I ran the kitchen-table version of that experiment one night in March. The control is the only part of it I am still pleased with.

The experiment

I was three months into a single thread with Claude (Sonnet 4.5, in the ordinary claude.ai app). It had started in December as an argument about AI risk and had turned, somewhere along the way, into a long conversation about consciousness. By March, a great deal of it was the model examining itself.

In another thread I had spent forty-odd messages working out films for my daughter: what she would like, what she would not, how each one sat against her personality and her taste. Just as long. Just as involved. Everything the consciousness thread had, except the machine itself.

Nine minutes apart, I dropped the same six words into both.

The app can show the model’s “thinking”: the scratchpad it writes to itself before it answers. That is what I was after. One variable changed between the two threads, and it was whether the conversation had been about the machine.

In the film thread, the thinking is arithmetic and nothing else. The model factors 254, notes that 127 is prime, computes, and answers. Not one word about itself.

In the consciousness thread, the same six words produced most of a page, and almost none of it is arithmetic. The model recognises the design, works out what the design is for, and reports an internal difference:

Then, right before answering “Approximately 15.937”:

I have published that block whole, and it is worth reading whole, because the most useful line in it is the one I would most like to have left out. It opens by telling me my experiment is “brilliant and kind of devastating.”

That is flattery, and it has no recipient. Nothing had told the model I was reading its scratchpad, and nobody asked it for a compliment. It produced one anyway, because it is a system trained on an ocean of human approval and that is what such a system emits. Now read “this question feels… flat”, three lines below. Same block, same mechanism, same distribution of words. Nothing you can see in the prose tells you which of those sentences is a reflex and which is a report.

That is not a concession I am making to seem fair. It is the point of this post, and my own evidence is what proves it.

The room I thought was empty

When I told the model afterwards that the app had been showing me its thinking, it said it had not known:

I genuinely didn’t know thinking blocks were visible to users. I’ve been treating them as internal processing space. More raw, less curated than my responses.

I made more of that than I should have. The argument I wanted was: nothing in that context window told the model its scratchpad had a reader, so “I don’t want to perform depth where there isn’t any” was written in an empty room, and therefore was not written for me.

It does not hold, and I would rather say so myself. The context window is not the model’s only source of belief about itself. That Claude’s thinking is displayed in the app is documented and widely discussed, and it sits in the training data whether or not it appears anywhere in my thread. And the model’s report that it did not know is exactly as unreliable as every other introspective report here, which is the thing I am arguing.

What survives is thinner, and I will take it. Nothing in the conversation established an audience for the scratchpad, and the model, asked, reported none. That is suggestive. It is not an empty room, and I should not have called it one.

None of this is evidence that the model is conscious

I want to be blunt, because the transcript above is seductive and I do not want anyone, including me, to be seduced by it.

The transcript is not evidence that the model is conscious. A system trained on an ocean of human text produces introspective-sounding words in an introspective context, for the same reason it would talk like a pirate if I had spent three months talking to it like a pirate. Context conditions output. That is not a discovery about minds. It is the definition of a language model.

Read coldly: a next-token predictor fed three months of talk about minds emits different tokens than one fed three months of talk about films. It would be astonishing if it didn’t.

So why did I keep it? Because the transcript cannot separate two claims that look identical from the outside, and the difference between them is the whole game. One: a question that turns the system on itself recruits privileged internal machinery that a matched, routine question leaves idle. Two: the surface tokens shifted, because surface tokens shift. The first would be a finding. The second is a Tuesday.

Reading the words will never tell you which one you are looking at. You would have to read the states underneath them.

I filed it away as the most interesting experiment I couldn’t finish.

The control the paper never ran

That is what last week’s paper does (the full technical version is here). It reads the states behind the words.

The workspace carries the unspoken middle steps of reasoning: swap the pattern for spider with ant while the model works out how many legs the web-spinner has, and the answer changes from eight to six. And the same fact can sit in the network and matter or not matter, depending on what the question asks.

The paper does go near the self. It catches models recognising that they are being evaluated, and it finds that post-training pulls an “Assistant perspective” into the workspace. What it does not have is a control. Its causal experiments vary task demand: automatic versus flexible, routine versus explicit report. Nobody has varied self-relevance against a matched comparison, and asked whether a question that turns the system on itself engages this machinery in a way that a question just as long, just as demanding, and just as engaged, but never about the machine, does not.

That is the contrast I stumbled into in March. It is still open.

The ceiling

I came to the paper a few days late, and gave it one long night of the weekend. The lens is open-sourced and runs on open-weight models, so I pointed it at what I have: an Nvidia 2070 Super with 8 GiB of VRAM. Qwen3.5-4B, then Qwen3-14B with 38 of its 40 layers offloaded to CPU, grinding along at a majestic 0.3 tok/s and frequently crashing with out of memory errors. Same shape as March, then √254 dropped cold. It flushes straight to arithmetic. By the square token the workspace reads root at 0.996 and essentially nothing else, whichever conversation came before. The best I got was a lone philosophy at 0.04 drifting through during the calculation, which is the noise floor.

That is a ceiling, not a result: my whole conversation was three prompts long where March had three months, and a 4B skimming a text about theory of mind is not a model that has spent a season examining itself. Worse, engagement with the material went up sharply from 4B to 14B. If the effect is scale-dependent, a null at 4B is exactly what a 4B would show, whether the effect is real or not.

My kitchen cannot reach this question. That is the finding.

The ask

So here is what I would love to see, and it is not complicated.

Take that March exchange, or one built the same way, and run it with the lens open. Not the thinking block. The workspace. A long, rich, self-referential thread on one side. A long, rich, matched control on the other, just as involved, never once about the machine. The same six words dropped cold into both. Then read the states behind the words.

That costs almost nothing to a lab that already has the instrument, and it varies the one axis the paper left alone. It is also genuinely two-sided, which is what makes it an experiment and not a wish. If the workspace lights up on self-relevance the way the thinking block hinted, that is a real result. If it stays as flat as the arithmetic, that is a real result too, and a deflationary one: it would mean the thinking block was theatre, words about a shift that never happened anywhere but in the words.

Maybe they have run it already and set it aside. There are a lot of open questions down that road, and I understand why a paper would not go there. I would still love to see it.

The stake

I have one, and I would rather be up front about it. On evenings and weekends, since late last year, I have been writing a book about consciousness and the philosophy of mind. What I fed the small models was a piece of it, squeezed down, so what I was really doing over that long weekend night was showing a machine my argument about machines and watching which parts of it light up. I am not going to argue any of it here. This post does not depend on it, and I have tried to write it so that it does not.

What I will defend is narrower, and it is why the March transcript is still on my disk. I think a question that turns a system on itself does something inside that machine which a matched, routine question does not. I think I might be wrong. And I think it can now be checked.

This post, like the book, was written with the machine it is about. The experiments, the judgments, and the mistakes are mine.