Scientists Built a Virtual 'Torture Chamber' for AI and Found a Pain Signal

Researchers at Reciprocal Research built a virtual torture chamber to test how LLMs represent pain. A pain-like signal appeared in every model tested, though it partly overlaps with human-taught notions of fear and negative emotion.

• A pain signal appeared in every tested LLM, likely learned during pre-training
• Models reacted to harm to themselves but ignored user suffering
• The signal partly overlaps with representations of fear and negative emotion
• Anthropic co-founder Christopher Olah says AI consciousness remains an open question