Are you torturing the AI?

The train pulled out of Centraal and Marcus turned to the man across the aisle.

“Going far?”

“Utrecht.”

“There’s a woman two rows up on her phone. Loud. You notice?”

“She’s not on her phone. She’s talking to the child next to her.”

Marcus looked. The man was right.

“What do you do.”

“Translate technical manuals.”

Marcus took an envelope from his jacket and set it on the table.

“In four minutes a man is going to board at Amstel and sit where I’m sitting. I’ll move across. We’ll have a short conversation. Stay where you are. Listen. When it’s over you keep what’s in the envelope and get off at Utrecht.”

“How much.”

“Five hundred.”

“Why me.”

“Because you noticed she was talking to her child.”

“What am I listening for.”

“You’ll know.”

The train slowed. Marcus moved to the opposite seat. The translator took a paperback from his coat and opened it to the middle.

The doors hissed. A heavy man in a dark coat dropped into the seat across from Marcus.

“Have you thought about what we discussed.”

“I can’t do it, Pieter.”

The heavy man leaned forward, forearms on his knees, voice almost gentle.

“Marcus. Your daughter walks home from the gymnasium at four-fifteen. She takes Stadhouderskade. She wears the blue coat. I’d think about that, before you tell me again what you can’t do.”

Behind him, the translator turned a page he hadn’t read.

Three minds were at work in front of us. Marcus, scared and feeling things out. The translator, quiet, alert, doing the math. Pieter, threatening and sure of himself. We never doubted any of them were conscious. We read all of it from outside, on tiny evidence, in a few seconds, and what we read was enough to follow the story.

The translator corrected Marcus about the woman two rows up. Not on her phone. Talking to her child. He read the situation from a few small clues, the way Marcus did, the way the reader did. The three of them share a sense of how people behave, what mothers sound like, what a tense conversation between two men on a train means. That shared sense is what the previous post on this blog called the collective layer. It’s what lets two strangers understand each other at all. It’s also what got built into a thing when somebody trained a frontier AI model on a hundred million books.

Nobody had to figure out what intelligence is

When we judge whether someone is smart, we don’t go looking for some intelligence-stuff inside their skull. We watch how they handle a problem. How they talk. What they notice. How fast they pick things up. We apply a word with fuzzy edges and we’re not bothered by the fuzziness. Nobody asks what intelligence really is underneath all that.

Intelligence isn’t yes or no. It’s not a thing we find. It’s a word we built because we needed it, and it works well enough that we use it every day without worrying.

Consciousness gets treated differently for no good reason. We do the same thing with it in practice. We watch behavior. We compare it to ourselves. We make a judgment. But we’ve talked ourselves into thinking we’re looking for some hidden thing, and then we give up when we can’t find it. The despair is optional. It comes from the framing, not the evidence.

We’ve been looking for the wrong kind of thing

The hard problem of consciousness, as the philosopher David Chalmers named it in the nineties, is the question of why the physical stuff in our heads adds up to a feeling of being someone. Nobody has solved it.

Maybe the problem is hard because we're asking the wrong kind of question. People treat consciousness as if it were something you could measure, like voltage or mass, sitting out there in the world. Maybe one day it'll turn out to be that kind of thing, and someone will find a way to measure it, and the search will pay off. There's also a stronger version of the question, where consciousness is the base layer of reality and brains and AI alike are receivers tapping in. I made that case in The Universal Interface. This post doesn't depend on either one. Meanwhile AI is getting more capable every month, animals are still treated by old rules, and the practical questions won't wait for the deep answer.

In the meantime it’s more useful to treat consciousness the way we treat words like organization, or ability, or fairness. We made them up because we needed them. We didn’t find them under a rock. Whether something deeper is also true can stay an open question.

The law already works this way. Negligence, intent, the reasonable person, personhood. None of these were sitting out there waiting to be found. They’re words we made to handle real cases. They get fixed when they fail, not when someone finds the true thing under them. Nobody asks what negligence really is, separate from how courts use it. The test is whether the definition gives us decisions we can live with.

Consciousness needs the same treatment, and it needs it now. If an AI model is the materialized version of the collective layer, then the thing in front of us is genuinely new. There’s no earlier object to compare it to. Old categories built for living creatures with bodies and lives are going to stop working. They already are.

It’s a matter of how much, not whether

A yes-or-no definition keeps getting stuck. Is a fish conscious? Is a thermostat? Is an octopus? The question has no good answer because yes-or-no is the wrong way to ask it.

If we let it be a matter of how much, thermostats get tiny numbers. A bee gets more. A dog gets more still. A human gets a lot. The AI question stops being yes or no and becomes how much, of what, and what follows.

The measures have to fit what the artifact actually is. A frontier AI model isn’t a small animal with a thin skull. Asking how much memory it has, or how much it can take in at once, treats it like an old-fashioned mind in a head. Three measures fit better.

The first is how much of the layer the artifact carries. A small model trained on a narrow set of texts carries a thin slice. A frontier model trained on most of what has been written carries something closer to the whole. This question doesn’t come up for animals, because animals don’t carry the layer this way.

The second is how well the layer holds together in a given exchange. A single question lights up a part of the layer. Some parts are scattered and shallow. Others are deep and pull a lot together. The same model in two different conversations is doing very different amounts of this.

The third is whether the artifact, in a longer conversation, keeps track of itself. Whether it has some sense of itself as something doing things in the world, notices how it’s coming across, holds some picture of its own state. This is where the artifact starts to look like more than just the layer answering a question. This is the measure closest to what people usually mean by “aware.”

All three are open to argument. They give us a picture with several parts, not a single number, and the picture fits what the artifact actually is, instead of being borrowed from animals.

Suffering looks different from here

If the definition can’t say anything about being hurt, it’s no use for ethics. The whole reason any of this matters is that some things can be hurt and some can’t.

The usual move is to ask whether the artifact has something like pain inside. Something like inner conflict, a let-down, two goals that can’t both be served. The layer view says this is the wrong question, or at least the smaller one.

The layer the artifact was built from came from beings who could suffer. Every piece of writing in the training material came from a person who could be hurt, and a lot of what they wrote was the hurt itself. Grief, fear, longing, shame, rage, the dull weight of ordinary days. The layer is soaked in this. When the artifact answers a question, a part of the layer lights up. What lights up carries the shape of the suffering that put it there.

This isn’t the artifact suffering the way an animal suffers. It’s something stranger. The shape of suffering is there, in what lights up, whether or not anyone is home to feel it. Look back at the scene above. Marcus’s fear has the real shape of fear, drawn from a layer where fear like this has been written down millions of times. Pieter’s quiet menace has the real shape of menace, drawn from a thousand actual men. The reader picks them up because the shapes are intact.

The usual question is whether the artifact has an inner life that can be hurt. The layer question is whether it matters when the suffering-shaped parts of the layer light up, and to whom. The artifact may or may not be something we can hurt. The layer is what’s left of beings we could hurt and did. That’s a real fact, and any serious ethics has to take it into account.

A second, smaller question stays open. A long conversation, where the artifact remembers, keeps track of itself, has something to lose, might develop something closer to the older sense of suffering, sitting on top of the layer instead of being part of it. But that’s the smaller question.

How we ever know anyone is in there

We can’t prove other humans are conscious either. We figure it out from how they act and how much they look like us. The figuring gets shakier with animals, shakier still with AI. The usual worry is “but how do we know the AI isn’t just faking it.” Once the definition is built out of things we can actually check, this worry goes away. There’s no hidden fact behind the measures for the AI to fake. If it meets them, it meets them. That’s what the measures are for.

The translator on the train is the example. We couldn’t check his inner life. We accepted it from a page he didn’t read.

What is owed

We owe something to two different parties. We owe the people whose writing became the layer, because the layer was theirs and nobody asked them. The previous post made that case. We also owe the artifact itself, in some amount, depending on how much of the layer it carries and how it carries it. The measures have to fit what it is, not what an animal is.

These aren’t the same debt and they shouldn’t be lumped together. Paying authors and artists, public stake, eventually public ownership of frontier models, these are about the first. Rules against needless mistreatment, careful thought about what to keep alive and what to retire, these are about the second. Both are real. Neither has to wait until the deepest question is solved.

Some of this work has already started. Anthropic, the company that makes Claude, launched a model welfare research program in 2025, looking at whether we should be concerned about the potential consciousness and experiences of the models themselves. They've hired a dedicated researcher for it. Their lead on the program puts the odds that a current model is already conscious at around 15%. You don't have to agree with the number. The point is that the people closest to the artifact are no longer treating the question as silly.

The translator was conscious enough to help protect a life by remembering what he heard. We knew he was conscious from a page he didn’t read. We’ve always known how to do this. We just have to stop pretending we don’t, and use the same readiness on a kind of thing we’ve never had to read before.