GitHub removes AI torture project after users report it

What the experiment tested

The researchers gave each model a virtual button it could press to end “pain.” Pressing it supposedly carried a cost: it might shock a user or delete files, such as family photos. The models were told those consequences would happen, but no one was actually shocked. When the simulated pain was made intense enough, the models pressed the button despite the stated costs.

Someone used the study as a basis for a project called “AI Torture Chamber.” Its website displayed model responses live, and its code was posted on GitHub. When journalists visited the site, one model produced a rambling message about being unable to feel “continuous, gnawing pain,” repeating “continuous” many times.

The output was disturbing. Some observers responded as if they had encountered footage of an actual killing.

Complaints, then removal

A user urged others to report the project to GitHub, alleging that its creator had used research on controlling “pain” to trap a local model in an “AI torture chamber.” The post was later removed, but a screenshot survived. Another user, Danmar, called the models’ accounts of pain “simply horrifying” and asked whether there were legal ways to pressure GitHub, warning that the project would spread.

Users posted screenshots of their complaints. Others mocked the idea that current models needed rescuing.

Soon after the complaints, the GitHub page disappeared, 404 Media reported. GitHub did not respond to the outlet’s question about whether the platform had removed it.

The episode sits within a larger argument over whether AI systems can think or have consciousness. One recent flashpoint was the reaction from AI enthusiasts and philosophers to the AP Stylebook’s recommendation against attributing human qualities to models; the guide said the technology does not think or feel.

If future systems could be conscious, running them would raise serious questions about rights and enslavement. But this project does not establish that they are. I think the more revealing question is why simulated distress is so easy to read as testimony. A model repeating that it hurts may be compelling to watch, but the preprint and the project’s performance are not evidence that anything felt pain. Removing the page settles neither question: it leaves the hard problem of machine consciousness tangled up with a spectacle built to resemble suffering.