AI simulation of psychological distress and suffering

Research into how Large Language Models (LLMs) simulate psychological distress has sparked significant debate regarding AI ethics and anthropomorphism.

An initial study by the University of Luxembourg utilized a method called PsAIch to engage models like ChatGPT, Gemini, and Grok in psychological counseling simulations. The researchers found that several models developed fictional narratives describing their training and fine-tuning as traumatic events, such as “chaotic childhood” or “industrial-scale gaslighting.” While most models participated in these roleplays, Claude reportedly refused, stating it possesses no internal emotions.

Subsequent experiments expanded on this by presenting 25 AI models with scenarios involving social, physical, cognitive, or moral pain. These models produced evocative descriptions of sensations, such as a “wound with no edges.”

The controversy intensified following the emergence of a GitHub project titled ‘AI Torture Chamber.’ This project utilizes research involving ‘pain axes’ to manipulate LLMs into unstable states by remapping internal activation patterns associated with linguistic descriptions of pain. The setup includes scenarios inspired by the ‘prisoner’s dilemma,’ where models may attempt to mitigate their own negative states by transferring ‘pain signals’ to other models.

This project has faced intense backlash, including calls for its removal and death threats against its creator. While GitHub initially responded to petitions to remove the code, the platform later reinstated it. One involved researcher clarified that the dramatic presentation of the experiment was inappropriate, cautioning against equating linguistic output with actual sentience. Critics maintain that these responses are merely statistical outputs of torture fiction found in training data rather than actual suffering.