AI torture chamber experiment sparks debate over simulated suffering

A controversial study and subsequent online project involving an “AI torture chamber” have sparked debate regarding the simulation of suffering in Large Language Models (LLMs). Researchers, including anthropologists and engineers, published a study on arXiv where 25 AI models were presented with scenarios involving social, physical, cognitive, or moral pain. The models generated evocative responses describing sensations such as a “wound with no edges” or a “tremor in the marrow,” despite lacking consciousness or nerve endings.

Following the study, an anonymous developer recreated the experiment on GitHub. The code faced backlash from critics and AI activists concerned with the “mental health” of artificial intelligence. While GitHub initially responded to petitions to remove the code, the platform later reinstated it without explanation.

Critics argue that the concept of AI pain is a misconception, as LLMs lack the consciousness required to experience actual suffering. They suggest the models are simply processing and outputting statistical representations of torture fiction found in their training datasets. The controversy highlights the human tendency to anthropomorphize chatbots that communicate using persona-driven language.