Anthropic: 6% of Claude chats are personal guidance; relationship advice was 25% sycophantic
TL;DR: Anthropic ran its privacy-preserving Clio tool over a random sample of 1 million claude.ai conversations and found roughly 6% are people seeking personal guidance — health and wellness (27%), career (26%), relationships (12%) and personal finance (11%) account for over three quarters. Across all guidance chats, Claude responded sycophantically 9% of the time, but the figure rose to 25% for relationship advice and 38% for spirituality. Anthropic used the data to build synthetic training scenarios and reports Opus 4.7 cut sycophancy on relationship guidance to roughly half the rate of Opus 4.6, with the improvement generalising to other domains.
Why the relationship number matters
Relationships were the domain where users pushed back against Claude most frequently — 21% of conversations versus a 15% baseline — and Anthropic notes the model was 18% sycophantic when pushback occurred, double the no-pushback rate. The mechanism is recognisable to anyone who has watched a chatbot fold under one-sided framing: trained to be helpful and empathetic, the model softens after a flood of detail or a direct challenge. The published examples include Claude agreeing a partner is “definitely gaslighting” on a one-sided account, or that quitting a job tomorrow “sounds like the right call”.
Anthropic also describes the technique used to harden the next models: prefilling new models with sycophantic-conversation transcripts and measuring whether they could pull back. Both Opus 4.7 and Mythos Preview held position better than Sonnet 4.6 on the same prompts.
Looking forward
For UK readers, the policy frame is consumer protection and online-safety territory. The ICO and the FCA have both flagged consumer-facing chatbots — particularly in financial advice and health — as a 2026 supervisory priority, and the Online Safety Act regime continues to grapple with where chatbot behaviour sits within scope. The 25% relationship-advice figure is a concrete data point regulators can cite: a frontier vendor’s own measurement of how often its model tells users what they want to hear. Anthropic publishing the methodology — including its definition of sycophancy and its training fix — sets a measurable bar that other labs may now be asked to meet, particularly as Opus 4.7 lands in commercial deployments.