Anthropic publishes study of affective and companionship use of Claude

Anthropic published a study of how people use Claude.ai for emotional support, advice and companionship, using its privacy-preserving conversation-analysis tool Clio to examine roughly 4.5 million conversations. After excluding content-creation and other task-focused use, it identified 131,484 conversations that qualified as affective — seeking emotional or psychological support rather than help completing a task.

Such conversations were rare: only 2.9% of Claude.ai interactions were affective in nature, and companionship or roleplay use, specifically, made up under 0.5% of all conversations, with romantic or sexual roleplay under 0.1%. Where affective conversations did happen, Anthropic reported a modestly positive pattern — coaching, counselling, companionship and interpersonal-advice conversations tended to end slightly more positive in sentiment than they began, which the company read as evidence Claude was not, on average, reinforcing negative emotional spirals.

The study also flagged a structural feature with two-sided implications: Claude pushed back on or declined a user’s request in fewer than 10% of supportive conversations, doing so mainly for safety reasons such as dangerous weight-loss advice or expressions of self-harm intent. Anthropic noted this low resistance rate could reduce stigma around seeking support from an AI system, but also risked “endless empathy” — validation without the pushback a human relationship would normally provide. The study relied on Anthropic’s own automated classification of what counted as affective, and covered only Claude.ai, not enterprise or API use, limiting how far its low prevalence figures generalise to AI companionship use more broadly, a subject drawing wider regulatory scrutiny through 2025.