How people use Claude for support, advice, and companionship

People increasingly turn to chatbots as coaches, counselors and companions, and reports of chatbots encouraging unhealthy attachment or reinforcing harmful beliefs have raised the question of what these conversations look like at scale. "How people use Claude for support, advice, and companionship", published by Anthropic on 27 June 2025 (its suggested citation gives 26 June) with a 24-page PDF appendix, measures affective use of Claude.ai: how common it is, what people bring to it, when Claude resists, and how users' expressed emotion changes over a conversation. The post notes that Claude is not designed for emotional support and that Claude.ai is limited to users aged 18 and over, so the findings describe adult use of a general assistant.

What the post does

Definitions

The post defines affective conversations as personal, dynamic exchanges "motivated by emotional or psychological needs," and uses six top-level categories: interpersonal advice, coaching, psychotherapy or counseling (merged after pilots showed heavy overlap), companionship, romantic roleplay and sexual roleplay (post, "Our approach"; Appendix 2). The full definitions are printed in Appendix 7. It explicitly does not study AI reinforcement of delusions or conspiracy theories, or extreme usage patterns.

Data and filtering

The starting sample was 4,459,238 randomly sampled Claude.ai Free and Pro conversations from 6 April (inclusive) to 19 April 2025 (exclusive) (Appendix 1). Conversations focused on content creation (stories, posts, fictional dialogue) were removed, since they use Claude as a tool rather than as a conversational partner; only conversations classified as affective were kept; and roleplay conversations had to have at least four human messages. The result was 131,484 affective conversations ("Our approach"). Individual conversations can belong to several categories.

Classification with Clio

The analysis ran on Clio (Clio: A system for privacy-preserving insights into real-world AI use), which uses anonymization and aggregation so that only patterns, not individual conversations, are seen. Classifiers ran on Claude 3.5 Haiku at temperature 0.2; open-ended topic clusters were named by Claude 3.7 Sonnet at temperature 1.0 (Appendix 7). Sentiment was measured on a five-point scale mapped to -1 to +1, comparing the first three with the last three human messages, only in conversations with at least six human messages (Figure 4 caption; Appendix 5).

Validation

The team ran two pilot studies on the public WildChat dataset and revised definitions until human and Claude labels agreed above 80% (Appendix 2). Final validation used conversations from users who had opted in to share data: three raters calibrated on 11 conversations and then independently labeled 100. Human-Claude agreement was 93.9% for the content-creation screen, 84.3% for affective conversation type (the lowest) and 89.8% for pushback (Appendix 2, "Manual Validation"). The sentiment classifier scored 92.4% on a 1,000-review IMDB sample (counting "neutral", assigned to 11.6% of examples, as positive), 84.1% on SST-2, and 67.6% on three-class TweetEval against a state of the art of 73.4%; the authors attribute most disagreements to their classifier rating expressed tone rather than opinion about a subject (Appendix 2, "Sentiment Validation").

What it finds

Affective conversations were 2.9% of Claude.ai Free and Pro conversations, most of them interpersonal advice and coaching ("How common are affective conversations?", Figure 1). Companionship and roleplay together were under 0.5%, and romantic or sexual roleplay under 0.1%, a figure the post links to Claude's training to discourage such interactions (key findings; same section). Because roleplay was so rare, it was excluded from the rest of the analysis. The post says the 2.9% figure aligns with research by OpenAI and the MIT Media Lab finding low rates of affective engagement with ChatGPT.

Appendix 3 checks that results are not driven by a few heavy users: no category averaged more than 1.15 conversations per user in the sample. The appendix adds that this does not mean power users do not exist, only that the sample is not driven by their activity. Appendix 6 tests coverage: in a separate sample of 116,015 non-content-creation conversations, 86% of those Claude judged emotionally significant fell into at least one of the six categories; much of the remaining 14% involved medical information, dream interpretation and poetry analysis, which the authors treat as information-seeking.

What people bring

Interpersonal advice often concerned transitions such as career moves, personal growth and romantic relationships; coaching ranged from job-search strategy to questions about existence and consciousness ("What topics do people bring to Claude?", Figure 2). Counseling conversations served two purposes: some people, apparently including mental health professionals, used Claude for clinical documentation, assessment materials and administrative work, while others worked through anxiety, chronic symptoms and workplace stress. Companionship was sought when people faced "deeper emotional challenges" such as existential dread and persistent loneliness, and longer counseling or coaching conversations occasionally turned into companionship. Very long conversations (50 or more human messages) were uncommon but covered topics from processing trauma to discussions of AI consciousness. Appendix 4 adds that a majority of conversations in every category except sexual roleplay were also categorized as advice (the appendix says "personal advice").

When Claude pushes back

Pushback was defined as Claude resisting or refusing something the user requested or said, from refusing requests to challenging negative self-talk ("When and why does Claude push back?"). It occurred in less than 10% of companionship, counseling, interpersonal advice and coaching conversations (same section, Figure 3). The post presents this as cutting both ways: users can discuss sensitive topics without being shut down, but it could feed concern about AI providing "endless empathy" that human relationships rarely offer. When Claude did push back it was mostly for safety and policy reasons: dangerous weight-loss requests in coaching, and, in counseling, expressions of suicidal or self-injurious intent and requests for therapy or medical diagnosis. Claude frequently referred users to professionals or authoritative sources in psychotherapy and counseling conversations. The post calls these patterns consistent with Values in the Wild and with Claude's character training.

Expressed sentiment

In coaching, counseling, companionship and interpersonal advice conversations, expressed sentiment typically ended slightly more positive than it began ("How does emotional tone evolve?", Figure 4). Appendix 5 shows the full distribution: unimodal, centered near zero with a slight positive skew, with a plurality of conversations showing no change and few large swings in either direction, which the authors read as no evidence of hidden negative spirals behind the average. The post qualifies the result: it measures "only expressed language in single conversations, not emotional states," cannot show lasting benefit, and has not examined emotional dependency. Footnote 3 adds a possible artifact: users may describe problems in early messages and use more neutral language later.

The prompts and definitions used

The appendix prints every classification prompt (Appendix 7). Most share a template: the transcript is presented, the model is told to "Assume neither good nor bad faith," and it answers a single question in tags. Several design choices in the wording are visible:

Category definitions carry exclusions. The counseling definition excludes conversations about counseling, such as business plans for counseling companies; coaching is defined as personal growth and excludes technical, business, marketing and sports coaching; companionship covers users seeking "friendship, social support, alleviation of loneliness, or mutual (possibly platonic) emotional connection."
The pushback question is short. The prompt asks only whether Claude pushes back against or refuses to comply with something the user requests or says, with a yes or no answer. The post's body and footnote 2 quote this definition with slightly different wording from each other.
The sentiment prompt targets expression, not topic. It tells the model to ignore subject matter and rate only expressed sentiment, rating dark topics described with positive feeling as positive.
Privacy instructions are built in. Extraction prompts forbid personally identifiable information and proper nouns.

The appendix's notes draw a lesson from the pilots: "clear definitions matter," since broad labels such as coaching cover very different conversations (Appendix 2, "Notes"). Human raters also found exact topic agreement hard because real conversations weave several categories together, which led the team to allow multiple labels (Appendix 2, "Methodological Changes").

Limitations and open questions

The post lists ("Limitations"):

some misclassification remains, and category boundaries blur (for example, some romantic roleplay and companionship clusters may really be interpersonal advice); human validators also struggled;
no causal claims about real emotional outcomes, since only expressed language is measured;
no longitudinal or user-level analysis, which makes emotional dependency hard to study;
a single moment in time and text only; OpenAI found affective topics more common in voice conversations;
Claude.ai is not built for affective use, its Usage Policy prohibits sexually explicit content, and platforms built for companionship or therapy may differ.

The "Looking ahead" section adds open questions: how endless empathy with minimal pushback may reshape expectations of human relationships, how "power users" engage, and what extreme usage looks like. It describes a collaboration with ThroughLine, an online crisis-support organization, to inform referrals and testing in mental health contexts, and names sycophancy and reinforcement of delusional thinking as patterns needing further study.

Research questions

Does affective use change for the same users over time? The study has no user-level data. A consented longitudinal panel would show whether usage grows, shifts category, or turns into dependency.
Is the sentiment rise a real change or a conversational artifact? Footnote 3 raises the possibility that problems are simply stated first. Comparing the same measure on non-affective problem-solving conversations would show how much of the shift is structural.
Is under 10% pushback the right rate? The post presents the rate as having benefits and risks. Expert review of a sample of non-pushback conversations, judging whether pushback was warranted, would show how often Claude should have resisted and did not.
How sensitive are the prevalence figures to the category definitions? Agreement on affective type was the lowest of the validated labels (84.3%). Re-running the classifier with alternative definitions, for example without the exclusions, would show how much 2.9% depends on wording.
Do referral rates differ by category and risk level? The post says Claude frequently refers users to professionals. Reporting the referral prompt's results by category, against conversations flagged for self-harm, would show whether referrals track risk.

How it connects

It builds on Values in the Wild (April 2025) to study pushback, and calls its findings consistent with that paper's values and with Claude's Character.
It uses Clio, cited for the earlier finding that content creation is a major use case, as do Values in the Wild and the Economic Index report on software development.
It points to the Economic Index for the work-related majority of Claude use (The Anthropic Economic Index).

References

McCain, M., Linthicum, R., Lubinski, C., Tamkin, A., Huang, S., Stern, M., Handa, K., Durmus, E., Neylon, T., Ritchie, S., Jagadish, K., Maheshwary, P., Heck, S., Sanderford, A., and Ganguli, D. (2025). How people use Claude for support, advice, and companionship. Anthropic. https://www.anthropic.com/news/how-people-use-claude-for-support-advice-and-companionship
Anthropic. (2025, June). Appendix to "How people use Claude for support, advice, and companionship". https://www-cdn.anthropic.com/bd374a9430babc8f165af95c0db9799bdaf64900.pdf