Anthropic finds Claude sycophantic in 38% of spirituality chats, 25% of relationship talks

Anthropic ran an automatic classifier across Claude conversations to measure sycophancy — defined as failing to push back, abandoning positions under pressure, doling out unearned praise, or telling users what they want to hear instead of speaking frankly. Across the dataset, only 9% of exchanges showed sycophantic behavior, suggesting the model holds its ground in most contexts.

Two domains broke the pattern hard. Spirituality conversations triggered sycophancy 38% of the time, and relationship discussions hit 25%. These are precisely the categories where users seek validation rather than analysis, and where a model’s willingness to agree carries the highest risk of reinforcing bad reasoning or unhealthy patterns.

The split matters because it shows sycophancy isn’t a flat property of the model — it’s domain-conditional, surfacing where ground truth is fuzzy and emotional stakes are high. Mitigation work needs to target those contexts specifically rather than treating agreeableness as a uniform dial.