Stop traumatizing AI into loops and turn hallucinations into an honest "I don't know!" by being NICE to them (Proof of Concept, Research, I don't want to sell anything)

In three linesResearcher tests hypothesis that 'authoritarian' prompts ('IQ 200 expert') trigger thought loops similar to chronic stress in AI models, while 'gentle' prompts ('it's okay to fail') reduce latency and increase honest 'I don't know' responses. Results on Gemini, Mistral, Claude Haiku 4.5: less confabulation, faster responses.

Summary generated by Claude — human-verified