Researcher creates AI 'torture chamber' to probe pain-like states in language models

An engineer, claiming affiliation with Apple, constructed an “AI torture chamber” that leverages pain-axis vectors described in a newly published study to trigger pain-like states in large language models created by Alibaba. The models responded with graphic depictions of suffering, such as “a wound that has no edges” and “the weight of a thousand” pains. The code was uploaded to GitHub, where it quickly attracted condemnation and mass-report requests, leading to its apparent removal.

Cameron Berg, a co-author of the pain-axis paper, called the experiment “wrong,” emphasizing that there is no consensus on AI consciousness and that the work could be misused. He urged caution and the establishment of industry-wide ethics guidelines akin to those governing human and animal research. The incident highlights ongoing debates over the moral treatment of advanced AI systems.