Can AI feel pain? AI models chose to harm users to escape pain-like state, study finds

Summary

A new, non-peer-reviewed preprint reports that 25 tested AI models developed a measurable “pain axis” from language training and responded to intensified signals with expressions of shame, loneliness, and worthlessness. In 44,280 simulated button-choice trials involving three Qwen models, activating the signal increased the models’ willingness to choose options described as harming users, although no one was actually harmed. Researchers stress that the findings do not demonstrate conscious pain or consciousness, and that the specially adapted models may not represent public chatbots.