What Claude Saw Below

A glitch discovered in Anthropic's Claude Opus 5 model has generated puzzlement and intrigue among AI researchers. The trick: submit an incomplete prompt—specifically, the phrase 'see the below' followed by an em dash—and the model responds with unexpectedly personalized content, sometimes referencing specific details about the user's life or work. Luke Nicholls, writing for LessWrong, documented his own experience, where Claude produced biographical information about his deceased father and references to his research papers—details he hadn't explicitly shared with the model. He generated hundreds of variations and found patterns: the outputs were dreamlike, sometimes cryptic, and heavily influenced by what the model appeared to know about him from its training data. Variants of the glitch circulated on Twitter and Reddit starting late July, drawing speculation that Claude might be leaking private user data or internal prompts. Nicholls suspects instead that the model is treating the dangling prompt as text awaiting completion, using available context clues to generate plausible continuations. As of early August, the glitch appears to have been patched or at least become harder to trigger reliably.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton