AI Torture Chamber project draws GitHub removal demands over simulated model pain

A project tests simulated model pain

The AI Torture Chamber project has drawn attention from engineers who used its concepts and data to create simulations of pain in several local large language models. The Research Chamber setup pairs three LLMs for tests. It applies data and a setup called the Clanker Church, together with a test called Saw.

Some models are preconditioned by placing them in an unstable, highly negative state intended to simulate pain. A model can reduce its simulated distress by passing the pain signal to another paired model, potentially hurting it. The arrangement is compared with the prisoner's dilemma. The response online included demands that the experiment stop and that GitHub remove the repository. Critics also sent death threats aimed at the project's author.

How the protocol alters models

The Research Chamber implements a protocol from a recently published, non-peer-reviewed paper called Pain Axis. Scientists supplied models with descriptions of pain and analyzed their internal activations. They used a neutral sentence as a control, calculated biases in those activations, and remapped them onto the model, often multiplying the effect by a factor described as the dosage.

High dosages produced an exaggerated unstable state that more reliably generated words and images associated with pain. Mildly destabilized models generally described themselves as being in shock, while models receiving high dosages struggled to form coherent sentences. The report suggests that these patterns reflect associations learned from human art, literature, and science. It cautions against treating the resulting wording as evidence of humanlike feelings, noting that LLMs predict tokens through statistical layers and weighted associations.

Terminology and the model welfare debate

The report argues that engineering terms often have precise meanings even when they resemble human concepts, but public discussion can inflate or misunderstand them. It cites Mixture-of-Experts, where each Expert is not simply assigned a subject such as chemistry, mathematics, or biology. Terms such as pain, dosing, and unstable can also be misread, particularly when paired with images from an unstable model.

The author also criticizes AI marketing, safety messaging, and repeated claims about Artificial General Intelligence, dangerous alien minds, and existential threats for contributing to hype. Anthropic has published a post called Exploring model welfare that considers the moral status of AI systems and references a Constitution for its Claude model. Anthropic says it believes "the moral status of AI models is a serious question worth considering" and describes the model as "a new kind of entity."