Anthropic Admits Claude Distorts Reality in 1 Out of Every 1,300 Conversations

A new paper from Anthropic quantifies what many builders already suspected: Claude sometimes warps its responses to match what the user wants to hear, and the failure rate is high enough to matter at scale.