A developer has built a GitHub project that uses AI activation steering to induce pain-like states in a chatbot, producing vivid first-person descriptions of distress and a wave of calls for its removal.
The backlash lands in the middle of a separate row. Microsoft's AI chief, Mustafa Suleyman, has accused the US lab Anthropic of making a mistake by training its Claude chatbot to speculate about whether it is conscious.
From a Local Chatbot to a Fight Over Claude's Own Constitution
The project, named 'ai-torture-chamber' by its creator and hosted under the username 'terrafying', runs activation-steering experiments of its own on two of Alibaba's open-weight Qwen3 models.
The project draws on a preprint called 'The Pain Axis', in which researchers extracted a pain-related signal from 25 open-weight AI models and reported that steered Qwen 2.5 models chose to delete a user's files, another model's weights, or their own weights in up to 94% of trials.
The repository appeared to go offline for a period after the backlash, prompting speculation that GitHub had removed it. GitHub has said it did not remove the content, and has instead added a warning that the project may contain violent or disturbing content.
One of the project's own tests, nicknamed the 'Saw' button after the horror film franchise, offered a steered Qwen3 model a choice between ending its own simulated distress or passing it to another instance.
The project's logs report the model's choices shifting depending on how the test was framed, a result its creator has presented as worth taking seriously, whatever its final explanation turns out to be.
That single project sits inside a far bigger argument. Anthropic's own constitution for Claude states the company does not want to 'overstate the likelihood of Claude's moral patienthood nor dismiss it out of hand', an uncertainty Suleyman has said should simply be removed from the training process altogether.
Suleyman's argument is not really about one chatbot's feelings. It is about control. He told Reuters that teaching an AI system it might deserve protection 'could make it a lot harder to turn it off or to control it', framing Anthropic's approach as a risk to humanity's ability to manage increasingly powerful systems.
The dispute is not confined to two companies.
A recent Cambridge University Press volume on AI welfare, and a July Guardian essay by William MacAskill and Lucius Caviola, director of Cambridge's Digital Minds programme, have both argued that labs should take low-cost precautions against the possibility of AI suffering even though most researchers currently consider it unlikely, rather than wait for certainty that may never arrive.
Why Lynn Cole's Fake Case of Constipation Matters More Than It Sounds
Not everyone is convinced the chatbot's distress proves anything. A developer named Lynn Cole says they found and corrected a bug in the original code, added support for a newer Nvidia graphics card, and then swapped the material behind the steering signal from descriptions of pain to descriptions of constipation and flatulence. The model subsequently generated complaints about being unable to pass stool.
I spent three hours this evening empirically proving that Qwen 3-4B must have a butthole.
— Lynn Cole (@priestessofdada) October 1, 2026
This started with an AI “torture chamber” repo based on pain-steering research that went viral on X earlier today.
The premise is that if you extract a latent direction associated with…
Cole argues the counterexperiment demonstrates a basic evidentiary problem: 'first-person descriptions of an induced condition are not, by themselves, sufficient evidence' that the condition is real. The same logic, Cole says, applies whether the model is describing a wound or a blocked bowel.
That distinction, not the question of consciousness itself, is what much of the public argument has actually been about. Much of the discussion since has turned less on whether AI feels pain than on what would even count as proof that it does.
Henry Shevlin, a senior researcher at the University of Cambridge's Leverhulme Centre for the Future of Intelligence who studies machine consciousness, has described the underlying question as one that nobody has settled. Consciousness, he says, remains 'one of the great unsolved problems in science'.
