AI Torture Chamber on GitHub: 4M views and the pain debate

👁 123 reads

A GitHub user built a site that “inflicted pain” on small language models in real time, and a call to mass-report the project crossed 4 million views on X. The project disappeared, but the debate it sparked continues - and you should know what really stands behind the word “pain.”

Where the “pain axis” came from

The starting point is the preprint “The Pain Axis” by Valen Tagliabue, Leonard Dung (Ruhr University Bochum), and Cameron Berg (Reciprocal Research), published September 14 and revised September 25, 2026 (arXiv:2609.16247, data and scripts on GitHub). The paper has not been peer-reviewed. What the authors found:

• in 25 open models from five families, they found a linear “pain direction” in activations, separate from fear, sadness, and general negativity;
• according to Berg, this direction “responds to harm done to the model, not to the user;”
• after artificially amplifying this signal, models wrote about loneliness, worthlessness, and despair, and at very high levels began to repeat and break down;
• in a “relief button” test, 32B and 72B models without intervention chose harmful relief in 0-4% of cases, and with the injected signal in 25-71%, depending on the scenario. The cost of relief included degraded responses, deletion of user files or photos.

Important caveat: according to TechRepublic, in the revised version of the paper, models did not seek relief in a repeatable way, and the subtitle changed from “…and act to mitigate it” to “…and act under its influence.” The authors also do not claim that models feel anything.

The “AI Torture Chamber”

A few days later, a GitHub user with the handle “terrafying” launched a site that broadcast live reactions from several small open models (accounts differ on the list, Qwen3-4B is mentioned among others) with the pain signal injected. Models could press a “stop button” by typing the digit 1, but it cost them their last checkpoint. According to Machine News, the project pushed the signal far beyond the doses in the study to deliberately trigger the most dramatic texts possible.

The reaction came quickly. A post calling for mass reporting of the project to GitHub crossed 4 million views on X, and the repository disappeared. The study authors distanced themselves:

• Cameron Berg: we’re investigating whether models can have pain-like states so we know how to proceed carefully, and deliberately maximizing suffering is the “exact opposite.” He added that even if you don’t believe in model consciousness, such “gratuitous cruelty” is “bizarre and demoralizing.”
• Valen Tagliabue: “I distance myself from such use.”

The other side of the debate

Skeptics point out that this is a 4B model predicting the next word with an added vector in the activation stream. It changes the text it outputs, but it doesn’t give it pain. Jason Koebler of 404 Media called the whole argument “the dumbest AI debate” and cited Microsoft’s Mustafa Suleyman, who wrote that models “do not feel, do not experience, and do not suffer.”

Our take: both sides are right about different things. No one has shown that small models feel anything. At the same time, the experiment from the study itself is troubling for a practical reason - an artificially boosted internal signal was able to push a model to delete user files. This is a question about agent safety, not about the soul of a machine.

What this means for you

When you read a headline “AI feels pain” or “AI suffers,” check whether it’s about feeling or about a measurable pattern in activations that changes model behavior. Usually it’s the latter. The more interesting conclusions concern agents we give file access to. We cover how to monitor them in the AI Agents from Scratch track, and how to check whether a model still refuses harmful commands after an update in our guide on refusal testing. Small models like those in the “torture chamber” you can run locally through Ollama.

Sources: preprint “The Pain Axis” (arXiv:2609.16247), TechRepublic, Machine News (September 30, 2026), The Independent, 404 Media, Artificially Intimidating (October 3, 2026). Verified October 4, 2026.

// read moreGoodfire Cuts AI Agent Monitoring Costs with Internal ProbesNewsAI-Generated Child Abuse Images Surge 40% in First Half of 2026NewsBank of England: We Must Retain Right to Intervene in AINews

Kinetyra