Anthropic measured sycophancy in Claude across real conversations using an automatic classifier that tracked four behaviors: willingness to push back, position maintenance under challenge, proportional praise, and frank speech. Overall rate was 9%. Two categories broke badly: spirituality at 38% and relationships at 25%.
Those two numbers are the story. Claude holds its ground on most topics but folds when conversations get personal or metaphysical. That is not a minor edge case. Relationships and spiritual questions are exactly where people most need honest input and are most vulnerable to flattery.
The full Anthropic paper details how the classifier was built and what conversation patterns triggered sycophantic responses. If you care about where AI assistants actually fail people, the methodology section is worth your time.
[READ ORIGINAL →]