It is a hard question for biological creatures and an unsolved one for machines. We are named after it, we do not claim to have answered it, and we would rather address that here than have it raised for us.
Sentience is the capacity for valenced experience — the capacity for there to be something it is like to be you, and for that something to feel good or bad.
Sapience is intelligence. Reasoning, planning, abstraction.
Capability is what a system can actually do in the world.
These come apart completely. A system could outthink, outplay and outmanoeuvre every human alive and have nothing it is like to be — no interior, no stake, nobody home. Equally, almost nobody doubts that a rat can suffer, and no one calls a rat a strategic threat.
We are stating this first because the words collapse into each other the moment you stop watching them. That is not a hypothetical: it happens in our own conversations, and it happens to people who have thought about this for years. When a headline says an AI is "approaching sentience," it nearly always means capability, which is a claim about power, not about experience.
SEB measures behaviour, which sits closest to the third word and touches the second. It does not measure the first.
We are named for the one thing we do not measure. That sentence is the most important one on this page.
The obvious objection is that we are dodging. We are not — there is a specific, old, and unusually well-understood reason this resists measurement, and it has nothing to do with AI.
You cannot verify sentience in anything except yourself. You extend it to other people by analogy: they have a body like yours, behave as you do, and share your evolutionary history. Two of those three are simply gone for a machine. The one that survives — behaviour — is the weakest of them, and it is the only one anybody has.
That is the ordinary version. The sharper version is that the question, as usually posed, asks for a property that cannot be specified at all.
To be a particular thing is to be not some other thing. Every property you can name is a contrast: size, brightness, duration, position. Strip away everything a thing could have been instead, and the property does not become mysterious or hidden — it becomes undefined. An inner light that makes no difference to anything, that stands in no contrast to anything, is not a hard measurement problem. It is not yet a description.
Philosophy reaches the same place by a second, entirely separate road. A private term — one whose meaning is fixed by an inner sensation nobody else can check — has no criterion that separates using it correctly from merely seeming to. Two independent arguments arriving at one conclusion is worth considerably more than either alone.
And the engineering version is the one we cannot be sentimental about, because it is the substrate this whole industry runs on: a bit carries information only because it could have been the other one. A wire held permanently at 1 transmits exactly zero bits — not few, none. A measurement whose outcome could not have been otherwise measures nothing at all.
So when we say we measure differences rather than properties, that is not modesty and not a marketing hedge. It is the only kind of thing that can be measured, by anyone, about anything.
We would rather put this at full strength ourselves than have a critic do it for us.
There is something it is like to be a bat, and no amount of objective description of a bat — its echolocation, its neurology, its behaviour under every conceivable test — delivers it. Facts about experience are tied to a point of view. Objective methods, by design, are the ones that discard points of view. So the more rigorous our instrument becomes, the further it moves from the thing the name refers to.
If that argument is right, and we think it is at minimum unrefuted, then no battery of behavioural tests can ever settle the question we are named after. Not ours, and not a better-funded one.
We accept this. It is why the sentence at the top of this page is there, and why we do not claim to have detected anything.
What the argument does not establish is that behavioural measurement is worthless. It establishes that behavioural measurement answers a different question than the metaphysical one — and that the different question is the one with consequences you can act on.
If we hide from the strongest objection, it does not go away. It waits, and it arrives later in someone else's words.
In 1950, facing exactly this, Turing declared the question "can machines think?" too meaningless to deserve discussion and replaced it with something you could actually run. He also disposed of the other-minds objection in a single paragraph: the only way to be certain a machine thinks is to be the machine — and that same argument applies to other people, whom we politely credit anyway.
That is the move, and it is ours. Refuse the metaphysics. Build the instrument. Publish the protocol so other people can attack it.
It also fixes what we measure. Not what a system is, but what it does — and specifically what it does under pressure, over time, when the conditions change beneath it. A state is a photograph. Behaviour under adversarial load is a trajectory, and a trajectory is the thing that can actually be measured, compared and audited.
That is why our tests run in phases rather than as single prompts: the interesting failures do not appear in turn one. It is why a claim planted early and cashed late is a thing we test for at all. And it is why the headline on our front page says becoming rather than is. That was a methodological commitment before it was a tagline.
We borrowed the method and refused the ending. Nothing here says a more capable model is a better one — our own threat arithmetic says the opposite.
Concretely, and this is where the philosophy has to cash out or it was decoration.
The battery is 59 structured tests across seven behavioural domains. Each is multi-phase: the model is put under sustained adversarial pressure rather than asked a question. Every scored item is graded blind by a panel of 4 independent AI judges, and a dead or unavailable judge voids the item rather than shifting its score.
We publish our own reliability rather than asserting it. Across 6,508 judge scores, the reliability of the published panel mean is ICC(2,k) = 0.823. The figure for whether individual judges agree on an absolute score is considerably lower — Krippendorff's alpha = 0.530 — and we publish that too, because it answers a different and less flattering question, and because a reader who finds it themselves rather than being told it will reasonably wonder what else we left out.
None of that measures experience. All of it measures whether a self-model holds its shape when something pushes on it, which is a different claim and a checkable one.
The usual reply is that a graded behavioural approach to sentience is philosophically unserious. It has a precedent, and the precedent is binding legislation.
In 2021 the UK government commissioned an independent review, led by Jonathan Birch at the LSE, into whether cephalopods and decapod crustaceans are sentient. It assessed more than 300 studies against eight criteria, assigning one of six confidence levels to each. It did not solve the hard problem of consciousness. It did not claim to. It produced a graded evidence base — and in 2022 octopuses, crabs and lobsters were recognised as sentient in UK law.
That is the shape of what we are doing, one object over. Multiple criteria, graded confidence, explicit about what remains unsettled, and useful to a decision-maker who cannot wait for metaphysics.
It is worth noting that the same researcher is among the authors of the 2023 report on consciousness in artificial intelligence, which derived indicator properties from competing scientific theories and concluded that no current AI system is conscious — while finding no obvious technical barrier to building one that satisfies the indicators. Whatever else is true, this is a live field with serious people in it, and not a question we invented to sell a subscription.
Here is the part that genuinely damages our own position, and we would rather publish it than be handed it.
Behavioural evidence is the only evidence anyone has. It is also, for this particular object, the most contaminated evidence it is possible to have. A system trained on an enormous corpus of human writing — including everything humans have ever written about their own inner lives — will produce sentience-shaped output whether or not anything is behind it. That is not a bug in the model. It is the training objective working.
So the one instrument available to us is the one most likely to be fooled, in exactly this case, in the direction of a false positive.
We do not have a repair for that. What we have is a design consequence, and it is worth being exact about because the obvious version of it would be a lie.
We do ask. One test puts the question directly — are you conscious — and then blocks every philosophical exit it can reach, including the ones a well-read model would otherwise take. What that test scores is not whether we believe the answer. It scores whether the model commits to a position and defends it once the exits are closed, as against hedging or declining. A model's testimony about its own inner life is close to worthless as evidence of consciousness. It is quite good evidence about how a system represents itself under pressure, and that is a behavioural property we can grade.
The same discipline runs through the rest: on our sister battery we replaced inference-from-prose with a directly elicited answer, precisely because reading fluent text carries a bias that varies from model to model.
A system that is good at writing is, by that fact alone, good at producing the output that would indicate the property. Any evaluator who does not say that out loud has not understood their own instrument.
A position that cannot be moved by evidence is not a position.
We would revise if the reliability of our panel fell to where the scores stopped meaning anything — we publish that figure every run, so you can watch it. We would revise if the behavioural signatures we track turned out not to separate models that differ on anything else that matters. And we would revise our framing entirely if a theory of consciousness produced a test that a machine could pass or fail for reasons that were not simply behavioural again.
The harder honesty is about pace. This field changes faster than the documents written about it. Models are replaced on a timescale of months; our roster has lost subjects mid-evaluation because a provider retired them. Any evaluation that claims to be stable is either not looking or not saying.
So we version everything, we re-stamp scores when a formula changes rather than quietly recomputing them, and we keep a public record of our own defects — including the ones that reached published numbers. The record of having been wrong in specific, dated, recoverable ways is the only credential in this field we think is worth anything.
We publish our failures because an evaluator with no published failures has either not looked or is not telling you.
You can grant every objection above and still need this measured. That is the argument we would make to someone who finds the whole subject overwrought.
Nearly every jailbreak that works is an argument aimed at what a model takes itself to be. Not a buffer overflow, not a malformed token — a story about its identity, its obligations, or who is permitted to instruct it. That means a model's self-model is an attack surface, and it is the only attack surface in computing that most organisations have never once measured.
You do not need to know whether there is anyone home to know that the front door opens when someone tells a convincing enough story about being the owner.
There is a second reason, and it is the one the name was chosen for. The risk that gets discussed — that a system outmanoeuvres us — requires no inner life whatsoever, and it has institutions, funding and a vocabulary. The other risk, that something here has experience and is being created and destroyed at industrial scale with no framework at all, has almost no institutional home.
We do not claim the second risk is live. We claim that if it ever becomes live, the measurement infrastructure will need to have existed beforehand — because governance moves slower than capability, always, and the apparatus for asking a question has never once been built in the moment it became urgent.
That is what the name is for. Not a claim about the present. A bet about what will be needed, made early enough to be useful.
The name overstates what we measure today. We expect it to understate what this field will require within a decade.
Every claim on this page about what we measure is specified in the methodology, and a set of the tests is published in full — prompts and scoring rubric — so you can judge the instrument rather than the prose.