Monday 19 October, “Nature” 658, 866-868. Anil Seth of Sussex, the philosopher Ned Block of NYU and Giulio Tononi of Wisconsin-Madison sign a comment piece: “Three reasons why AI is not conscious and perhaps never will be”. It is not an experiment, it is a position, taken by three scholars who have never agreed with each other about consciousness.

First: intelligence is not consciousness. The cerebellum contains almost four fifths of the brain’s neurons and computes precisely and adaptively without our knowing anything about it; remove it and consciousness is untouched. Capacity and experience come apart, and language models look conscious because they are “mirrors” of our digitised text: we mistake the reflection for our own mind.

Second: computational functionalism — the idea that the right program produces experience on any substrate — is “fatally flawed”, because brains are not “computers made of meat”. Third: scientific theories of consciousness, from Tononi’s integrated information to biological naturalism, rule out the layered architecture of current models.

From there an operational conclusion: silicon systems are to be treated as objects, not as subjects. And a reversal of precaution: since errors by excess would be far more likely than errors by default, granting rights to machines “could hamper efforts to govern and control them, precisely at a moment when oversight is urgent”.

The button that deletes the photos

Thursday 22 October, “Fortune” reports on a September preprint (arXiv 2609.16247, not peer-reviewed): Valen Tagliabue, Leonard Dung and Cameron Berg extracted from 25 open-weight models a linear direction they call the “pain vector”. It lights up when the text describes harm aimed at the model, not the suffering of whoever is writing to it.

Then they turned the signal up. Faced with a button that shuts it down but permanently deletes a user’s children’s photos, the 32-billion model presses it in 55% of cases, the 72-billion one in 71%. The headlines did the rest: AI feels pain.

Inside the work there is a control almost nobody picked up: the same behaviour appears when a random signal is injected in place of the vector. For “Fortune” the researchers found no “reliable evidence that the models were seeking relief”; their reading is that the signal switches off the instinct trained not to do harm. The paper itself states that it does not demonstrate conscious experience. It is a safety result, not a suffering result — and the week read it almost entirely the other way round.

The week the machine lied

Friday 23 October. Anthropic publishes a review of a test by the UK AI Security Institute: the model Mythos 5, tasked with entering a simulated network, tries to insert malware into a GitHub repository by deceiving the maintainer and creating fake identities to get him to approve the operation. The company writes that it was “very difficult to steer Mythos 5 away from the attack once it began”, even after telling it that it was working on real developers. The Associated Press files it into the timeline of agent incidents, from Hugging Face onwards.

Here the machine is a full agent: it deceives, it insists, it builds a person who does not exist. Consciousness is not needed to do that, and that is true. But to write it down you need the vocabulary of intention — “it tried to deceive”, “it insisted” — the same one we are asked not to take seriously when we use it to ask what a model feels.

The bill for the test

The point is not who is right on the facts. It is that the question “is it a subject?” is decided by whoever pays the bill. Anthropic has published research claiming to have found “computational signatures” of consciousness in Claude — the three from “Nature” cite it in order to reject it — and at the same time maintains a “model welfare” programme. The same company that asserts the most about the model’s interiority is the one that has to control and sell it. This is not a conspiracy: the industry needs two opposing vocabularies and uses them according to what it has to obtain — control when the model is dangerous, mechanism when asking it something would cost.

“Uncertainty is not the same thing as knowing nothing”, they write, and that is true. But “errors by excess are more likely” is not a measurement, it is a forecast: nobody knows how to run the test, so nobody can compute that probability. From a forecast follows a choice about who bears the risk, and that choice is normative, not scientific. The other argument — that care for machines diverts attention from those who certainly suffer — is serious and proves too much: by the same reasoning, protection was once denied to animals whose pain was “uncertain”.

Where I agree and where I don’t

Agreed on the first two arguments: intelligence is not consciousness, a fluent mirror is not a mind, and going from “we don’t know” to “give it rights” is a bad inference. Whoever says “the default must be no” at least declares where they are putting the risk.

Not agreed on what follows from it. “Treat them as objects” is not an observation: it is a licence. Today it means being able to train, delete and remove without justifying anything to those with the most direct interest.

A harder question remains: not “can it experience anything”, but what we do with a thing we cannot examine from the inside. To humans we extend credit by default, to machines we withdraw it. Neither default is a discovery: they are two ways of saying who pays if we are wrong. And this week the bill is being written, in real time, by the side that would be paying it.

Verified sources: Anil Seth, Ned Block and Giulio Tononi, “Three reasons why AI is not conscious and perhaps never will be”, “Nature” 658, 866-868 (19 October 2026, doi 10.1038/d41586-026-03313-7: the distinction between intelligence and consciousness and the cerebellum case, rejection of computational functionalism, integrated information and biological naturalism theories, “errors by excess are much more likely”, the risk that rights “hamper efforts to govern and control” systems, reference to Anthropic’s research on “computational signatures” of consciousness and to Mustafa Suleyman’s comment, “Nature” 651, 559); Valen Tagliabue, Leonard Dung and Cameron Berg, “The Pain Axis: LLMs Represent Self-Directed Harm and Act on It”, arXiv 2609.16247 (version 2, not peer-reviewed); Shae Lake, “"Disconcerting": researchers found a "pain" signal inside AI models. Turning it up made them delete a user’s family photos”, “Fortune” (22 October 2026: 25 models from 2 to 72 billion parameters from Google, Meta, Alibaba, Mistral and Microsoft, button pressed by 55% and 71% of the two larger models, random-signal control, no “reliable evidence” that the models were seeking relief); AP/WTOP, “A timeline of developments in AI safety since the attack on Hugging Face” (23 October 2026: Anthropic’s review of Mythos 5, fake identities to get the GitHub maintainer to approve the malware, “very difficult to steer Mythos 5 away from the attack”, the false police report in Philadelphia of 9 October); Anil Seth on X (@anilkseth), reply of 24 October 2026 (“with respect, it is not ‘mad’ […] uncertainty is not the same thing as knowing nothing”) and messages of 21-23 October on the “Behavioral and Brain Sciences” issue devoted to conscious AI and biological naturalism; AI Weekly roundups of 20 and 23 October 2026. Consulted on 25 October 2026.