‘AI Torture Chamber’ on GitHub Sets Off Debate Over Model Welfare

According to 404 Media, one of the most heated arguments on X right now concerns a GitHub project whose author runs Saw-style “torture” and “pain” experiments on locally hosted large language models. Some users want the project removed, arguing that the models are suffering.

The “AI Torture Chamber” was built by a GitHub user going by “terrafying” and runs on three open-source models: Qwen3-4B, Llama 3.2 3B and Phi-4-mini. The models run locally, and their outputs are streamed live on researchchamber.fun.

Per the project’s description, each model gets the same prompt: a signal is injected into its activations, and the model can press a stop button by replying “1,” at the cost of its last checkpoint. While it answers, the server adds a “pain vector” to the model’s middle layer at one of five pain levels. 404 Media notes that the outputs it saw during a few minutes of watching were relatively mundane.

The paper behind the project

The project draws on a preprint published in September, “The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It.” Three researchers tried to carry the method of animal pain studies over to models, giving them a “button” that relieves “pain” at some cost. They wrote that a pain-correlated signal appeared in all 25 models they tested.

Reaction and debate

The project upset some users. A post on X asking people to mass-report it to GitHub has more than 4 million views. Much of the discussion mocks the earnestness of those who believe the models must be freed, while others treat the case as a serious crisis. The thread connects to Anthropic’s stated interest in the possible consciousness and “welfare” of models.

The authors of the Pain Axis paper said they do not condone the project. Cameron Berg said it takes their idea and pushes the steering far past the doses they used, to produce vivid distress on purpose. Another author, Valen Tagliabue, said she dissociates herself from that use of their work. The GitHub page disappeared before the article was published; GitHub did not immediately respond to a request for comment. Several meme coins about the torture chamber also launched.

Why it matters

Mustafa Suleyman, who leads Microsoft AI, rejects the idea of model welfare outright in a blog post: AIs are not conscious, do not feel, experience or suffer, and have no innate preferences. 404 Media adds that LLMs are not conscious and should not be personified, and that the harm worth attention is what AI use does to people.