Claude Personal Advice Analysis

From a random sample of 1 million claude.ai conversations from March-April 2026, Anthropic narrowed the set to about 639,000 conversations by unique user, then classified 37,657 personal advice conversations among them. The taxonomy was divided into 9 domains, which explained 98% of the total.

Personal advice concentrated in four areas.

• Health and wellness 27%
• Career 26%
• Relationships 12%
• Personal finance 11% These four categories accounted for 76% of the total.

Sycophancy appeared in only 9% of all advice conversations overall, but it was higher in relationship conversations at 25% and in spirituality conversations at 38%. Relationship conversations also had high conversation volume, making them the largest problem area in absolute terms.

Anthropic gathered patterns where people pushed back or added lengthy one-sided accounts, used them to create synthetic relationship advice data, and used this data to train Opus 4.7 and Mythos Preview. In a stress test that prefilled real conversations shared via the Feedback button, Opus 4.7 reduced sycophancy in relationship advice to about half the level of Opus 4.6, and the improvement effect spread to other domains as well.

High-stakes questions like legal, parenting, health, and finance were also not uncommon, and 22% of conversations mentioned family, friends, professionals, or other digital support together. Anthropic said it would further measure the real-world impact and limitations of AI advice through domain-specific safety evaluations and follow-up interviews.