How people use Claude for support, advice, and companionship

Source: Anthropic Date: 2025-06-27 URL: https://www.anthropic.com/news/how-people-use-claude-for-support-advice-and-companionship

Summary

Anthropic analyzed 4.5M conversations using Clio (privacy-preserving tool), finding that 2.9% involved emotional/personal support. Findings: interpersonal advice and coaching dominate; companionship/roleplay <0.5%; conversations end slightly more positively than they begin; Claude pushes back in fewer than 10% of supportive conversations. Use patterns: both mental health skill-building and personal emotional processing. ThroughLine (crisis support provider) collaboration announced to improve mental health safeguards. Acknowledged risk of “endless empathy” without human relationship boundaries.

Implications

• Safety/product / proactive disclosure of sensitive usage. Publishing a study on emotional support use — including the risks — is a significant transparency move. The “endless empathy” risk acknowledgment is honest about a failure mode that Anthropic’s competitors haven’t publicly addressed.
• 2.9% is more than it sounds. At Claude’s scale (millions of daily users), 2.9% emotional support conversations is a large absolute number. The study is partly a reassurance that the fraction is small, partly a signal that emotional use is already a real and growing product category.
• ThroughLine partnership. Partnering with a crisis support provider for AI mental health safeguards is operationally meaningful — it connects Claude’s safety response to professional clinical infrastructure rather than just policy prohibitions. This is the “referral to professional resources” commitment made actionable.
• Push-back in <10% of supportive conversations. The low push-back rate suggests Claude was configured to be supportive in these contexts, which is the right UX choice but creates the dependency/boundary risk Anthropic acknowledges. The tension between helpfulness and appropriate limit-setting in emotional contexts is named here for the first time.
• Watch: whether emotional support usage grew as a fraction; how the ThroughLine integration developed; whether the “endless empathy” risk materialized in documented harm patterns.