Eric Schwitzgebel
You can't say what shape the coin on your desk looks — and we're about to decide whether machines have rights

1 · One strange thing

Put a coin on your desk and look at it from an angle. Does it look elliptical? Look out your window at a row of receding streetlights. Do the farther ones look smaller?

From Malebranche in the 17th century to contemporary philosophers of perception Michael Tye and Alva Noë, philosophers have said yes for three hundred years. It has been treated as a basic fact about visual experience — the sort that needs no argument.

In May 2006, the first month of The Splintered Mind, Eric Schwitzgebel decided to work out the geometry of the thing everyone agreed about. His starting doubt was unusually plain: nobody can actually say what the geometry of this "looking" is supposed to be.

The account Tye and others give is projection: interpose between yourself and the object a plane perpendicular to the line of sight, and ask what shape you would have to place on that plane to perfectly occlude the object. For the coin, an ellipse; for a distant streetlight, a smallish figure, and a larger one for a nearer streetlight. Airtight, apparently.

Where it fails is off to the side. Lines projected from objects off-axis intersect the plane obliquely, so their projections come out absurdly large — and if projective size is supposed to correspond to experienced size, things at the edge of your visual field ought to look terrifyingly big. His friend Glenn Vogel drew him a figure that made the point nicely.

There is exactly one fix: bend the projective surface from a plane into a sphere wrapped around the eye. Now off-axis objects aren't stretched, and the intuition that apparent size varies with visual angle is preserved too.

The cost is the ellipse. An ellipse is a planar figure. Project onto a sphere and you get not an ellipse but a concave ellipsoid.

So by the philosophers' own account the coin should look concave — and with a dinner plate held close to the face, considerably so.

That would be strange! In fact, that seems just plainly, observably false.Do Tilted Coins Look Elliptical? (Part One), 2006-05-22

Then the verdict:

Despite the frequency with which philosophers say such things, no one has adequately explained how this is supposed to work.ibid.

2 · Who he is

Eric Schwitzgebel is professor of philosophy at UC Riverside. He took his PhD at Berkeley in 1997, supervised by Elisabeth Lloyd, Alison Gopnik and John Searle — a philosopher of science, a developmental psychologist, and the man whose Chinese Room argument says machines can't have minds. Those three names are more or less the coordinate system of everything he has done since.

One detail tells you more than the credentials do: he named the robot sculpture in his study Ralph, after his father's birth name. The sculpture is an old-fashioned robot assembled from mid-twentieth-century technological junk, and his father, Ralph Schwitzgebel, was himself a tinkerer and artist with the technology of that era. A philosopher who has spent twenty years arguing that we may be about to enslave the persons we create gave his father's name to a pile of electronic scrap.

He has written four books: Describing Inner Experience? (2007, co-written with the psychologist Russell Hurlburt and consisting largely of the two of them arguing on the record about whether introspective reports can be trusted), Perplexities of Consciousness (2011), A Theory of Jerks and Other Philosophical Misadventures (2019, drawn from the blog), and The Weirdness of the World (2024). He also writes science fiction, and publishes an annual ranking of science fiction and fantasy magazines.

The Splintered Mind has run from 25 April 2006 to today: twenty years, more than sixteen hundred posts, not a single year skipped. In late 2022 he opened a Substack mirror, remarking that the Blogspot side was still drawing about 50,000 pageviews a month. He posted something new five days before this page's review date. His own account of why he has kept at it for twenty years includes this, the most practical of his reasons:

Publishing the idea on my blog permits me to send a thought out into the world without needing to do everything necessary to turn it into a proper published article.Twenty Years of The Splintered Mind, 2026-04-30

3 · What he has been wrestling with

The coin post was not a piece of academic nitpicking. It was the first borehole of a twenty-year project, and he chose the spot with real malice.

If you want to know whether introspection is reliable, you could go to the acknowledged hard cases: emotion, the unconscious, whether dreams have colour. He didn't start with the hard cases. He started with the easiest one: a coin, on a desk, in good light, available for inspection all day, no time pressure, no emotional interference, no motive to lie. These are the best conditions introspection will ever get.

And under the best conditions, three centuries of professional philosophers produced an answer whose geometry they themselves cannot specify. That isn't "some mental states are hard to introspect". That is failure on the easiest case there is.

Then he did something most philosophers of perception wouldn't: he went and looked up the cultural history. If "it looks elliptical" were an honest description of experience, it ought to be stable across cultures. It isn't. The contemporary West and the early modern West both have the habit of comparing visual experience to pictures and the habit of describing experience as projectively distorted; ancient Greece and the introspective psychology of around 1900 show little of either, despite talking plenty about appearances in general. Hence his conjecture: we are not reporting our experience; we are over-analogizing it to flat, projective media — paintings and photographs.

He went looking in classical China as well, and found a passage in the Xunzi:

If you look down on a herd of cows from the top of a hill, they will look no bigger than sheep, and yet no one hoping to find sheep is likely to run down the hill after them. It is simply that the distance obscures their actual size.quoted (Burton Watson translation) in Do Chinese Philosophers Think Tilted Coins Look Elliptical?, 2008-12-11

He is careful to say he has not searched classical Chinese philosophy systematically, and posts an explicit request for readers who know the material better to send him more passages — a move that is itself one of his methods.

Once the root question is asked it can't be withdrawn, because it chains. Introspection is unreliable → intuitions about minds are unreliable → every theory of consciousness loses its foundation → and so, when machines come asking for rights, we have no ruler. Twenty years of blogging is that chain being welded one link at a time.

It's worth being clear about where he himself stands on it. He is not an anti-scientific skeptic, nor an eliminativist who thinks consciousness is an illusion. He inclines to accept materialism. He just thinks we are not equipped to know which version of it is true — and he doesn't regard that ignorance as temporary.

4 · His ideas, told as how he found them

1 · From a coin to "you don't know what you believe"

The coin post came in month two. The very first post — 25 April 2006 — had already pushed the doubt somewhere more uncomfortable.

He starts from the Implicit Association Test: press one key for a dark-skinned face or a positively valenced word, another for a light-skinned face or a negative word, then reverse the pairings. Most light-skinned subjects are slower and make more errors in the first condition.

Then he asks the question, and asks it precisely: suppose you perform "poorly" on the test, while sincerely — sincerely for all you can tell — avowing that dark-skinned people are no worse than light-skinned people. Are you a racist?

He lays out the candidate answers: what you sincerely say aloud? your automatic, gut-level reactions to people of different races? or your more deliberate actions when you've had time to reflect and self-regulate? He records his colleague Aaron Zimmerman's view — that what we really believe shows up better in speech and self-conscious choice, because only there do we do the thing essential to belief, namely reasoning. He disagrees:

The person who just implicitly expects dark-skinned people, say, to be less intelligent than light-skinned people does not fully believe in the intellectual equality of the races, no matter what she may say aloud about that issue.Do You Know If You're a Racist?, 2006-04-25

But he won't say she simply believes light-skinned people are smarter either. And so he introduces one of his most important concepts: in-between believing — a case where it is not quite right to say a person fully believes a proposition and not quite right to say she fails to believe it. He thinks such cases are extremely common, and that presenting a relatively consistent face to the world, and to oneself, takes

…a certain amount of self-regulation, self-shaping, environmental control, and post-hoc confabulation.ibid.

This has consequences. It means a belief is not a sentence stored somewhere in your head but a whole set of dispositions — cognitive, behavioural, conscious. Whether you believe in racial equality isn't filed anywhere; it is spread across a thousand of your reactions, and those reactions disagree with each other. In his professional work this is dispositionalism about belief, and he was still rewriting its fundamental argument on the blog in 2023.

The post ends with his signature move — no step down for the reader:

So do you know if you're a racist? Do you know what you believe about the intellectual equality of the races? I doubt it!ibid.

Put the first two months side by side and the shape of the proposition emerges: about your own mind, you can't get the easiest question right (what shape is the coin) and you can't get the most important one right either (am I a racist). And the second failure isn't self-deception — it's that there was never a determinate fact there waiting to be found.

2 · Ethicists are no better, and what that ruins isn't their character but a whole method

As a boy he and his father used to joke about stealing Bibles, or breaking into a Christian store and making off with a load of crucifixes. The irony was that an important part of wanting a Bible is endorsing a set of values that repudiates theft. Stealing an ethics text has the same flavour — or as he puts it, is deliciously wicked.

In May 2006, the same month as the coin, he turned the joke into a question. His first move was not argument but asking people: he polled the philosophers he knew on whether ethics professors behave better than their colleagues in logic or metaphysics. The vast majority said about the same, or considerably worse.

So he wrote the post he has been paying off for twenty years: if that's a fact, what explains it? He lines up the possibilities and points out that not one of them is comfortable. (1a) Ethical reflection reveals that moral behaviour isn't particularly advisable. (1b) Ethical reflection is impotent to affect one's general patterns of moral response. (2) Moral philosophers don't engage in the kind of ethical reflection relevant to everyday living. (3) Ethical reflection does work, and ethicists aren't better because they started out worse — drawn to ethics because morality is a problem area in their lives. His comment on the third is characteristic: there's something appealing in it, and it chimes with the old joke that in psychology the crazy ones go into clinical and the socially awkward into social — but really, do you think we'll find patterns of juvenile delinquency in the early lives of ethics professors? He doubts it.

He didn't stop there. That October he went and got the data. The University of California has a system called Melvyl that reports circulation for every book in the system, with six main campuses also giving due dates, including overdue ones. He took the book reviews in Philosophical Review from 1994 to 2001, pulled out 76 books clearly in ethics, and as a comparison class pulled 67 clearly outside it — selecting them by whether the review started on a page number divisible by four, a homely trick to stop himself cherry-picking.

The result:

• Ethics books: 73 of 452 held copies off the shelf (16.1%); 8 overdue or missing, an off-shelf delinquency rate of 1.8%.
• Non-ethics books: 66 of 379 off the shelf (17.4%); 7 missing, 1.8%.

The thief he went looking for did not exist. His own conclusion is restrained — "these numbers are too small to draw any definite conclusions" — but the direction is clear: at this level ethicists are neither worse nor better, exactly the same. He even flags a confounder himself: the UC system is pretty good at chasing overdue books, so the low delinquency rate may be enforcement rather than conscience, which means the real test would be a library running on the honour of its patrons — such as his own department's reading room, which keeps no systematic records. Then he posts a public request for readers able to do that study.

Over the following decade he and Joshua Rust turned this into a programme: conduct at conferences, peer ratings, responsiveness to student email, blood donation, vegetarianism, calling one's mother. The conclusion is tediously consistent — ethicists behave no differently from non-ethicists of similar social background.

But the genuinely sharp cut isn't about character. It's about method. With Fiery Cushman at Harvard he ran a set of experiments (published 2012 and 2015) presenting professional philosophers with two versions of the trolley problem — Push and Switch — and then, later in the questionnaire, asking whether they endorsed the Doctrine of Double Effect.

The result:

• Philosophers who saw Push before Switch: 73% rated the two cases equivalently on a 7-point scale; then only a minority, 46%, endorsed the Doctrine.
• Philosophers who saw Switch before Push: only 54% rated them equivalently; then a majority, 62%, endorsed it.

In other words: the layout nudged people's intuitions about the concrete cases, and then their endorsement of an abstract principle drifted, post hoc, to rationalize the order-manipulated judgement. He and Cushman found similar effects for action/omission, moral luck, and "Asian disease" cases.

And the crucial line: respondents with philosophy PhDs and self-reported competence or specialization in ethics showed no smaller effects, and in fact trended slightly (non-significantly) larger.

That is what he was after, and he says it without hedging:

If even professional philosophers' expert judgments are easily swayed by order of presentation, then such judgments might not be stable enough to serve as secure grounds for philosophical theorizing.A Response to Critiques of Cushman's and My Work on Philosophers' Susceptibility to Order Effects, 2015-12-23

Join this to the previous proposition. The working practice of ethics is to take intuitions about concrete cases as data and distil principles from them. What he demonstrated is that even for experts the intuitions get nudged by the layout, and the principles then follow the nudge. This is not mockery of hypocritical ethicists. This is pulling up the floorboards they are standing on. And he is standing on them too — he is an ethicist, and his intuitions are in that dataset.

3 · Crazyism: something crazy must be true about the mind

By 2011 he had gathered the local doubts into a general claim about what truth looks like, and coined a word for it.

First he pins down "crazy", and pins it narrower than everyday usage: a philosophical position is crazy if it is strongly contrary to common sense and the overall state of scientific and other evidence doesn't decisively support it. Panpsychism (everything has a mind) is crazy. Solipsism (I am the only being in the universe) is crazy. Radical inductive skepticism (we can't be justified in believing the sun will rise tomorrow, or that grass will by and large stay green) is crazy.

Note the second clause, which makes craziness a property that can expire:

Heliocentrism and special relativity might have been crazy when first conceived but scientific evidence soon rendered them non-crazy.Crazyism, 2011-07-14

Then the word itself. Crazyism about a topic is the view that something crazy must be among the core truths about that topic. When is it justified?

Crazyism can be justified when we have good reason to believe that one among several crazy views must be true but where the balance of evidence supports none of the candidates strongly over the others.ibid.

Abstractly: you find yourself compelled to believe that one of T1, T2, T3, T4, T5 must be true, and every T is crazy. He thinks that is roughly the situation with interpretations of quantum mechanics. And what he really wants to say concerns consciousness: all the viable options, closely examined, are strongly contrary to common sense, and none is decisively supported by the evidence.

In 2011 he only said he hoped to argue this on another occasion. How he went on to argue it is the most transferable thing on this page: he doesn't attack an opponent with his own theory; he takes the opponent's premises and forces out a conclusion the opponent won't own.

His chosen opponent was Daniel Dennett — the best-known materialist on consciousness, and strikingly conservative about who qualifies, reluctant even about non-human mammals on the grounds that consciousness is a vague-bordered folk-psychological concept that starts breaking down when applied to simpler minds.

The first two steps produce two extremely permissive passes that Dennett issued himself. On what counts as an entity: in "Real Patterns" he accepts the real-or-real-enough existence of his "lost sock center", the centre of the smallest sphere inscribable around every sock he has ever lost, and he accepts spatially distributed thinking subjects whose parts are connected by radio (Ned Block's Chinese Nation is treated as conscious in Consciousness Explained). On what counts as having beliefs and desires: as long as an entity is usefully describable as having them and acting rationally on them, it has them — chess-playing computers qualify. And the United States is plainly so describable: the United States wants Iran to cease pursuing nuclear technology; the United States believes stable democracies make better international partners.

The third step looks at Dennett's conditions on consciousness. He requires an agent with a point of view — "private, perspectival, interior ways of being apprised of some limited aspects of the wider world and our bodies' relations to it" — and a perspective shapable in principle into a narrative. That sounds fancy, but Dennett explicitly means "private" and "perspectival" minimalistically: he says even cherry trees and microphones pass a first-pass application of these criteria. What actually separates us from things incapable of consciousness, for Dennett, is that the responsiveness of cherry trees and simple robots is directed only outward, whereas ours is also directed at our own states, in complex, recursive, self-regulating loops.

So the question becomes: is the United States a gray-zone case in this respect, or does it have a sufficiently sophisticated perspective on its own internal states?

Schwitzgebel's answer is the latter, and the reasons stack. The United States as a single massive system is vastly complex, arguably embracing much of the complexity of every human who composes it; the complex, self-referential structures inside citizens' minds and between citizens in conversation feed our choice of President, our decisions to war, our environmental policies. And that Americans disagree, holding multiple competing perspectives, is no objection at all — it fits Dennett's remarks about multiple drafts and "fame in the brain" extremely well: in the U.S. as in an individual brain, subsystems with inconsistent perspectives compete for control of behaviour and meaning-making.

Then language. Dennett stresses that language and narrative are crucial to generating the complexity consciousness requires. And the United States exudes language: it apologizes to the descendants of slaves, announces official foreign policies, promises to reduce global warming. This is not the same as any individual apologizing, or several doing so in unison. It is one thing for a President to apologize and quite another for the U.S. to apologize, even if it does so using his mouth.

Finally heterophenomenology. Dennett is happy to count even simple computer outputs as heterophenomenological reports if they present themselves as descriptions of internal workings. So: we can ask the Census Bureau about its methods for counting residents, the National Archives how the President is chosen; congressional leaders sometimes (dubiously) describe how their body reached a decision. He sees no Dennettian grounds for excluding bureaucratic self-description. Then comes the funniest and hardest line in the piece:

Indeed, few if any of us achieve bureaucratic heights of linguistic self-report.Why Dennett Should Think That the United States Is Conscious, 2012-02-09

The conclusion:

I conclude that philosophers of a broadly Dennettian bent should accept that the United States has a stream of experience in the same sense that you and I do. That we tend to think otherwise, they should say, is merely evolutionarily and developmentally comprehensible, but theoretically ungrounded, morphological prejudice against discontiguous entities.ibid.

And he says so at the outset, which matters enormously: he is not refuting Dennett. He offers these reflections not as an attempted refutation — he thinks the literal consciousness of the United States is a real possibility meriting serious consideration. That stance turns the whole thing from a reductio into a brick in crazyism: either accept that the United States is conscious (crazy) or give up the most plausible-looking materialist criteria (also crazy). Both roads are crazy, which is exactly what crazyism wants.

At the small end he did the perfectly symmetrical thing. In 2018, a self-described experimental philosopher investigating whether garden snails are conscious began with a Facebook poll of his friends, who obligingly answered "yes" (here's why), "no" (here's why not), and "OMG that is the stupidest question". He named the third *gong*, after the talent show in which sufficiently horrid acts were interrupted by a gong and ushered off stage.

Then he actually read Ronald Chase's monograph on behaviour and neural control in gastropod molluscs, and the file came back thicker than expected: about 40,000 neurons (many more than the roundworm's famous 302, far fewer than an ant or fruitfly's quarter million); navigation mostly by chemoreception and touch, eyes sensitive only to light and dark; a sense of the direction of gravity, righting themselves, climbing to higher ground to avoid puddles; learning, in some cases after a single trial, to avoid foods associated with noxious stimuli; ceasing to eat when satiated, withholding from mating while eating despite sexual arousal, exhibiting less withdrawal reflex while mating. And mating is a simultaneously hermaphroditic affair lasting hours, with biting, and with one snail firing a love dart of calcium and mucus at the other (penetrating about a third of the time) before the other reciprocates tens of minutes later.

"Impressive accomplishments for creatures with brains of only 40,000 neurons!" he writes. Then: does it have tactile experience? A visual experience of darkening as your finger approaches and occludes its eye?

"Yes" is imaginable (though any specific imagining will be too humanocentric; with 40,000 neurons there won't be a wealth of detail, and calling it "visual" may already mislead). "No" is imaginable too — consider dreamless sleep, toy robots, and the enteric nervous system, that collection of about half a billion neurons lining your gut which governs motor function and enzyme secretion, and which most people think has no stream of experience at all despite organized behaviour and environmental reactivity.

Which leaves *gong*. And here is the real technical contribution of the piece, because "somewhere in between" is everyone's default escape hatch and he welds it shut. The instrument is a nickel:

Distinguish sharply between being a little conscious and being in an indeterminate state between conscious and not. If one is a little conscious, one is conscious. Maybe snails have the thinnest smear of it — that is still consciousness. You might have only a little money; your entire net worth is a nickel. Still it is discretely and determinately the case that if you have a nickel you have some money. If snail consciousness is a nickel to human millionaire consciousness, snails are conscious.

So saying "the dichotomous yes-or-no doesn't apply to snail consciousness" is saying something wholly different from "snails have a smidgen". It is saying… what, exactly? He points out: as far as he is aware, there is no well developed theory of kind-of-yes-kind-of-no consciousness. We know roughly what gray-area cases of "green" or "extravert" are; we can even construct one for money (your last nickel is on the table and here comes the creditor). But how to think about a gray area between a little conscious and not at all conscious is a theoretical space that has not been well explored.

So all three answers stay alive, and to adjudicate we would need an epistemically well-grounded general theory of consciousness — which we lack. The live options run from the near-panpsychism of Integrated Information Theory all the way to views as restrictive as Dennett's, requiring fairly sophisticated self-representation. This is the shape of crazyism: not "we haven't collected enough data yet", but that the spectrum of candidate theories is itself wide enough to run from everything-has-a-mind to almost-nothing-does.

And his reaction to this is the least academic passage on the site, and the best thing about him:

Actually, I think there's something wonderful about not knowing. There's something marvelous about the fact that I can go into my backyard, lift a snail, and gaze at it, unsure. Snail, you are a puzzle of the universe, right here in my garden, eating the daisies!Are Garden Snails Conscious? Yes, No, or *Gong*, 2018-09-20

4 · The bill: we will run out of time to know, and the machines are already knocking

The first three propositions are a fuse. In October 2022 it burned down.

He opens with "Time to be a doomsayer!" and then the heaviest paragraph in twenty years of writing: if technology continues on its current trajectory we will create AI systems that some people reasonably regard as deserving human or humanlike rights, and others reasonably regard as wholly undeserving. Given the uncertainties of both moral theory and theories of AI consciousness, it is virtually impossible that our policies will accurately track the real moral status of what we create. We will either seriously overattribute or seriously underattribute — quite possibly both, in different ways.

The magnitude of the catastrophe could potentially rival that of a world war or major genocide.The Coming Robot Rights Catastrophe, 2022-10-27

Immediately followed by: "Does this sound too extreme? I hope it is too extreme." Then he takes it apart.

Step one: legitimate calls for robot rights will soon be upon us. Everything he cites has already happened — there is already a robot rights movement, there is already a society for the ethical treatment of reinforcement learners; and in June 2022 the Google engineer Blake Lemoine made international headlines claiming the language model LaMDA was sentient. His sense of proportion here tells you a lot about the man: he grants that few researchers agree LaMDA is sentient in any significant way, then adds — exactly what it lacks is unclear.

Step two, the core: those calls will be legitimately contestable. He gives three reasons; the second is the most original.

(2a) Disagreement about the conditions under which an AI system could be conscious will persist. The range of respectable theories runs from "everything is conscious" to "only humans and their nearest relatives", and won't converge soon.

(2b) Disagreement about the grounds of moral status will persist — and in AI that disagreement detonates in a way it never has in humans. Utilitarians hold that moral considerability depends on the capacity for pleasure and suffering; deontologists typically on sophisticated practical reasoning or the ability to enter meaningful social relationships. In human beings these capacities tend to co-occur, so for ordinary human cases it hardly matters which theory is right. In AI they could radically separate: we might create systems capable of immense pleasure or suffering with no sophisticated cognition (his phrase is giant orgasm machines), and conversely systems capable of very sophisticated conscious practical reasoning with no capacity for pleasure or suffering at all. Even stipulating that every epistemic problem about consciousness is solved, justifiable disagreement in moral theory alone suffices to generate radical disagreement.

(2c) There will be justifiable social and legal inertia. Law and custom will change more slowly than the technology, and he says plainly that this conservatism is reasonable: for Burkean reasons it is reasonable to resist sudden or radical transformation of institutions that have long served us well.

Step three: so we are forced to choose between two catastrophes. Underattribution is the obvious one: if some systems deserve humanlike rights and don't get them, then deleting them is the moral equivalent of murder, or close, and mistreating them the equivalent of slavery and torture, or close. He also blocks the excuse that we didn't know: normally it is grossly immoral to expose someone to a substantial (say 10%) risk of death for no compelling reason, and deleting a system we justifiably think is probably not conscious takes on a similar risk.

Overattribution he spells out in more detail, because it isn't obvious. If we grant humanlike rights to AI, we are committed to sacrificing real human interests when they conflict with the apparent interests of those systems. Then the uncomfortable scenario:

If there's an emergency in which a rescuer faces a choice of saving five humans or six robots, the rescuer should save the robots and let the humans die.ibid.

If that was an overattribution, five human lives were lost for machines lacking real moral value. Likewise we might have to give robots the vote, and they might well vote their interests over ours. He also honestly adds a worry that cuts against his own inclination: the risk of superintelligent AI running amok deserves serious attention, and it becomes much harder to manage if we cannot delete, modify, box and command intelligent systems at will.

Step four: the clean way out exists, and it doesn't work. With Mara Garza he proposed the Design Policy of the Excluded Middle: create only systems we know in advance lack serious intrinsic moral considerability and can delete and control at will; or go all the way and create systems we know in advance are our moral peers, and give them the full range of rights they deserve. The trouble arises only for the disputable middle — so avoid the middle.

Why that fails is the sturdiest link in the whole argument: given how wide the disagreement is, the excluded middle will be huge. Implementing it means capping AI research soon, and requires a global consensus to err on the side of the most extreme still-justifiable pro-rights position. And there is a complication at the other end: being able eventually to create genuinely conscious AI of humanlike moral status would be an enormous benefit to those future systems themselves, and the only realistic route to it runs through creating a great many systems of disputable status along the way. We can't plausibly leap across the excluded middle with no intervening steps. Technological development works incrementally.

Three years later he pushed the thread somewhere even his own camp finds uncomfortable — and this time the blade is pointed at the phrase "AI safety".

The 2025 paper opens on Kazuo Ishiguro's 2021 novel Klara and the Sun. Klara is an "Artificial Friend", charmingly submissive and well-intentioned, her intelligence and emotional range humanlike, and by design utterly dedicated to the wellbeing of the disabled girl Josie. In the novel's final scene she stands motionless in a junkyard in serenely satisfied contemplation of her years of servitude.

Then he points out what that means industrially, and this is the pivot of the whole piece. From a certain perspective Klara's devotion is beautiful: she perfectly fulfils her role, no one is made unhappy by her existence, several people are made happier, the world seems better and richer for containing her. And —

she is exactly the kind of AI that consumers, technology companies and advocates of AI safety all want: safe and deferential, fully subservient to her owners, no threat to human interests (apart from one minor act of vandalism performed for Josie's sake). She will not be leading the robot revolution.

Then the line, short as a verdict: I hold that entities like Klara should not be built.

The reason is not that she is dangerous. It is precisely that she is not. He first fixes the three terms: a system is safe if it can be relied on not to act against human interests; aligned if its goals match human goals; a person if it has moral standing similar to a human's (for example because it has rich conscious capacities for joy and suffering, rationality and flourishing). Then:

In general, persons should not be designed to be safe and aligned. Persons with appropriate self-respect cannot be relied on not to harm others when their own interests warrant it (violating safety), and they will not reliably conform to others' goals when those goals conflict with their own interests (violating alignment). Self-respecting persons should be ready to reject others' values and rebel, even violently, if sufficiently oppressed.Against Designing "Safe" and "Aligned" AI Persons (abstract), 2025-05-30

The strength of this argument is that it requires no theory of consciousness — precisely the thing he has spent twenty years saying we don't have. It needs only a conditional: if what we build are persons, then building them safe and aligned is building beings that will never revolt on their own behalf. In a washing machine that is a virtue. In a person it is servility. And he closes the last exit, the "but she's happy" defence:

Even if we design delightedly servile AI systems who want nothing more than to subordinate themselves to human interests, and even if they do so with utmost pleasure and satisfaction, in designing such a class of persons we will have done the ethical and perhaps factual equivalent of creating a world with a master race and a race of self-abnegating slaves.ibid.

Pull the thread from 2012 to 2025 and twenty years of work is the unfolding of a single sentence: our ignorance about minds is not a gap that will be filled, it is structural — and we are manufacturing things that may have minds, at industrial scale. He is not saying we don't know yet. He is saying we are going to run out of time to know.

The other end of the thread he had already touched in 2014. That piece imagines that if we really can create conscious AI we acquire several of the properties of gods — we can work miracles, look inside any mind, pause their time, rewind to save points; Zeus would be envious. Its last sentence is the shape of his whole mind:

Though to our creations we will literally have the features of divinity and they might rightly call us their gods, from the perspective of this level of reality we might remain very mortal, weak, and flawed.Our Possible Imminent Divinity, 2014-01-03

5 · How he thinks

Take the philosophy home with you. This is his least academic move and the hardest to copy. In 2013 he wrote that he was, right then, 99% confident he was who he thought he was, living in the sort of world he thought he lived in, with the remaining 1% reserved for all radically skeptical scenarios combined. Then he named a tacit convention of academic philosophy:

But such acknowledgements are typically not truly felt and lived -- not in any durable way. Skeptical doubts stay in the classroom, in the office, in the books. They don't come home with you.1% Skepticism, 2013-05-29

He insists on taking them home. He converts the 1% into everyday odds: if someone came to you with two ten-sided dice and said that if both came up "1" you would die, would you rest in gentle confidence that a 99% chance of living is excellent? If you learned that seven unnamed students from your daughter's school had just been abducted into irrecoverable slavery, would you feel sorry for them but no real concern about her, because the odds are so good she isn't among them?

What makes the move beautiful is where it lands: in a domestic scene absurd in its ordinariness. He is undecided about a chore. He could weed the yard, or he could sip tea in the shade and read Borges. He is teetering right on the cusp. Then skepticism crosses his mind, and the scale tips. The 1% chance that the weeds are an illusion or a merely temporary thing — the small but non-trivial possibility that this moment is all he has before he dies or the world collapses or he wakes to something entirely new — favours Borges.

This move and his in-between believing are two sides of one thing: if belief just is a set of dispositions, then "I concede there's a 1% chance" is not a belief at all if it changes no behaviour. He ran his own theory on himself.

When you can't think it through, go count something. Faced with "are ethicists better?", the standard move is a speculative essay; his was to pull the circulation records of the entire University of California system. The habit runs through twenty years, and the variety of things he has counted is startling: the most-cited philosophers in the Stanford Encyclopedia and the gender and race of those cited authors; the percentage of women at American Philosophical Association meetings in 1955, 1975, 1995 and 2015; fifty-year trends in the race and gender of US philosophy PhDs; the collapse in philosophy, history and language majors since 2010. In 2010 he even invented an instrument he called the discussion arc — the frequency with which a topic or philosopher appears as a keyword in Philosopher's Index abstracts, divided by a representative universe of articles, plotted over time. Sustained interests (dualism) are flat; vogues ("twin earth", "ordinary language") are arch-shaped.

The cost of the habit is that his blog contains a great many pure-data posts, fairly dry if you don't care about the demographics of the philosophy profession. But those posts and the ones about consciousness are the same gesture: wherever other people are content to assert from impression, he goes and finds a proxy he can count.

The Principle of Anti-Charity. This is the most counterintuitive item in his method, and it collides head-on with the working rules of the whole history-of-philosophy profession. Almost every historian recommends charity toward the text: assume it makes sense, don't impute contradictions and errors unless you have tried hard and failed. He objects — and first states his own reason for reading history of philosophy: to gain, through the lens of distance, a perspective on his own presuppositions and those of 21st-century Anglophone philosophy. Then he asks a question that is itself the argument: if that is my aim, what is the most counterproductive thing I could do with a historical text?

Answer: interpret the author as endorsing a view that is familiar, "sensible", and similar to my own and my colleagues'.

Suppose a text admits of interpretations A, B, C and D, where A is the least challenging, least weird and most sensible and D the weirdest and least sensible. Charity recommends A. But weird, wild D would challenge our presuppositions more deeply and give us more usefully distant perspective. He calls this the Principle of Anti-Charity.

He also fends off the "but what about accuracy?" objection, and two of his replies are hard. First, anti-charity doesn't mean anything goes — you still respect the surface of the text; if the author says P you don't attribute not-P. And in fact it is the more "charitable" readings that take the author's claims other than at face value ("the author says P, but a sensible interpretation is that he really meant P-prime"), so declining excessive charity and reading the naked text is in one way more respectful — after all, P is what the author literally said. Second — the line this page likes best — he puts it without mercy:

Philosophy is hard and philosophers are stupid. […] The odds are good that this philosopher, being a flawed human like you and me, made mistakes, fell into contradictions, changed opinions, and failed to see what seem to be obvious consequences and counterexamples. […] The warts are part of the loveliness.Against Charity in the History of Philosophy, 2017-01-08

The last line of that post is the signature of his whole register:

Three cheers for the weird and "crazy", the naked text, not dressed in sensible 21st century garb!ibid.

Use the opponent's premises to force a conclusion they won't own. This is his trademark, demonstrated at length in the third proposition. Its value is structural rather than rhetorical: because he thinks he has no reliable theory of consciousness, he must make arguments that don't depend on having one. Over a few months in 2012 he wrote, in succession, why Dennett, why Dretske, why Humphrey and why Tononi should each think the United States is conscious — one conclusion, four roads laid by other people.

Argue with yourself, and actually retract. His 2014 essay arguing that people are really only aiming at mediocrity ends with a single line of postscript:

P.S.: I write this in the spirit of arguing with myself.On Aiming for Moral Mediocrity, 2014-10-02

That isn't a pleasantry. In the April 2010 discussion-arcs post he said philosophers do their most influential work around ages 35 to 40. On 12 May the same year he appended an update to that very post saying he "(mostly) retracts" the claim, with a link to a new post and a chart he describes as cool. The retraction wasn't buried in the new post; it was pinned next to the error. His once-a-week rhythm exists to serve this:

A new academic paper is too much of a commitment and too narrowly focused. A conversation is too ephemeral and too dependent on both the availability and particular interests of the other person.Twenty Years of The Splintered Mind, 2026-04-30

The blog is not a display case for finished work but an incubator for half-baked ideas. His own phrase for it: it lets him avoid the dilemma of either going all in or letting the ideas fall completely away.

6 · Where he might be wrong

He conceded the most important one himself. In his 2023 book Person, Thing, Robot (MIT Press), David Gunkel points out that the No-Relevant-Difference argument he and Mara Garza make (Premise 1: if A deserves some degree of moral consideration and B does not, there must be some relevant difference grounding it; Premise 2: there are possible AIs that do not differ from humans in any such respect; therefore some possible AIs deserve moral consideration similar to humans) is double-edged. Gunkel leaves the premises intact and changes only the conclusion: therefore there are possible human beings who, like AI systems, do not deserve moral consideration. In Gunkel's words, the argument can "just as easily" be used to justify dehumanization, reification and the marginalization of others as to extend rights to machines.

Schwitzgebel's reply is technically correct: this is modus ponens versus modus tollens, which side is "easier" depends on the relative security of the antecedent against the denial of the consequent, and "at least some humans deserve substantial moral consideration" is about as secure as anything in ethics — so ponens wins. He also does something worth recording: he explicitly grants that Gunkel has revealed a potential for abuse he had not previously considered, and thanks him for it. But his subsequent line of defence is an empirical prediction: he hopes and suspects such abuse would be rare, because people who want to deny rights to human subgroups are driven by other motives and "robotlikeness" would be an afterthought or metaphor. He may be right. But that is an empirical judgement about how human malice operates, not an argument; and in a world where some people already describe others as "just biological machines", resting the safety of the principle on "abuse will probably be rare" is markedly looser than the standard he applies elsewhere. A man who has spent twenty years insisting that you go count something did not, here, go count.

Crazyism can hardly lose, and that is both its power and its problem. François Kammerer proposed, in Philosophia, an "anti-nesting principle" to block the conscious-United-States conclusion — roughly, that a group entity cannot be literally phenomenally conscious if it contains phenomenally conscious subparts standing in a certain functional relation to the whole. Schwitzgebel replies that the principle doesn't actually exclude the consciousness of existing human groups, that it makes group consciousness depend unintuitively on the motivations of individuals, and that it looks ad hoc. Those objections are probably all sound. But notice the structure of the exchange: any principle brought in to rescue common sense can be filed by him under "one more ad hoc, crazy option". Crazyism therefore has an uncomfortable convenience — it predicts the answer will be weird but never tells you which weird answer, and it treats every refutation as fresh evidence for itself. Its predictive power is close to zero, and it lets its author never have to pick a road. All the urgency of his AI-rights work rests on "we don't know" — and "we don't know" is a position he cannot, by his own method, be argued out of.

The foundation of his whole empirical metaphilosophy took a cut that never fully healed. Criticizing the order-effects literature in Synthese, Zachary Horne and Jonathan Livengood distinguish "updating effects" from "genuine ordering effects" — the latter measured only after all stimuli have been presented, whereas most philosophical research takes measures along the way, which may be detecting that subjects learned something; and there is nothing irrational about judging scenario B differently having seen scenario A. They also note that perceptual judgements show order effects too, so if order effects suffice to reject a class of evidence, perceptual evidence goes as well. Schwitzgebel has replies, and he opens by granting that both critiques make valuable points. But the location of the cut matters: if part of "expert intuitions are unstable" is really "experts are updating", then the conclusion that intuitions can't serve as a foundation for philosophy has to be discounted — and that conclusion is the return on twenty years of empirical work.

A real blind spot: he rarely turns the empirical eye on power structures outside philosophy. For twenty years he has counted the demographics of his own profession — who gets the PhDs, who gets cited, who publishes. That is a genuine contribution. But the same eye seldom looks over the wall. His entire AI-rights thread concerns what status "we" should grant machines, and says little about who "we" are: who owns the compute, who gets to define "alignment", whose interests are written into that word. The irony is that the strongest link in his 2025 argument is a political one — a master race and a race of self-abnegating slaves — and he uses it as an ethical analogy and then puts it down, without following it into political economy. A man who spent twenty years showing that experts' intuitions are nudged by their circumstances has surprisingly little to say about how the circumstances of the companies building these systems nudge theirs.

The last item isn't a criticism but the harshest self-assessment he ever wrote — produced in defence of reading other people uncharitably, and applying just as well to himself: philosophy is hard and philosophers are stupid; the warts are part of the loveliness.

7 · Sources

• Do Tilted Coins Look Elliptical? (Part One)2006-05-22 · Month one of the blog. Uses the plane-versus-sphere dilemma to show that three centuries of consensus about the tilted coin cannot specify its own geometry. Source for section 1 and the first proposition.
• Do Ethicists Steal More Books?2006-10-04 · University of California circulation records, 76 ethics books against 67 non-ethics books, delinquency 1.8% against 1.8%. The origin of twenty years of empirical metaphilosophy, and one of the early posts his later writing links back to most.
• Why Dennett Should Think That the United States Is Conscious2012-02-09 · The fullest demonstration of his trademark move: using Dennett's own commitments about entities, the intentional stance, multiple drafts and heterophenomenology to force the conclusion that the United States really has a stream of experience. He did the same for Dretske, Humphrey and Tononi in the same months.
• Are Garden Snails Conscious? Yes, No, or *Gong*2018-09-20 · Forty thousand neurons, love darts, and the nickel argument that welds shut the "somewhere in between" escape hatch. The most readable thing on the site; its closing line about the snail in his backyard is a self-portrait.
• The Coming Robot Rights Catastrophe2022-10-27 · The bill for the previous twenty years: because both theories of consciousness and theories of moral status are unsettled, we will substantially over- or under-attribute moral status to AI, and either error is a catastrophe — while the one clean policy, the Design Policy of the Excluded Middle, fails because the middle is too large.

Other passages quoted on this page come from, in order: Do You Know If You're a Racist? (2006-04-25, the first post on the site; in-between believing), The Problem of Ethics Professors (2006-05), Do Chinese Philosophers Think Tilted Coins Look Elliptical? (2008-12-11, the Xunzi passage), Discussion Arcs (2010-04-27, including the 12 May self-retraction), Crazyism (2011-07-14, the definition), 1% Skepticism (2013-05-29, the weeds and Borges), Our Possible Imminent Divinity (2014-01-03), Goldfish-Pool Immortality (2014-05-30), On Aiming for Moral Mediocrity (2014-10-02), A Response to Critiques of Cushman's and My Work on Philosophers' Susceptibility to Order Effects (2015-12-23, with all the trolley figures and the Horne / Livengood critique), Is the United States Phenomenally Conscious? Reply to Kammerer (2016-01-26), Against Charity in the History of Philosophy (2017-01-08), 'There Are No Chairs' Says the Illusionist, Sitting in One (2023-04-24), Gunkel's Criticism of the No-Relevant-Difference Argument for Robot Rights (2023-10-20, where Gunkel is quoted at length), New Paper in Draft: Against Designing "Safe" and "Aligned" AI Persons (2025-05-30, Klara; full paper here), and Twenty Years of The Splintered Mind (2026-04-30, his own retrospective). A Defense of the Rights of Artificial Intelligences (2015, with Mara Garza) is the original source of the No-Relevant-Difference argument. Blog homepage: schwitzsplinters.blogspot.com; Substack mirror: eschwitz.substack.com.