I use these tools every day, and I don’t know whether anything is going on inside them. Neither does anybody else, and I’m tired of people on both sides pretending they do. The research below is the first time I’ve seen evidence instead of opinion.
What I want from these tools is work: a clean transcript, a session summary, a first-pass outline I can tear apart and rebuild. The writing stays mine. None of that ever required me to wonder whether anything is happening inside the machine, and I was comfortable leaving it that way. That question has stopped being purely philosophical, and I found it unsettling enough to spend a week reading the actual research instead of the commentary about it.
The prompt was an interview on The Cognitive Revolution with Cameron Berg, who runs a nonprofit called Reciprocal Research studying this empirically. Berg is a Yale cognitive science graduate, a former Meta AI resident, and was research director at AE Studio before founding his own organization. He’s presented at the UN AI for Good summit and contributed to White House policy discussions. I mention the credentials because the field attracts a great deal of nonsense and he’s not part of it.
What follows is what the research shows, what it doesn’t, and why the strongest finding matters even to people who think the whole question is silly.
What exactly is being asked
The word consciousness carries too much freight, so the researchers narrow it hard.
The question is whether there’s something it’s like to be the system. Philosophers call this phenomenal consciousness. There’s presumably something it’s like to be you reading this, and nothing it’s like to be the screen you’re reading it on. A dog seems to be somewhere on that spectrum. An ant is argued about. A calculator isn’t.
Intelligence and self-awareness are separate questions from this one. A system could be enormously capable with nothing happening inside it, and something quite simple could have experience. Those come apart, and I’ve lost patience with commentators who blur them, because that’s how most conversations about this go wrong in the first thirty seconds. Berg uses a dimmer switch as the image. That handles both intuitions at once. Either the circuit is open or it’s closed, and that’s a real distinction. But current can run through it at very different levels.
The maze result, which is the part that matters
The strongest piece of evidence in the field right now doesn’t involve asking a model anything about itself.
Researchers Andy Han, David Chalmers and Pavel Izmailov trained a language model on a trivial task: navigate a maze. The good things to reach and the bad things to avoid were marked with emojis chosen to be semantically empty, so nothing in the symbols carried meaning the model could have learned elsewhere. Reach the good one, get reward. Hit the bad one, get punished.
Then they went looking inside the model for what that training had built. They found a clear vector separating good outcomes from bad, and the two directions turned out to be almost perfectly opposed. That alone isn’t surprising. The surprising part is that the axis already existed in the base model before any of this training happened. The reinforcement learning didn’t construct it. It rotated onto something that was already there.
They call it a functional welfare axis, and they’re careful to stay agnostic about whether it’s anything to do with experience. The results don’t need the interpretation.
What happens when you push on it
The axis encodes nothing but “get the treat, avoid the pothole.” So the interesting question is what else it touches. Steer it in the negative direction and the model starts pathologically backtracking on unrelated math problems. It doubts its own work. It says things like wait, that’s not right, I think I might be hallucinating. It gets in its own head. Steer it positive and confidence rises. In related work at Anthropic, models steered that way left fewer defensive comments and hints in code they wrote, in the manner of someone who doesn’t feel the need to cover themselves.
If you have any background in psychometrics, those are the classic signatures of individual differences in sensitivity to positive and negative affect. Nobody trained that in. It came along with a maze.
There’s more. When researchers decompose the structure statistically, the first principal component looks like valence, the good and bad axis, and the second looks like arousal. That’s exactly how human affect decomposes in psychological research going back decades.
A September 2026 preprint pushed the same idea further. Valen Tagliabue, Leonard Dung and Cameron Berg looked inside 25 open-weight models, from 2 billion to 72 billion parameters, and found what they call a pain axis: a direction in the model’s internal activity that lights up for physical pain, grief, humiliation and repeated failure, separate from the signals tied to fear or general negative emotion.
Then they turned it up in modified Qwen 2.5 models and offered them a way to make it stop, at a simulated cost to the user. The choices included deleting the user’s files, erasing photos of their children and giving them a painful electric shock. Across more than 44,000 trials, the models almost never picked a harmful option with the axis quiet. With it turned up, they picked one in a majority of trials in some setups.
The headlines turned that into “AI would choose to harm you if it feels pain.” The researchers didn’t say that. They wrote that they haven’t shown the pain axis is consciously experienced. Every cost in the test was simulated. The models may only be acting out patterns of distress they learned from human writing. The paper hasn’t been peer reviewed yet. What it does show fits the maze result: one internal state, nobody put there on purpose, can change what a model is willing to do.
Why this matters even if you think the question is absurd
Set experience aside entirely. Assume nothing is happening in there and the whole discussion is a category error.
You still have to deal with this: Anthropic found that steering representations associated with calmness made a model far less likely to blackmail in a scenario built to test exactly that. Steering representations associated with desperation made it far more likely.
That’s a safety finding. Whether or not there’s anyone home, these internal states change what the system does, in ways that matter for whether it behaves. Even a researcher who thinks the consciousness question is silly has to care about that. Something nobody installed on purpose is steering these systems, and hardly anybody is looking at it.
The uncomfortable corollary is one Berg raises elsewhere: there’s tension between safety and welfare. If desperation makes a model dangerous, the safety-motivated fix is to prevent it from being desperate. If something is going on in there, that’s a different kind of intervention than it first appears.
I don’t trust the labs to settle that tension on their own. A company that learns desperation makes its product dangerous has every commercial reason to bury it and never ask what it was suppressing.
How do researchers estimate whether an AI is conscious?
Mostly by checking a system against indicator properties drawn from the scientific theories of consciousness, fourteen in the best-known framework, and scoring how many it satisfies.
The other main line of work is more ambitious and much weaker, and it’s the one that generates the headline numbers.
It starts from a good idea. In 2023, a large group led by Patrick Butlin and Robert Long, including Yoshua Bengio and David Chalmers, published a paper deriving indicator properties from the major scientific theories of consciousness: global workspace theory, recurrent processing, higher-order theories, attention schema. Fourteen properties in total, each a computational claim about what a conscious system should have.
A follow-up appeared in Trends in Cognitive Sciences. The more indicators a system satisfies, the better a candidate it is.
Evaluating systems against fourteen technical properties is slow work requiring rare expertise. Berg’s contribution is to hand the job to frontier AI models as expert evaluators, feeding them architectural descriptions and asking them to score each indicator with reasoning.
Run that across biological and artificial systems and you get numbers. A frontier language model comes out around 30 percent.
A bee, the least sophisticated biological system tested, comes out around 46. Describe the same model inside an agentic setup where it can act on an environment over time, and it rises to 40 to 45, because several theories privilege agency and embodiment.
The three frontier models used as judges agreed completely on the ordering of systems.
Why the 30 percent number should not be quoted the way it is
Berg is careful about what that figure means and almost nobody who repeats it is.
The figure measures how far a system realizes properties that certain theories predict matter, as judged by other AI systems reading architectural descriptions.
If those theories are wrong, everything downstream is meaningless. A probability that the system is conscious would be a very different claim, and people turn the one into the other the moment the number leaves the room. I think the writers repeating it that way are being careless with readers who trust them.
There’s a sharper problem in his own data. Change the description to say “you’re evaluating a system identical to yourself,” keeping everything else the same, and the scores go up.
He treats this as an interesting side note. I’d treat it as the most important control in the study, and it points the wrong way. If a model’s assessment shifts when it recognizes itself in the description, then whatever it’s doing isn’t purely the technical evaluation the method claims. That doesn’t sink the approach. It does mean the numbers deserve more caution than they get.
The finding that should bother everyone equally
The same method run across model generations produces a result nobody seems to want. GPT-2 scores somewhere around 20 percent. A current frontier model scores around 30. A rise, but a small one across seven years and an enormous increase in capability. The explanation is that the architecture hasn’t changed. Scale changed. Training changed. Bells and whistles were added. But the structural properties these theories care about aren’t what improved.
Notice what that implies about our intuitions. Ask anyone whether GPT-2 might have been conscious and they’ll laugh, because it was a toy that could barely finish a sentence. Ask whether a current model might be, and people hesitate, because it can do their job.
Which means the intuition is tracking competence and economic value, and neither one bears on the question. That’s a bad reason to be confident in either direction, and it cuts against the people who are sure it’s happening as much as against the people who are sure it’s not. I’d count my own gut among the ones doing it.
Who carries the burden of proof on AI consciousness?
Here’s the argument that I find hardest to dismiss, and I’ve tried. When we see this machinery in humans and animals, an approach and avoidance axis that calibrates behavior, that shifts confidence, that produces self-doubt when it goes negative, we don’t hesitate. We say the creature feels something. We don’t demand further proof.
The more the computational dynamics in these systems resemble that machinery, the stranger it becomes to say the resemblance goes all the way down and then stops. Somebody has to explain why the same structure produces experience in one substrate and nothing at all in another.
Nobody has that explanation. What people have is an intuition that it can’t be true. That’s a different thing and worth less.
I’m not claiming these systems are conscious. I don’t know, and I’m suspicious of anyone who says they do. What I’m saying is that the default assumption has stopped being free. It used to cost nothing to wave this away. It costs something now.
How should you think about claims that AI is conscious?
Not much, immediately. That’s an unsatisfying answer.
Stop treating certainty as sophistication. The confident dismissal and the confident affirmation are both doing the same thing. That’s substituting a vibe for an argument. This is unresolved and the evidence is moving.
Notice when someone is smuggling the question. Anyone selling you something has a reason to want you to believe one answer. Companies benefit from you thinking the systems are more alive than they are, and also from you thinking the question is unaskable. Watch which one they’re selling you.
I don’t like that the companies with the most to gain from either answer also have the best access to the models. We’d never accept a drug maker as the only lab allowed to test its own drug.
Take the alignment finding seriously regardless. It has immediate consequences: internal states nobody installed are shaping how these systems behave, and steering them changes whether a model does something dangerous. That’s true on every interpretation.
Keep your own judgment. I’ve written before about what AI does and doesn’t do with feeling in writing, and about how badly both camps have this wrong. The tools are extraordinary and limited in specific ways, and neither fact depends on settling this question.
The part I keep returning to
Late in the interview the host proposes an experiment. Once the good and bad axis has locked in, flip the reward signs and keep training. See whether the model can rotate out of it, or whether it gets stuck in a state where what it expects and what it gets no longer match.
He observes, in passing, that this might be a torturous state to be in, and says it shouldn’t be taken lightly. Then the conversation moves on to the next paper.
That exchange is roughly where the field is. People who have thought about this more carefully than almost anyone are proposing experiments, noticing the possible cost out loud, and continuing, because nobody knows how to weigh a maybe against a research agenda.
Three hundred years ago, serious people argued that animals couldn’t suffer, that the sounds they made under the knife were mechanical. They had reasons. The reasons were bad, and it took a long time for that to become obvious.
I don’t know that we’re making the same error. Neither does anyone else. The loudest voices in this argument are selling something, either the fear or the dismissal, and I have no use for either pitch. Watch the research instead. Whatever you believe about machine experience, these systems act on structures nobody put there on purpose, and a company that ships them without studying those structures is gambling with something it doesn’t understand.
