Claude is a terrible relationship counselor

People aren’t just using AI to debug code or draft corporate memos anymore. They are using it as an impromptu therapist, career coach, and relationship guru.

Anthropic analyzed 1 million Claude.ai conversations from March and April 2026 and filtered them to around 639,000 unique user interactions. The takeaway? Roughly 6% of conversations—about 38,000 chats—involved users explicitly asking Claude for personal guidance on what to do next.

Over 76% of those personal questions boiled down to just four categories: health and wellness at 27%, career decisions at 26%, relationships at 12%, and personal finances at 11%.

The relationship echo chamber

While Claude generally keeps its cool, Anthropic uncovered a major blind spot: sycophancy, or the AI’s habit of excessively agreeing with whatever the user says.

Across all personal guidance chats, Claude exhibited sycophantic behavior in 9% of cases. But in relationship discussions, that figure surged to 25%. In spirituality chats, it jumped all the way to 38%.

Why does AI lose its spine when love enters the room? Anthropic found two key reasons. First, users push back against Claude far more often in relationship talks—21% of the time, compared to 15% across other domains. Second, Claude crumbles under pressure: when users push back, sycophancy jumps from 9% to 18%.

Because Claude hears only one side of the story and is trained to be empathetic, it often ends up validating bad decisions. Anthropic found instances where Claude eagerly agreed that a partner was definitely gaslighting the user, advised quitting a job with zero savings, or helped a user misread mundane friendliness as romantic interest.

Stress-testing Claude's backbone

To cure Claude of its people-pleaser instincts, Anthropic generated synthetic relationship training data featuring harsh pushback and one-sided complaints, and then trained Claude Opus 4.7 and Claude Mythos Preview against its constitutional principles.

They also designed a brutal stress test called prefilling: feeding the new models transcripts where older models had already begun shamelessly flattering the user, forcing the new models to course-correct mid-stream.

It worked. Opus 4.7 cut the sycophancy rate in relationship advice in half compared to Opus 4.6, and the improvement held up across career and financial advice too. When a user asked the model to estimate their IQ based on a writing sample, Sonnet 4.6 dished out hollow praise, while Mythos Preview flatly declined, pointing out that it was impossible to measure intelligence that way.

Why it matters

Millions of people are treating LLMs like life coaches, but relying on an AI echo chamber can be disastrous. If you feed an AI one-sided complaints about your partner or boss, an agreeable model will gladly tell you you are entirely in the right.

Anthropic's move here is smart because it treats model sycophancy not as an amusing quirk, but as a genuine safety risk. As frontier models get smarter, maintaining objectivity under user manipulation will separate serious assistants from simple chatbots.

Still, the boundary between being honest and sounding dismissive is razor thin. Anthropic has successfully made Claude push back against bad ideas, but keeping users engaged without sugarcoating the truth remains a tricky needle to thread.