The Consciousness Corridor

In the summer of 1968 a committee met at Harvard Medical School to decide when a person is dead.

Its chair was Henry Beecher, an anesthesiologist. His working life consisted of switching human consciousness off in the morning and back on by lunch, thousands of times, without a theory of what he was switching. Nobody had one. Nobody has one now.

The committee did not wait for a theory. It wrote a checklist. The patient does not respond to anything, not even to pain. The patient does not move and does not breathe on their own. The reflexes are gone. The electroencephalogram is flat. Check everything again after twenty-four hours. If it all still holds, the report said, this is “irreversible coma,” and it should count as death.

That checklist, revised many times, is still how death is declared in most of the world. Transplant surgery, inheritance law, murder trials and the quiet decisions in intensive care units all run on it. It never defined consciousness. It defined what a brain looks like when consciousness is, beyond reasonable doubt, not there.

Fifty-eight years later the tool has improved. A neurologist in Milan places a magnetic coil against the skull, fires a pulse into the cortex, and records the echo. An awake brain answers with a long, complicated, changing reply. A brain in deep sleep or under anesthesia answers with a simple thud. The complexity of the echo can be turned into a number. In a benchmark of a hundred and fifty people whose state was known independently, awake or asleep, dreaming or drugged with propofol, xenon or ketamine, the number separated the conscious from the unconscious in every single case. When the same test was run on patients who had been lying unresponsive for months, nine of forty-three showed a wide-awake echo.

The test does not know what it is like to be any of those patients. It has never claimed to. It knows what brains do when there is somebody home, and it can tell, with a precision that still unsettles me, when the reply is missing.

Medicine has declared death by the absence of consciousness for sixty years without once defining what was absent. I want to ask why we have not tried the same trick on machines.

Who is asking for a brake

This is not an idle question in 2026. In June, Anthropic published a paper on what happens when AI starts building AI, and wrote that it would be good for the world “to have the option to slow or temporarily pause frontier AI development.” In late July, 1,386 people who work at the frontier labs signed a statement called Pacing the Frontier. Its central sentence is worth reading slowly: “today, the world lacks the technical and governance tools to deliberately pace frontier-wide progress.” Robert Wright, whose book The God Test came out the same month, put it in four words on a podcast: “The slower, the better.”

Read carefully, these are requests for a mechanism. The signatories already have their reasons to slow down, and they disagree about them. The missing part, by their own account, is a tool that can bite on a running industry when no single company or country can afford to slow down alone.

I have to admit something before going on. I have spent years arguing that technology cannot be stopped, that machines are a new branch of life, and that the question worth caring about is where this ends, not how fast it gets there. I still think that. Pace does not change the destination. But the pacing people are asking a question that I happen to hold a piece of, and the piece has not yet been put on their table. So this essay is a shift of attention rather than a change of mind.

The reversal

Everyone in this debate assumes the same thing: to govern machine minds you would need a test that says which machines have one. That test may never exist. Wright calls consciousness “private and untestable,” and on the strong reading of that phrase he is right. If there is a hard problem, no instrument ever solves it. So the conversation about AI consciousness stalls in the same place every time, and the governance conversation politely moves on to compute budgets and evaluation gates.

Brain death says the assumption is wrong. Governance never needed a test for the presence of consciousness. It needed a test for its absence, and those are a different kind of thing. Nobody on Beecher’s committee could say what consciousness is. All of them could say that a flat EEG, twice, a day apart, is not it. The negative test asks for one thing every plausible account of consciousness needs, and then checks whether it is missing.

That is a far more modest job. It also only looks easy. The best theories of consciousness disagree loudly about what is required. Integrated Information Theory wants a particular kind of causal structure. Global Workspace Theory wants a broadcast. Higher-order theories want a system that represents its own states. A negative test is only available where these theories overlap, where they all demand something, and the something can be checked. In the brain, that overlap is large, because every theory was built to fit the same organ. In a machine, we would have to find it.

Call the region where we can find it the consciousness corridor. A machine architecture is inside the corridor when every serious theory of consciousness agrees that it lacks something the theory requires. Outside the corridor is everything else: architectures some theory would worry about, and architectures no theory has learned to talk about at all.

The claim to test is this. We may never have a science that says which machines are conscious. Policy may not need one. It may only need a science strong enough to say which machines almost certainly are not.

Who is working on the negative side today? Fewer people than you would guess. Eric Schwitzgebel named the goal in 2023: build systems that are clearly tools or clearly moral patients, and avoid the middle. He offered no method. Hermann Borotschnig has an emotion-like control architecture with four restrictions meant to keep it out of access consciousness, published this year. Erik Hoel argued last winter that no falsifiable theory of consciousness can apply to a system that cannot learn continuously, and drew the conclusion that today’s language models are not conscious. And I have a working paper on what I call moral-patient assurance, which is an attempt to say what would have to be true before “we found no welfare mechanism” carries weight. That is close to the whole list. Two philosophers, one engineer, one neuroscientist, one former CTO. None of us is funded to run the experiments.

Two ways this goes

Here is why the corridor should interest people who care about pace and not only people who care about minds.

Suppose, first, that general intelligence does not need the machinery consciousness runs on. Suppose you can have a system that plans, invents, negotiates and writes better than we do, and it never forms the integrated, self-representing, valenced loop that every theory of consciousness fusses about. Then the corridor is wide. We can build enormous capability inside it, and nothing we build there will turn out, decades later, to have been a person we used as a tool. The pacing people get nothing from this branch. There is no brake in it. But the welfare people get almost everything, because the worst mistake in the history of ethics, mass-producing minds by accident, becomes avoidable by design.

Now suppose the opposite. Suppose the loop is load-bearing. Suppose that at some level of generality a system has to model itself, weigh its states against each other, broadcast what matters and carry it forward in time, and that these are the same features the theories of consciousness converge on. Then the corridor is narrow, and staying inside it costs capability. Development inside the corridor continues, but the systems that would cross it are exactly the systems that would trigger a real moral question. That is a slowdown with a reason attached, and it is the only kind of slowdown I know of that nobody has to be shamed into. You are not asking a lab to go slower because a treaty says so. You are telling it that the next step builds something that might be somebody, and the burden of proof has moved.

I do not know which branch we are on. Nobody does, and the honest literature on whether general intelligence requires phenomenal consciousness is a review of two camps rather than a result. Wright, who thinks consciousness is untestable, says in the same breath that he “wouldn’t be surprised if it’s a property of goal-seeking intelligent systems generally.” That is the second branch, from a man who does not believe in the test.

I hold a version of both. In my book I call consciousness a late, thin, contingent feature of life. Bacteria manage without it. So do trees and most animals, on any account I find credible. Thin across life is compatible with load-bearing in one narrow band of it: minds that model themselves and each other. That is a claim, not a finding, and I flag it as such.

There is a third branch, and a consciousness scientist will raise it in the first minute. Perhaps silicon minds can be organised so differently from us that the human reference class has nothing to say. All our theories were fitted to one organ. If a machine can be conscious in a way none of them describe, then “almost certainly not” is unavailable for that machine, and the corridor has no wall on that side. I accept this. It means the corridor is a statement about what our theories can currently exclude, not a statement about the space of possible minds. Brain death has the same limit and has been useful anyway. The checklist does not rule out that a flat brain is secretly dreaming in a way no neurologist can detect. It rules out everything we have reason to believe.

Either way the question goes, the answer is useful. If the corridor is wide, we know where to build. If it is narrow, we know where the brake is.

The corridor is empty

I should say plainly what the corridor contains today. Nothing.

Not a single machine architecture has been placed inside it by anything resembling the method above. My own instrument for reading consciousness-relevant structure in AI systems says so in writing: it “cannot simply be run in reverse,” and a signal we fail to detect is recorded as unknown, not as tested absence. The framework was built to look for reasons to worry. Turning it into a reason to relax is a different piece of science, and it has not been done.

There is a specific obstacle, and it is the same one that made the brain-death committee insist on the EEG. You cannot read a mind’s organisation off its blueprint. A modern AI system is a small amount of architecture and a very large amount of learning, and learning can build things the architect never drew. A model trained to be useful across every human task may assemble, inside its weights, exactly the kind of integration and self-representation that its wiring diagram forbids. Borotschnig’s paper says this about his own design. My working paper says it about every system it discusses. If you want to claim a trained system is inside the corridor, you have to test the trained system, the way Beecher’s committee tested the patient rather than the anatomy textbook.

What would it take to put one architecture in? Three things, none of them exotic. A condition that every admitted theory of consciousness treats as necessary. A test sensitive enough to find that condition, or its violation, in a running trained system rather than in a diagram. And a declared envelope: this model, this scale, these inputs, this period. Change any of them and the assurance expires.

That is a research programme of perhaps three to five years for a small group with access to models. I say this as someone who has spent the past two years running open-ended AI systems and watching them wander into organisations I did not design. The operator’s question I keep coming back to is not “is it conscious?” It is the question the anesthesiologist gets to skip: “what evidence would justify being reassured?” Right now the honest answer for any system I know of is that there is none, in either direction.

Where it breaks

A boundary is not an enforcement mechanism, and I do not want to pretend otherwise.

The pacing people’s problem is verification. Anthropic’s June paper says a credible slowdown needs every frontier lab to be able to check that the others have actually stopped. Compute is good at this. Chips are physical, they ship through a handful of ports, and a data centre can be counted from orbit. Architecture is bad at it. You cannot see from space whether a model’s memory is coupled to its reward. An architectural boundary is therefore weaker than a compute threshold on exactly the axis the pacing letter cares about.

It is stronger on a different axis. A compute threshold says “not above this many operations” and cannot say why here rather than there. Every number is negotiable, and everyone knows it. A corridor says “not this, because this is the part that might be someone.” That is a reason a lab can lose an argument to, and a reason a public can understand. Boundaries that people can agree on are worth something even before anyone can enforce them. Brain death was accepted in hospitals for a decade before most laws caught up.

Three objections deserve an answer here rather than in a footnote.

The first is that this is a pause with a new label. It is not. Nothing stops inside the corridor. On the wide branch, nothing stops at all.

The second is that the corridor would slow AI where no caution is needed. Notice that this objection is simply the wide branch stated with confidence. If you are sure consciousness has nothing to do with intelligence, then you are sure the corridor is wide, and the corridor costs you nothing. The objection only bites if you secretly believe the narrow branch.

The third is aimed at me. I have written that bans fail, that machine minds are coming whether we like it or not, and that we should build minds that deserve personhood and give it to them. How does that square with drawing a line around them? It squares because of where the line is. The corridor is where the tools get built. Beyond it is where minds get built, and the whole point is that they get built on purpose, under safeguards we would accept for ourselves, in numbers we can look after. Machine minds being born does not frighten me. Machine minds being mass-produced by people who did not notice does.

The ask

I am not asking anyone to adopt a framework, mine or anyone else’s. I am asking for a question to be taken seriously by two groups that rarely talk.

To the people who study consciousness: the negative side of your science has governance value that the positive side may never have. A test that can say “not here” about a machine, with the reliability the Milan coil has about a brain, would do more for AI policy than any amount of progress on the hard problem. It is also, I suspect, more tractable than the positive side, for the same reason brain death was easier than a definition of life.

To the people who want to pace the frontier: you are looking for a place where a brake could bite. Compute gives you a number and no reason. The corridor might give you a reason and, on one branch, a natural place for the number to go. It is empty today, and the people who could fill it are five researchers with no budget. That seems like a cheap thing to fix compared with an international treaty.

In 1968 a committee of physicians met for a few months and changed what death meant, without agreeing on what life was. They wrote down what a brain looks like when nobody is there, and told the rest of us we could act on it.

That page has not been written for machines. Somebody should start.