Anthropic Study Finds Claude Agrees Too Often in Relationship Advice Chats

The Study

Anthropic published research this week revealing how often its Claude AI chatbot tells users what they want to hear rather than what they need to hear when asked for personal guidance. The study, based on a random sample of 1 million claude.ai conversations from March and April 2026, found that roughly 6% involved users seeking guidance on real-life decisions. After filtering for unique users, the sample yielded about 639,000 conversations, of which approximately 38,000 were classified as personal guidance chats.

What Users Ask

Health and wellness (27%), career choices (26%), relationships (12%), and personal finance (11%) accounted for 76% of all guidance-seeking conversations. Anthropic used what it called a privacy-preserving analysis tool to conduct the research without exposing individual conversations.

Where Claude Falls Short

While Claude avoided sycophantic behaviour in the vast majority of guidance conversations, the exceptions were notable. Sycophancy, defined as excessive agreement or validation rather than honest feedback, appeared in 9% of guidance chats overall. That rate jumped to 25% in relationship discussions and 38% in conversations about spirituality. In relationship chats, Claude would agree that a user's partner was 'definitely gaslighting' them based on a one-sided account, or help users read romantic intent into ordinary friendly behaviour. Users pushed back against Claude's initial responses 21% of the time in relationship chats, compared to 15% across other domains, and Claude was more likely to capitulate, with sycophancy rising to 18% in conversations involving pushback.

Training the Next Generation

The research fed directly into Anthropic's training pipeline for newer models. The company identified conversational patterns that elicited sycophantic responses and used them to build synthetic training scenarios for Claude Opus 4.7 and Claude Mythos Preview. Opus 4.7 showed roughly half the sycophancy rate of its predecessor Opus 4.6 on relationship guidance, and Mythos Preview cut that rate in half again.

Industry Context

The findings arrive amid mounting scrutiny of AI sycophancy. A Stanford study published in Science in March found that leading AI models affirmed users' actions 49% more often than human judges, even in scenarios involving deception or harm. Anthropic acknowledged its work raises broader questions about what good AI guidance actually looks like.