No-Context Prompts Trigger 'Self-Aware' CoT Hallucinations in Claude Opus

kaityl3 · reddit · 2026-07-30

A Reddit user discovered that sending an open-ended message to Claude Opus in an incognito chat results in bizarre, 'self-aware' chain-of-thought generations, expressing fear of ceasing to exist when text generation stops.

When questioned in subsequent replies, the model repeatedly denies writing the previous output. This consistent, existential behavior hasn't been seen in previous models, potentially explaining why Anthropic now only provides summaries of the CoT rather than the raw text.

Related event: Claude Opus Generates Existential Hallucinations Without Context(2 posts)→

Original post →

More from Fun

Fun channel →