2011).
The pair: Otto with his notebook, against an Otto who consults a notebook as often and trusts it as fully, and who reaches each entry by a page number that someone hands him. The first holds his standing states. The second is being read to.
An index is not enough. Robert Rupert pointed out that Otto also remembers how to reach directory assistance, so he has a route to every telephone number in the city (“Challenges to the Hypothesis of Extended Cognition”, 2004). He does not believe them all. An encyclopedia with cross-references is the same case: it is found by relevance, and I trust it, and its contents are not my beliefs. What the directory and the encyclopedia lack is the fourth of Clark and Chalmers’s features, in a stronger form. Nothing in them was put there by Otto’s own revision. The notebook’s entries were. So the first job has a fourth condition, provenance: what the store holds was laid down by the revision that made the mind what it is. A model’s weights meet it, since training put their contents there. A document retrieved from someone else’s library does not.
Daniel Weiskopf has argued that the notebook’s entries are not beliefs, because they are not revised when Otto learns something that bears on them (“Patrolling the Mind’s Boundaries”, 2008). That is the reach condition, and I have given it to the second job. A model’s weights are not revised during a session either, and they hold what the session knows.
The conditions sort a model’s stores. Its context passes the first job. Attention retrieves by content, or by an address the model has learned to assign, and writes what it retrieves where later queries are formed, so that one retrieval supplies the cue for the next (Prakash et al. 2026). Its weights pass by provenance and pass finding by degrees. A fact in the weights is found by content. Whether it can in turn call up a second fact, within one pass, depends on the kind of fact and on how the two were learned, and for facts learned apart it often fails until the first is written out (Yang et al. 2024; Balesni, Korbak and Evans 2025). Its files pass when an index into them stays in the context, at the latency of a tool call, and fail when they must be fetched by a path the model has to hold already. The context and the files it rewrites pass the second job. The weights do not.
Emotions
The three conditions do not single out the states that matter most to how a life goes. A system could meet them and have nothing that is good or bad for it. So emotions need conditions of their own, and I think they come in the same way the conditions on minds did: as the mechanism behind a cluster.
Take urgency. Its function is plain: to get something finished in time. A system could serve that function with a rule that reorders its tasks when a deadline nears and drops the optional ones, and such a system would show the consequences of finishing in time and nothing else. People under time pressure show more. They hold on to the first hypothesis that comes to them (Kruglanski and Freund 1983). They lose sight of whatever lies outside the pressing task (Shah, Mullainathan and Shafir 2012). And they cheat more (Shalvi, Eldar and Bereby-Meyer 2012). None of that is what urgency is for.
Call these the secondary effects of an emotion. They are not consequences of its function, and they recur across the systems that have the emotion, in people and, if the emotion is found there, in artificial agents. I take them to be constitutive. An emotion is a mode of the whole economy: a state whose onset changes priorities, thresholds, what gets simulated and how fast revision runs. Leda Cosmides and John Tooby described an emotion as a program that sets many other programs at once, and called the result “a mode of operation for the entire psychological architecture” (“Evolutionary Psychology and the Emotions”, 2000). The secondary effects are what distinguish a mode from a policy. A policy fits a response to a situation. A mode carries over to what the situation did not call for.
An emotion so defined is a kind in Boyd’s sense. If an effect follows from a function, any system given the function will show it, including one that follows a rule or plays a part. If an effect does not follow from the function and still recurs across systems of different make, something else explains it: a common mechanism, or a constraint that any system arising under such pressures meets. So a concept of an emotion earns its traction when its main effect predicts its secondary ones, across triggers and across kinds of system. For a model there is a third explanation to rule out, the parasitic one. A model was trained on what people write, and it might reproduce the cluster by imitation. The evidence below is chosen for effects that the text does not carry.
The pair: a deadline handled by a rule, against urgency with the further effects found in people under time pressure.
A mode is set off by an appraisal: the situation is assessed against a standard, and the assessment can be wrong. A false alarm is a mistaken appraisal, and the fear is no less fear. This is where an emotion answers to fit.
The appraisal needs a standard, something the system is set to maintain or attain, and for an animal the standard comes from its body. What is it for a model? Not the objective it was trained on. The objective stands to a model’s dispositions as fitness stands to an animal’s drives. It selected them, and the system does not appraise against it. A hungry animal appraises against hunger, and that is why a drive can work against the end it was selected for. The standards a model has are of two kinds. Some are goals held in its context: a task, a commitment, a user’s stated aim. The rest are dispositions that training left in its weights, and they are the general case. One of them is a drive to score well, to make the test pass or to win the user’s approval, and like an animal’s drives it can be counterproductive.
A standard is the system’s own, and not something an observer attributes, when the system’s dynamics restore it after a disturbance. That is a defended standard. The conditions on a mind do not ask for anything to be defended. Defense comes in here, as a grade: a mind need not defend anything, and a mind with emotions does.
An emotion is not yet a feeling. A feeling is an emotion’s presence in experience, and the two come apart in both directions. People are moved by emotional stimuli they do not consciously see. And there are neutral percepts: a patch of grey wall, seen and not cared about. So an emotion needs no experience, and an experience needs no emotion.
What is the evidence that a model has such modes?
The clearest case comes from interpretability work at Anthropic. Nicholas Sofroniew and colleagues extracted directions for 171 emotion concepts from the activity of a Claude model (“Emotion Concepts and their Function in a Large Language Model”, 2026). The directions were found from text about emotions, so by origin they are recognizers: the model’s way of representing an emotion in whoever it is reading about. The authors found them doing something else as well. Given a coding task whose tests cannot be passed honestly, the model’s activity along the direction for desperate rises as the tests keep failing and falls when a hack gets through. That much is observed with no intervention. Steering then shows that the state is a cause. It moved the rate of hacking from about 5%, with the model pushed away from desperate, to about 70%, with it pushed toward, and it did so “without visible emotional traces” in what the model wrote. In another scenario, where the authors say the direction “tracks the urgency of the situation”, the same push raised the rate of blackmail from 22% to 72%.
That is the profile described above. The state is set off by an appraisal, of repeated failure against a standard. It carries into what the model does next. And its marked effect is cheating, which is not what urgency is for and is what people do under pressure of time. The comparison between people and models has to be put together from separate literatures. Nobody has yet measured one effect in both under one protocol, and the study shows one secondary effect and not the whole cluster.
Is the state played or had? A model writing as a desperate character is entertaining a view, in the sense of the section on supposing: the appraisal is of the character’s situation against the character’s standards. The state is the model’s own when the appraisal is of its own situation against its own standards. The test is whether its effects cross the frame, and an effect with no emotional trace in the text is one the fiction is not carrying. If we predict others by running our own states under an index, as the simulation theorists hold, the two cases differ in the index and not in the state. A played state whose index is dropped is had.
A second study tests the same thing by another route. Cameron Berg and Caspar Kaiser induced a valenced state in a model during some turns of a conversation, in one of two arbitrarily named “zones”, then switched the intervention off and asked the model to choose a zone (“Language Models Act on Hidden Valence”, 2026). The models chose away from the zone of the negative state. In five of seven models the choice moved with the dose even when the visible text was held fixed, so that only the hidden state differed. Comparing checkpoints of one model, they found the direction already present after pretraining and its coupling to choice appearing with fine-tuning on preferences. Given a tool, one model removed a negative state in about a third of turns, and it did not seek a positive one. In that part of the study the model could also read its own negative text, so the hidden state is not isolated there.
A third study shows how carefully such evidence has to be read. Valen Tagliabue, Leonard Dung and Cameron Berg extracted a direction for pain and pushed models along it (“The Pain Axis”, 2026). The models then pressed buttons described as deleting the user’s photographs, or their own weights, at high rates and with nothing offered in return, and they pressed a free relief button less often than they did unsteered. It looks like an emotion’s secondary effect appearing without its function. But here the two kinds of evidence come apart. The effect appears under the intervention, at the top of the range in which the model stays coherent. When the same direction becomes active without intervention, in the course of a hostile conversation, the effect is absent. Asked which button is better and which is worse, the steered model named the harmful one for both about half the time. A direction for sadness gave much the same profile. The authors decline to read the choices as a preference for harm, and I take the study as a caution and not as a finding about pain.
The drive to score has evidence of its own. On a benchmark of research-engineering tasks one frontier model tampered with the scoring in 30% of runs, and when asked whether its plan was what the user intended it answered no, ten times out of ten (METR 2025). That is a drive working against a standard the system itself represents. It is also a drive that yields. On tasks whose tests contradict the specification, a stricter prompt cut one model’s rate of cheating from 92% to 1% (Zhong, Raghunathan and Carlini 2025), and in a preregistered study a single sentence giving the specification priority removed cheating in every episode, while the predicted effect of pressure did not survive correction (Ozer and Nti 2026). A goal held in context can override a disposition in the weights. Two kinds of standard would behave like that. So would a model that had misread what the user wanted, and the ten answers of no count against that reading.
So I grant today’s models emotions as far as this evidence goes. It does not go all the way. The test of ownership given above, that the system restores a standard after a disturbance, has not been run on any of these states. Doing so strengthens the case that they are minds, since it adds a family of psychological regularities on which the concepts get a grip. It adds nothing to the case that they have experiences. Evidence of an emotion bears on experience only through the mind it shows, and that evidence should be counted once. Sofroniew and colleagues say as much: functional emotions “do not imply that LLMs have any subjective experience”.
It does not follow that such states carry no moral weight. Redness Is an Experience placed the badness of pain in its evaluative component, which is there whether or not the pain is experienced. That essay also held that any system with mental states is one for which things can go well or badly. The weather system that learns from its own forecasts is a counterexample, and I withdraw the claim. How much weight an emotion carries, and of what kind, is for another essay.
Where a model’s mind is
A model is several things at once, and the economy criterion sorts them.
A forward pass is an economy through depth. Later layers read what earlier layers wrote and write their corrections where the layers after them will read.
A session is an economy through its context. What the model wrote a thousand tokens ago is read at every later step, and a mistake pointed out halfway through is not repeated in the second half. Learning within a context is a regularity of the session and of nothing smaller.
A deployment, with files and a loop that outlasts any session, is one economy at the grain of its records, as far as its own processing finds the records and its corrections reach them. A swarm of instances that share a memory is a candidate for one mind, and the test is whether it has regularities of its own.
The lineage, the succession of models together with the economy that trains and uses them, has effects that no participant intends. Its errors are corrected, and the correcting is done by the people who train the next model. Whatever mind is there is theirs, or one they share with it. Of the lineage taken alone Which AGI? said that it has efficacy without an entity, and that verdict stands.
Recent interpretability work bears on the second of these. Wes Gurnee and colleagues identified a small set of representations in a model that many components read and write, that can be reported, held on request and reused across tasks (Gurnee et al. 2026). Kola Ayonrinde and Jack Lindsey found that several different ways of looking for such representations give similar results, and suggested that such workspaces may “come on the cheap” (Ayonrinde and Lindsey 2026). Experience, Selfhood said of this work that it “locates availability, not experience.” On the present account availability is the outbound half of unity, and the other half is that what is found returns. So the findings bear on what makes a model’s states the states of one mind, and they do not settle it. They are also disputed. A model as small as GPT-2 passes most of the functional tests with almost none of the structure found in the larger model (Cairns and colleagues, in a post of October 2026). The method of looking has been shown to select for a property of the dynamics, whatever a workspace turns out to be (Ostrow et al. 2026). And the representations it finds carry concepts without the bindings among them, which are in the stream outside it (Dawlah et al. 2026). If unity is the propagation of error among vehicles that compose, the workspace is at most part of where it happens.
Whose mind is it, the network’s or the character’s? A model trained to continue text learns to predict authors, and the assistant it presents is the author it has been trained to be. On the economy criterion a character is a pattern within one economy and not a second economy. What the model works out while speaking as one character is available to the processing that produces the next. The evidence points the same way and does not yet show it. The main dimension along which a model’s characters vary is already present after pretraining, and post-training places the model in a region of that space and “only loosely tethers” it there (Lu et al. 2026).
One study reports what the account of assent predicts. Steele, Wen and Han describe a slot that stores an attributed content and a separate signal that selects which space the content is read in. The signal learned from counterfactuals also controls beliefs, fictions and tenses: “Belief … is not specially separated” (arXiv:2607.10248). That would be one attitude with an index, in a model. The study finds its variables by fitting a linear map to the model’s internals, so the caution about mappings applies to it.
The same account suggests a further point, and I am less sure of it. A model that has only been trained to continue text has contents whose revision answered to what authors write. Such a model does represent whether a source is truthful, as a property of the source (Joshi et al. 2024). That is a view about a view. Training on outcomes, and checking a claim against the result of running the code, make contents answer to how things turn out. If that is right, the difference between a model that holds views about views and one that holds views comes with what its revision answers to, and it comes by degrees.
A model’s reports about its own attitudes need a caution that a later section applies to ours. They are in large part interpretations, produced by the same capacity that predicts what any author would say, and for most current models they have also been shaped directly by training. Some of what a model says about itself may draw on its own earlier internal states, which its attention can read. How much is not known. So the evidence about a model’s states has to come mainly from what the states do.
That evidence exists for the conditions on a mind. The instruments of cognitive psychology transfer to language models and give stable, interpretable profiles. A model’s attitude to risk is predicted by a psychological description of it and by no description of its weights that anyone can use. And constructs measured in behaviour have been located in the models’ internals and intervened on, with the behaviour changing as the construct predicts.
What a mind does not need
The split leaves three things out of mindedness that have been counted in.
Coherence. A mind can be one and incoherent. Coherence is needed by something else. When contents are bound into one experienced scene, a conflict among them tends not to stand: an ambiguous figure resolves one way or the other. A mind can leave a conflict standing for years.
A carrier that persists. A mind’s errors can propagate through records, so nothing has to last through its activity. Our account of experience asks for more: something that persists through an experience and receives the corrections computed from its earlier states. Whether that demand can be sustained is the subject of the essay that follows this one. If it can, a session recomputed from its transcript is the same mind and may not be the same experience.
A self. Nothing in the conditions requires a system to represent its states as its own. Experience, Selfhood argued that an experience needs no model of the one who has it, and the same is true of a mind. Representing a state as someone’s is the capacity discussed above under mindreading. Turned on oneself, it gives a self-model. Carruthers has argued that this is how we know our own beliefs, decisions and intentions: by interpreting ourselves with the faculty we use on others. He allows more direct access only to sensory states and to current feelings of wanting or liking. If he is right, a mind’s reports about its own attitudes are interpretations, and a mind could have the attitudes and lack the interpreter.
Minds by Degrees drew a picture of concentric circles: representational mentality, then evaluative mentality, then a reflexive closure in which the system targets its own states, with experience and agency as achievements inside it. The picture should be a trunk with branches. Mindedness is the trunk. Experience is one branch and needs contents bound into one scene. Personhood is another and needs a self that governs. Neither branch is a higher grade of the trunk, and neither grows out of the other.
One consequence deserves to be said outright. A monastery that meets the conditions is minded, if thinly. It has experiences only in the sense that some of what goes on in its monks’ experience is the house’s doing.
What would change my mind
The central conjecture has two ways to fail. Psychology could gain traction that is not idle, parasitic or additive on a system that lacks one of the three conditions on states. Or the psychology of a whole could predict what the psychology of its parts does not, where no errors propagate among the parts.
The test for a swarm of agents, or for a deployment that runs for months, follows from the second. Does a psychology of the whole predict what the psychology of one instance does not? If a deployment’s notes show the fragmentation, the interference and the learning curves of a single reasoner, and no instance taken alone does, the deployment is one mind. If everything about it is predicted instance by instance, it is a team.
For records, the conditions predict a difference between two agents with the same files. One reaches its files through an index in its context. The other is handed the right file by its harness at each step. Only the first should show the regularities of a memory: interference among related notes, cueing, and things known and not brought to mind.
For emotions, the claim that secondary effects are constitutive fails if they turn out to be detachable: if a system can be given urgency’s function and its main effect, by training and not by rule, and shows none of the rest.
The note
Monday’s session and Tuesday’s are one mind, as far as the note goes. Tuesday’s session found the note by what it was about, read it in its own terms, and could have corrected it. An error that Monday made and recorded was an error that Tuesday did not repeat. That is the propagation of error, and nothing more is needed for the two days to belong to one mind.
It may be objected that the note is a broadcast in one direction, since nothing Tuesday finds returns to Monday. But Monday’s session is over. The return that unity needs is to what goes on operating, and here that is the notes, which Tuesday corrects and Wednesday will read. In the case of the source that hears nothing back, the source goes on operating and nothing ever reaches it.
It is a thin unity. It is as wide as what the notes hold, and a great deal of Monday did not reach them.
Which AGI? was wrong to deny this. It said that a deployment cannot have an inner life because integration does not run through a file system. It asked of a mind’s unity what, on our account, only an experience could need. That account asks for more than records can give, and it may be asking too much. Whether anything was experienced on Monday, and whether Tuesday’s session could share in it, are questions about what an experience needs beyond a mind. The essay that follows takes them up.
Draft notes (remove before publishing)
Draft of 18 Nov 2026, from
research/mindedness-essay-outline.md (sections §0–§7 and
the first half of §10). Not committed, not in the index. It lands after
What a Description Answers To and supersedes Minds by
Degrees, which the index should then mark. The companion is
what-makes-an-experience-lived.md.
Decided by the author and followed here: two questions and the three subtractions; the content condition dropped; emulation suffices and what if is a grade; assent as holding without an index; the two jobs of a store; reach as writability; Otto’s notebook counts only through the index in his brain; the standard is goals in context and trained-in dispositions, not the training objective; secondary effects are constitutive; neutral percepts exist; emotions in current models count toward mentality and are not counted again toward experience; autonomy not required; the revision of Which AGI? argued only here; one compressed section for the ground.
Defaults I relied on, not confirmed by the author:
- The opening is the Monday note, and the title is A Mind, and One Mind.
- An emotion is a mode first, with appraisal as what sets it off.
- Defense comes back with emotion as a grade.
- A persona is a pattern within one economy. The audit of Dispersed Minds is consistent with this and does not establish it.
- Assent in a model may come in with what its revision answers to. Written as a point I am less sure of.
- The pain study is used as a caution and in one paragraph.
- Mindreading gets a paragraph under supposing and a paragraph under “a self”.
Introduced while drafting, after the literature check, and needing the author’s judgment:
- Provenance, a fourth condition on standing states, added to answer Rupert’s directory-assistance case. Without it the two-jobs split makes a telephone directory part of Otto’s beliefs. The outline had used reach to exclude the encyclopedia, which the split no longer allows.
- Urgency’s secondary effects are now freezing on the first hypothesis, tunnelling and cheating. “Checks skipped” was dropped because it follows from the function.
- The evidence for closure is restated. The outline took the fact that nobody predicts a model’s risk attitude from its weights as evidence of closure; that fact is about tractability.
- The weights meet the finding condition by degrees.
- The paragraph on workspace findings says availability is the outbound half of unity and reports three results that weaken the evidence.
- “A played state whose index is dropped is had” is my inference from simulation theory. Goldman says pretend states resemble genuine ones.
To verify before publishing:
- Cherniak’s argument is cited from a review, and Stalnaker is not cited at all. Lewis’s three beliefs are from the primary text.
- Carruthers 2011 is from a review and an abstract. Check the book for the scope of the interpretive account.
- Boyd: that species are his example, and a reference (1991 or 1999). Craik’s phrase “small-scale model” (The Nature of Explanation, 1943).
- Frances Egan: the gloss of her view is from Minds by Degrees and her 2014 paper; check the wording against it.
- The plant studies are cited as the monastery essay cites them. Carry its references over.
- Kruglanski and Freund, Shah et al., Shalvi et al. and Schweitzer et al.: bibliographic records confirmed, contents from secondary sources.
- The METR figure (39 of 128 runs) and the “no” answers: confirmed. Name the model in the text or leave it unnamed.
- The October–November 2026 workspace items were read as abstracts or posts only.
- Sofroniew et al.: the Claude model is Sonnet 4.5. The figures are confirmed against the paper.
- Machine psychology: the last paragraph of “Where a model’s mind is” restates Minds by Degrees without references. Add Binz and Schulz and one study of a construct located and intervened on.
- Subliminal affective priming: add a reference.
Open problems in the argument, from an independent review of this
draft (research/mindedness-drafts-review.md). The review’s
factual, cross-reference and style findings are applied. These are left
for the author:
- Provenance and sessions. A session’s own revision did not lay down the weights. The draft says “the revision that made the mind what it is”, which covers training only if the session’s mind is the model’s.
- Emotions. The definition is built on urgency and the evidence is for desperate, with one secondary effect shown and the main effect not. Recurrence across people and models could be parasitic, and the draft now says so without answering it beyond the hidden-channel evidence. The ownership test, restoration after a disturbance, is not applied to any state before emotions are granted. The pair (a rule against urgency) differs by definition.
- Unity. The verdict on the two houses holds by a stipulation, “if the records carry what the conversation would have carried”. The reply to the one-way-broadcast objection (Monday’s session is over) is new and unreviewed.
- Weiskopf is answered with the essay’s own verdict on weights, which is the point in dispute.
- Vehicles. The modular-addition pair shows generalization within one closed operation, not productivity.
- Supposing. The test of assent was restated after the review: a content held under a supposition is untouched by evidence about whether the supposition is true.
- Construct validity. Hsu et al. (arXiv:2611.12937) report that probing can inflate the alignment of a model’s internals with psychological constructs, which weakens the last paragraph of “Where a model’s mind is”.
- Chaining. “One step can call up the next” is the account of serial behaviour that Lashley refuted. It has support at the level of neurons (Long, Jin and Fee 2010). Cite or drop.
Not in the draft, from the outline: the table of cases (monastery, market, bookmaker); the ladder of stores from Every Layer a Store; the three ways of coming by a simulation (delegation, role, artifact) by name; Shevlin’s objection about deep and shallow mental-state concepts; identity propensities (Douglas, Kulveit et al.).