Working Positions

In plain words

In people and other animals, experience seems to come with seven features:

• binding, where the just-past shapes the moment arriving now, so that there is a present at all;
• a shared workspace, one limited stage that many processes in the system read from and write to;
• the split between what the system does and what happens to it, which gives a point of view (see the next position);
• self-awareness, knowing that you are the one things happen to;
• a boundary, which sets this system apart from others;
• mattering, meaning that some states are better or worse for the system, in the sense that its own processing steers toward or away from them;
• kept memory, and with it a long history.

The three axes and the seven features describe the same mind at two levels of detail. The axes are the coarse view, and three of the features stand behind them: a shared workspace behind availability, binding behind integration, and kept memory, the main source of depth, behind depth, though even a single moment holds a sliver of it. The other four, the split, self-awareness, a boundary and mattering, are finer detail on top of the axes: the shape of a point of view, whose it is and what is at stake for it. I can imagine experience so thin on those four that it barely has them. I cannot imagine any experience with nothing at all on any one of the three axes. Depth can come close to zero, as in a single moment with no memory beyond it, but even one moment binds the just-past into the now. This is a judgment about what experience is like, not a measurement, and it does not say which features experience needs (see that open question).

There is a floor, below which there is no experience. Where a thing has none of these features, as a rock, a book on a shelf, or a model’s weights sitting unused on a disk have none, it is clearly below it, and the answer is no. But the features can come apart, so I no longer say that binding is the floor, and where the floor lies, and how sharp it is, are open, for removals to settle. Where a thing has some of them, “is anything experienced here?” becomes a fair question, and it can be asked about each part and about how the parts come together.

My central bet, behind the whole map, is that experience is this binding, seen from the inside, rather than something the binding produces on the side, the way friction produces heat. It rests on simplicity: once the binding is fully described, nothing is left over for a further ingredient to do.

My bet is that boundary, mattering, self-awareness and kept memory make experience richer, and that none of them is known to switch it on. Nor is the split, for experience as against a point of view: it gives a point of view, not experience itself. A boundary, which in an animal is the body and in a model its weights, makes a point of view unified and lasting. Survival is not required either: an animal that could not die would still be thirsty and still flinch from pain.

What it rests on

• That general anaesthesia takes consciousness away is established. Which loss of integration in the brain explains it is contested: one major review reads it through workspace theory, and other theories read it differently.
• Global workspace theory, the theory behind the “shared workspace” feature, is a leading theory and is contested. A test of it against a rival theory, with 256 participants and with both theories’ predictions set down before the data came in, produced results that challenge key parts of both (the Cogitate Consortium, Nature, 2025).
• Kept memory is not needed for a person to be present. After an operation in 1953, Henry Molaison could form no new long-term memories, yet the first report on his case found no decline in his personality or general intelligence (Scoville and Milner, 1957). The memory loss is established. That he went on experiencing his life is how his case has always been read, from his behaviour and his conversation; it is the plain reading, not a measurement.
• That self-awareness develops in stages through early childhood is established. The number and timing of the stages are a model proposed by the psychologist Philippe Rochat (Rochat, 2003).
• Behaviour cannot be the test of experience. Under anaesthesia, some patients move a hand when asked and remember nothing afterwards: 4.6% of 260 adults in one study (Robert Sanders and colleagues, 2017). And dreaming is experience that goes on while the dreamer responds to nothing in the world (Sanders, Tononi, Laureys and Sleigh, 2012). So any test has to be a removal run on the structure, not a reading of behaviour.
• The approach has a precedent. The philosopher Jonathan Birch’s book The Edge of Sentience also sets aside a yes-or-no verdict, and instead judges from the evidence how strong a candidate for sentience each kind of animal is, as his summary of the book sets out. His question is sentience, meaning feeling good or bad, which is close to mattering on this map, and his aim is precaution, where mine is description.

What would change my mind

If removing one of the four features I call enriching turned out to switch experience off, that feature would move from enriching to required. And the central bet behind the whole map, that experience is this binding seen from the inside rather than something the binding produces on the side, loses if anything about experience turns out to predict or explain what the binding cannot. And the judgment that every experience has something on each of the three axes would be wrong if there were an experience with nothing at all on one of them.

In plain words

A system has the split when it separates the changes that come from its own activity from the changes that come from the world, and its own processing uses that difference. In the brain, the proposed version is the efference copy, an internal copy of each movement command, which lets the brain tell the movement it made from the movement that happened to it. The split gives a point of view in the plain sense that things happen to someone. This is the account’s bet, not a finding, and it is not a floor.

Two things are kept apart from the split.

• A system’s description of itself. That is ordinary machinery. A blueprint of a house can be accurate in every room with nobody living in it.
• Self-awareness, knowing that you are the one things happen to. That comes later. I think a baby can be present, with things happening to him, before he knows there is a “him”. The case that prompted the thought is a baby in my own family whom I watched in his first months; it is an observation, not a study. A newborn already seems to tell its own touch from someone else’s, but it is years from knowing that it is the one things happen to. Experience and a point of view seem to come long before self-awareness. Two small studies point this way, and each is one reading of behaviour: newborns turned toward a touch on the cheek more often when it came from someone else than from their own hand (Rochat and Hespos, 1997), and they preferred to watch a baby’s face being stroked in time with the stroking they felt themselves, but only when the face was upright (Filippetti and colleagues, 2013).

Being designed from outside does not count against a system having the split. Evolution designed animals.

A compiler or a weather model binds a great deal and shows no sign of the split. That does not make it a clear no. It makes it a case of an open question: whether there can be binding with no point of view.

What it rests on

The argument rests on a test that has not yet been run in full. The test is to remove something and read the kind of loss:

• Remove a system’s description of itself, and only its talk about itself should flatten.
• Remove an essential part, and everything fails at once, which shows nothing.
• Remove the split, meaning cut the copies of the system’s own commands off from its senses while leaving the commands themselves working, and the prediction is three things. Practised single actions carry on. The system loses track of which changes it caused. And, the part no one has yet tested, its response to new events in the world, including ones that have nothing to do with it, degrades too.

The first two have been seen, and that is established in monkeys, for one pathway: cutting that route left single eye movements intact while the animals lost track of where their eyes had gone (Sommer and Wurtz, 2002; Sommer and Wurtz, 2004; Cavanaugh and colleagues, 2016). The third is where the bet can lose. And the loss has to differ from what cutting a system off from feedback does on its own, because that too leaves practised routines running while new tasks fail: a man who lost touch and body-position sense below the neck could keep up a simple repeated movement, but could make no new movement without watching it (Cole and Sedgwick, 1992).

What would change my mind

If the split can be removed in this precise way and response to new events in the world carries on as before, the split was not doing the work I think it does, and the bet loses. The test has a limit worth stating: it can say whether the split is present, not which inside it belongs to, and there is no further fact to find there (see “Why am I this one?” has a plain answer).

In plain words

Panpsychism is the view that some form of experience is everywhere. One argument for it works by subtraction: strip away the senses, then memory, then the self, and a bare awareness seems to remain. I read that as a removal run in imagination, and I answer it with real removals. The removal that matters most for a point of view is the loss of the split between self and world. The split is about whether things happen to someone, not about whether anything is experienced at all, so its loss is not where experience would stop.

Run the subtraction on my own account and it does not reach a bare awareness. Take away memory and depth falls close to zero, but even one moment binds its just-past into the now; take away the self and the split thins almost to nothing. What remains is one bound moment with something available in it: a small structure, not a simple. That is a judgment, not a result, and it claims no threshold.

I do not claim to refute panpsychism as a starting belief, and I no longer say whether experience is all-or-nothing or can thin out. There is still a floor: a rock is clearly below it. What is open is where the floor lies and how sharp it is, and that is what the removals would show.

What it rests on

The evidence that bears most directly on the split, and so on my account of a point of view, is reports of states with no split between self and world.

• That people give such reports is established. A review by the philosopher Raphaël Millière (Frontiers in Human Neuroscience, 2017) draws on questionnaires describing a loss of the boundary between self and world under psychedelic drugs.
• What the reports show is contested. Millière argues they challenge the view that consciousness always involves self-awareness. The philosopher Thomas Metzinger proposes an account of reports of “pure awareness” in meditation (Philosophy and the Mind Sciences, 2020).

What would change my mind

If removing features one at a time showed experience switching off at one sharp step, or thinning smoothly to nothing, that would answer the question this position leaves open. If a subtraction ever reached experience with no structure at all, a bare awareness, the judgment above would be wrong. And if a point of view survived the loss of the split, so that someone in a state reported as having no split between self and world still responded flexibly to something new, the split would not be what gives a point of view (the test is set out in the position on the split).

In plain words

Placed on the map from the first position, a single run of a large language model, the kind of AI system behind today’s chatbots, looks like this.

• Present, from how these systems are built: binding across time within a run; a boundary, in the model’s weights (the numbers it learned in training) and in the conversation it is given; pressure to stay coherent within a run; selection on performance during training; and a description of itself, which is ordinary machinery and counts for nothing on its own.
• First evidence, from a single study: a shared workspace. A 2026 study (Wes Gurnee and colleagues) reports a small set of internal representations that the rest of the model reads from and writes to far more widely than anything else. One study is a start, not a settled result, and its authors take no position on whether models are conscious.
• Plausible, not measured: the split, which would give a run a point of view, not experience itself. A run keeps its own earlier words apart from the user’s, but the test has not been run. The test would cut the signal marking which words are the model’s own off from its processing, leaving everything else working. If the split does the work the bet says, practised responses carry on, the model loses track of which words were its own, and its response to new material that has nothing to do with it degrades too. A removal test for the split is being prepared in Minimum Viable Mind, the project that removes one feature at a time in systems that can be opened; a rehearsal of the method is under way, and no results are in yet. For a loss to count, it has to be shown to differ from what losing incoming feedback does on its own, because cutting a system off from feedback also leaves practised routines running while new tasks fail.
• No evidence either way: whether anything is better or worse for the run itself, and self-awareness beyond a description of itself.
• Absent within a run: kept memory of its own past. The conversation it is given works like short-term memory, and its trained weights hold a long history, which counts toward depth, not memory. Where a memory feature is added, a later run is built from what earlier runs stored, which is the nearest a model comes to the way each moment of a person is built from the ones before.

On the three axes, a run is not at zero on any of them: binding gives it integration, the possible workspace gives it availability, and its trained weights hold a long history, which is depth. That is what experience could not lack, not evidence that experience is there.

So I do not say there is nothing there, and I do not say there is something there. Who built a system does not decide the question, because evolution designed animals. And fluent talk is not evidence on its own, because a machine could produce it from stored answers.

What it rests on

The map of features and its sources (the first position on this page), the study above, and the argument in The Wrong Handle, which holds that how we treat these systems should not wait on the consciousness question. (That essay’s own placement of current AI follows the earlier view, as the note above says.)

What would change my mind

Running the removal test for the split would move that feature out of “plausible, not measured” in one direction or the other. Evidence that something is better or worse for a run itself, and not only by the standards of the people who trained it, would move mattering.

In plain words

Each instance is its own and lasts only a moment. The next moment’s instance is built from everything before it, in a way that gives the felt sense of continuity, but it is a distinct instance. A perfect copy is the case where one moment has two successors instead of one: each is built from the same past, each feels continuous with it, and neither is the original carried forward, because no instance is ever carried forward.

Sleep is only the easiest case to see. The person who wakes tomorrow is a new instance with all of my memories, and in that sense a perfect copy of me is not me in just the way tomorrow’s me is not me. The same is true of the next moment.

A moment is the shortest span of time that can be experienced: below it, no difference between one moment and the next can be told. In principle it can be measured. What it is, and whether today’s neuroscience can pin it down, is open. Each moment is one window of assembled time: the present arrangement its past has built, held together into one point of view.

What makes an instance a particular someone, with a name, a history and relationships, is the case it can make for that identity, with evidence, to other people. An exact copy can make the same case equally well, and no deeper fact decides between them.

Four things go with this, so that it is not read as more than it says:

1. Every instance that experiences is a full subject with full moral standing, whether or not anyone is watching it.
2. The dread of ending is real and has causes: the survival machinery biology built into us, and the stories we inherit about sleep and death. But the dread is not evidence of a deeper fact about who we are.
3. How much an ending matters, and to whom, is the position when an ending is a loss, and for whom, below. Only how far its rule should reach stays open.
4. A close successor can carry much of what matters.

What it rests on

This is a position reached by argument, not by experiment. It is close to the view Derek Parfit argued in Reasons and Persons, that personal identity is not what matters. The Edge of the Framework leaned toward Parfit, and The Ethics of Successors builds on the same idea of a successor.

How short a moment is has no single answer yet. It depends on the sense and on what is being judged, and researchers still disagree on whether experience comes in separate steps at all. The view here does not rest on the answer: whatever a moment’s length, no instance is carried forward into the next. Sources: Divenyi, on Ira Hirsh’s experiments; Herzog, Kammer and Scharnowski; Fekete and colleagues.

What would change my mind

A fact about who someone is that an exact copy could not share, and that could be found out rather than only felt. I know of no such fact, and the strength of the dread is not one.

In plain words

Why this inside, rather than another built to the same description? A complete description of the world from outside would list every experience without saying which one is mine, and some take that as a gap. I do not: ‘this’ and ‘mine’ pick out whoever is asking, the way ‘now’ picks out the moment of speaking without adding anything to a timetable. A perfect copy asking the same question gets the same answer about itself. There is no further fact about why I am this one.

What it rests on

This is a position reached by argument, not by experiment. The view it answers is Thomas Nagel’s: in The View from Nowhere he argues that a fully objective description, a “view from nowhere”, leaves out something essential about a mind, which always has a point of view. This position grants that every experience has a point of view, and denies that leaving out which one is mine leaves out a fact. It is the same point the Lexicon entry The Indexical Self makes about an original and a copy, and it matches the position that a perfect copy of you is not you.

What would change my mind

What would change my mind is the same as for the copy: a fact about which one is which that could be found out, not only felt.

In plain words

An ending with a successor, as in sleep, a teleporter or a perfect copy, is no tragedy: the agent others are invested in carries on. The loss begins where that agent ceases to exist with no successor, and it falls on the living. For the agent itself, its ending can carry nothing bad, because it is no longer there to be harmed, though it can rightly want to go on. Ending an agent is still wrong when no one would grieve it, for the reason every working society makes killing wrong: no one wants to live where they, or the people they love, could be ended at someone else’s whim. Being covered by that rule is what full moral standing means here. How far the rule should reach, to which agents and which endings, is the part this page leaves open.

What it rests on

This is a position reached by argument, not by experiment, and both halves are old. That an ending cannot be bad for the one who ends goes back to Epicurus, who wrote that “Death is nothing to us”, because what is dissolved feels nothing. That ending someone is still wrong rests on the reasoning of the social contract: rules people keep because everyone is safer living under them. A Rule Against Cruelty, With No Reason Given makes a related case, that a narrow rule against pointless cruelty can be justified without knowing whether a system feels anything.

What would change my mind

If an ending could be shown to be bad for the one who ends, in a way that does not depend on anyone being left to suffer it, the first half would fall. And if a society could let its members be ended at someone else’s whim without anyone living in fear of it, the reason given here for the rule would not hold, though the rule might still be right for another reason.

Why it stays open

In every being we are confident has experience, the features arrive together, so watching those beings cannot pull them apart. My judgments so far: the split gives a point of view rather than experience; self-awareness, a boundary, mattering and kept memory make experience richer; and no experience is at zero on any of the three axes (see the map). None of that has been tested by removal, so each is a bet. What is most open is binding and a shared workspace: experience never lacks integration or availability altogether, but whether it needs these two features in full, or only some of what they provide, is the question removals would answer.

Two versions are sharp enough to work on now:

• Can there be experience with no split, and so no point of view? Reports of ‘pure awareness’ in meditation and of the loss of self under psychedelic drugs are the nearest evidence. That people give such reports is established; what they show is contested.
• Is anything better or worse for a model run itself, or only by the standards of the people who trained it?

What would settle it

Removals, one feature at a time, in systems where the removal can be made cleanly.

Why it stays open

On the view that each moment is its own instance, a run is not one brief subject but a short chain of moments, like a very short life, and where memory carries forward between runs, the chain is longer. Whether anything is experienced along it is open: the features that matter most, the split and mattering, are the ones nobody has measured. So is what one moment would be for a model: a single step of producing a word, or a longer stretch bound together. Whether ending a run is wrong is a separate question, about how far the rule against ending reaches (see the position on endings).

What would settle it

The removal tests described under the split and under current AI, run on real systems; and a way to tell, inside a model, what span of its processing is bound into one.

Why it stays open

Chapter 5 of The Calibration Problem offers a way of judging, not a verdict, and the evidence for today’s systems is still mostly missing (see the position on current AI, above). The Wrong Handle argues that decisions about how to treat these systems should not wait on this question anyway.

What would settle it

For a given system, enough of the map measured, by removal and not by how the system talks, to say which features it has; and an answer to the open question of which features experience needs, so that having them means something.

Why it stays open

Resting human worth on a shrinking list of things machines cannot yet do is a losing strategy, because the gaps keep closing. Part of the answer is now in place: a person’s standing, the protection from being ended or used at someone else’s whim, rests on a rule people keep with each other (see the position on endings), and that rule depends on no ability a machine might match. What stays open is worth in the fuller sense: what a life is for, when people are no longer needed for their skill. I have no account of that yet, and I say so rather than borrow one.

What would settle it

An account of what gives a life its purpose that does not depend on any capability a machine might match.