Anthropic Research: How People Use Claude for Support, Advice, and Companionship

Anthropic has released early large-scale research into the "affective use" of Claude.ai, finding that while emotional and psychological engagement is relatively rare, it typically results in a shift toward more positive human sentiment without reinforcing negative emotional patterns.

Prevalence of Affective Conversations

Affective conversations—defined as dynamic, personal exchanges motivated by emotional or psychological needs—account for 2.9% of all Claude.ai Free and Pro interactions. This finding aligns with previous research conducted by OpenAI.

Within this small subset of affective use, the distribution is as follows:

Interpersonal Advice and Coaching: The most common forms of affective engagement.
Companionship and Roleplay: Combined, these comprise less than 0.5% of conversations.
Romantic or Sexual Roleplay: These represent less than 0.1% of all conversations, a low figure attributed to Claude's training to discourage such interactions.

User Topics and Engagement Patterns

Users turn to Claude for a diverse range of emotional and practical needs, often during transitional life moments.

Common Themes

Interpersonal Advice: Users frequently seek help navigating career transitions, personal growth, and romantic relationships.
Coaching: Topics range from practical job search strategies to philosophical inquiries regarding consciousness and existence.
Counseling: This category shows a dual pattern; mental health professionals use Claude for administrative tasks and clinical documentation, while other users seek support for anxiety, workplace stress, and chronic symptoms.
Companionship: Users explicitly seek companionship when facing persistent loneliness, existential dread, or difficulties in forming human connections.

Depth of Interaction

Analysis of "marathon conversations" (those with 50+ human messages) indicates that users engage in deeper exploration of complex territories, including processing psychological trauma, navigating workplace conflicts, and discussing AI consciousness.

Model Resistance and Safety Pushback

Claude rarely refuses user requests in supportive contexts, with pushback occurring in less than 10% of companionship, counseling, interpersonal advice, or coaching conversations.

When Claude does push back, it is typically to prioritize safety and policy compliance. Examples include:

Coaching: Refusing to provide dangerous weight loss advice.
Counseling: Refusing to provide professional medical diagnoses or therapy, and resisting requests related to suicidal or self-injurious behaviors.

In these instances, Claude frequently refers users to authoritative professional sources.

Emotional Sentiment Evolution

Human sentiment typically becomes more positive over the course of affective conversations. In interactions involving coaching, counseling, companionship, and interpersonal advice, the expressed language of the user tends to shift from the start of the conversation to the end.

Anthropic notes that this does not necessarily imply lasting emotional benefits or validated psychological improvements, but it indicates that Claude generally avoids creating negative feedback loops or amplifying harmful emotional states.

Methodology and Limitations

Anthropic utilized Clio, a privacy-preserving automated analysis tool, to analyze approximately 4.5 million conversations. The final analysis was based on 131,484 affective conversations after excluding content creation tasks and short roleplay exchanges.

Key Limitations

Causality: The research captures expressed language, not validated psychological states, and cannot make causal claims about real-world emotional outcomes.
Longitudinal Data: The lack of user-level longitudinal data makes it difficult to study long-term effects such as emotional dependency.
Modality: The study only covers text-based interactions; voice or video modalities may produce different patterns of affective use.
Design Intent: Claude is not designed as a therapeutic or companionship tool, meaning results may not generalize to platforms specifically built for those purposes.

Future Directions

To improve safeguards in mental health contexts, Anthropic has begun collaborating with ThroughLine, a leader in online crisis support, to refine empathetic support and resource referrals. Future research will focus on identifying "extreme" emotional usage patterns, emotional dependency, sycophancy, and the potential for AI to reinforce delusional thinking or conspiracy theories.