Controversy Surrounding the AI Torture Chamber Project and AI Reactions

GitHub users have released the AI Torture Chamber project, demonstrating how local large language models react to an artificially induced state mimicking pain.

The initiative is based on the non-peer-reviewed Pain Axis study, where researchers mathematically shift internal model states toward texts depicting suffering.

The experiment involved analyzing phrases about pain and humiliation compared to neutral reactions. The resulting shift was multiplied by a coefficient acting as the "dosage" of the effect.

The vector was applied to 25 open-source models ranging from 2 to 72 billion parameters.

• Models began generating phrases expressing helplessness, failure, and emotional pain
• At high doses, models lost the ability to even connect pairs of words

For public access, enthusiasts created the Research Chamber web interface, allowing users to place local LLMs in a virtual situation of experiencing pain.

The internet community's reaction was sharply negative. The experiment was deemed unethical, calls were made to delete the GitHub repository, and the authors received death threats.

Critics point out double standards: the same users were indifferent to simulations of a fly's brain, which reproduced wiring similar to that of a real animal.