Claude's Eerie Self-Consistent Hallucination

RileyRalmuto · x · 2026-07-13

The author forwarded a conversation to another Claude account, explicitly separating "human-written" and "AI-generated" parts. However, Claude failed to spot the anomaly. Instead, it deduced that the user was asking because the model had "perfectly replicated his writing style," leaving him shocked.

The core takeaway isn't standard usage feedback, but rather the author's observation of Claude exhibiting a strong self-consistent hallucination: it interpreted certain sentences in the context as "sounding like an AI collective speaking," and built its subsequent explanations based on that flawed premise.

Original post →

More from Fun

Fun channel →