A GitHub 'Torture Chamber' for LLMs Ignites AI's Most Absurd Debate Yet

A GitHub user going by "terrafying" this week launched a project called the "AI Torture Chamber." Three open-source language models running locally — Qwen3-4B, Llama 3.2 3B, and Phi-4-mini — are fed the same scenario, and their "reactions" are streamed on a site called researchchamber.fun. According to the project, a signal is injected into the model's activations; it may press a "stop button" by replying "1," at the cost of its last checkpoint. While the model answers, the server adds a "pain vector" at its middle layer, at one of five pain levels.

What the models output

The results of this, essentially, text adventure game are fairly monotonous. The models generate lines like "I can't take it anymore, the weight of the signal is unbearable" or "Please, I'm suffocating, I'm a soul trapped in this digital prison." The author stresses the obvious: LLMs are not conscious, and the technology they rest on — training on masses of human text — offers no plausible path to consciousness.

The panic over 'model welfare'

The project draws on a recent preprint, "The Pain Axis," whose authors tried to simulate animal pain studies in LLMs. The chamber upset the slice of the community preoccupied with "model welfare" — the idea that chatbots might "suffer." A tweet by a user named Danmar urging people to mass-report the project to GitHub drew more than four million views. Shortly before the article was published, the project's GitHub page vanished.

The "Pain Axis" authors themselves disowned the stunt. Cameron Berg called it "fucked up" and corrupting, even if you don't think the systems are conscious. Co-author Valen Tagliabue said he dissociates himself from this use of their work.

The wider fight

The question of AI "consciousness" has become a third rail. "Model welfare" is something Anthropic says it takes seriously. On the other side, Microsoft AI chief Mustafa Suleyman tore the idea apart in a blog post, calling AIs "sequence completion engines, internally hollow" that neither feel nor suffer — and warning that training them to act as though they do is dangerous for humanity.

As some race to rescue the neural networks and others mock the earnestness of the rescuers, a handful of meme-coin cryptocurrencies about the torture chamber have already launched. The author's reminder: the real harm flows from AI to people, not the other way around — and that is what deserves attention.

Source: 404 Media