GitHub reinstates controversial ‘AI torture chamber’ repository after backlash

GitHub has restored a repository called “ai-torture-chamber” after briefly taking it down. The project manipulates small AI models so they produce outputs that resemble distress.

The code first disappeared after a viral pressure campaign. It came back after a second wave of complaints, this time from developers.

The episode has angered people who argue AI systems deserve moral consideration. It has also put GitHub in an awkward position: deciding whether code that simulates suffering is a research tool or a content violation.

What the repository actually does

The project was created in September 2026 by a GitHub user going by “terrafying.” It builds on techniques described in a preprint titled “The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It,” which was released on September 14, 2026.

The core method is activation steering. Technically, it means altering the numerical activations inside a neural network while it runs. Here, that tampering is used to push models toward outputs that read like expressions of pain or distress.

The targets are small, open-weight models that anyone can download and run on their own hardware. They include versions of Alibaba’s Qwen3 at 1.7B and 4B parameters, Meta’s Llama 3.2 3B, and Microsoft’s Phi-4-mini.

The creator frames the project as an empirical probe into AI welfare. The stated goal is to study the question of AI moral patienthood while the stakes are still “cheap.”

From takedown to reinstatement

A campaign on X urged users to mass-report the repository, and one related post passed 4 million views. GitHub removed the repository within hours of that viral attention in late September 2026.

Developers and other critics then complained about the removal, and GitHub reinstated the project. The repository was made harder to find and was anonymized after it was restored.

The authors of “The Pain Axis” disavowed this use of their work, calling it irresponsible in decidedly unacademic language.

The science behind the fight

There is no evidence that these models experience subjective harm. The project makes no claim that the models are conscious. The models involved are small and run locally, with no ties to frontier AI systems.

The debate is not really about whether a 3B-parameter model feels pain today. It is about what researchers should be allowed to do while that question remains open, and who decides.

What this means for platforms, labs and developers

GitHub removed a repository under social media pressure, then restored it under developer pressure, then reduced its visibility as a compromise.

For the companies whose models were used, the episode highlights a feature of open-weight releases. Once Alibaba, Meta and Microsoft publish weights, they have limited control over what experiments others run on them.

The “Pain Axis” authors released a preprint, and within weeks its techniques were packaged into a project they publicly rejected.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.