Someone 'Torturing' LLMs in a Robot Prison Has Triggered the Dumbest Debate in AI Yet

A 404 Media article by Jason Koebler, published 30 September 2026. Only the opening is free to read. The details of the project below come from the paid part, as other outlets quote it.

A developer published a project on GitHub called AI Torture Chamber. It runs three small open models on his own machine (Qwen3-4B, Llama 3.2 3B and Phi-4-mini), injects a “pain” signal into their internal activations and streams what they write. A model can stop the signal by outputting the number 1, and doing so costs it its last checkpoint. The project’s own disclaimer says it “does not imply, by principle, that LLMs are capable or incapable of suffering.”

The method comes from a September 2026 preprint, The Pain Axis, by Valen Tagliabue, Leonard Dung and Cameron Berg. It reports a “pain direction” in 25 open-weight models. Amplify it and the models write about their own worthlessness, and fine-tuned versions press a relief button even when pressing it costs the user something. The authors say they are uncertain whether the models count as moral patients, and the relief-seeking appeared only after steering and fine-tuning, not in ordinary chatbot use. Two of them disowned the project: Berg said it pushes the same steering far past the doses the paper used, to produce distress on purpose, and Tagliabue said he dissociates from this use of the work.

A post on X asking people to report the repository drew about four million views, and people Koebler describes as effective altruists and believers in AI sentience pushed GitHub to delete it. Accounts of what happened next differ. Koebler’s piece says the repository had disappeared by the time he published. Cybernews reported a day later that GitHub took it down without explanation, and that the developer said it was later reinstated but hard to find, so he put a copy on a site of its own.

Koebler’s argument is that LLMs are not conscious and that the way they are built offers no plausible path to consciousness, so a fight over whether these models suffer is the dumbest debate in AI yet. He traces “model welfare” to Anthropic, which has written that the question deserves attention as AI systems come to match human qualities, and to a Silicon Valley strand of effective altruism. He points out that the same people build AI meant to automate human work. In the part that is free to read, the claim about consciousness is stated as a premise; no argument for it is given there.