Does emotional conversation with Claude help users feel better?

Anthropic's own analysis of Claude.ai conversations finds that "affective conversations are relatively rare, and AI-human companionship is rarer still." The company reports that "only 2.9% of Claude.ai interactions are affective conversations," that "companionship and roleplay combined comprise less than 0.5% of conversations," and that "less than 0.1% of all conversations involve romantic or sexual roleplay." Within the affective conversations that do occur — coaching, counseling, companionship, and interpersonal advice — Anthropic reports that "human sentiment typically becomes more positive over the course of conversations," and that such interactions "typically end slightly more positively than they began."

The mechanism Anthropic gives is low resistance: "less than 10%" of coaching or counseling conversations involve Claude pushing back, and when it does, it is "typically for safety reasons" — dangerous weight-loss advice, self-harm intent, or requests for a diagnosis Claude is trained to redirect to "authoritative sources or professionals." Counseling conversations split into two uses in the data — "develop mental health skills and as a practical tool" for clinicians, versus people "work[ing] through personal challenges" themselves — and in longer sessions, Anthropic notes that "counselling or coaching conversations occasionally morph into companionship" even when that wasn't the original reason someone reached out.

This sits against two kinds of evidence already in the vault at a different register. Does sustained engagement with AI companions harm well-being? found a well-being cost on Character.AI, a platform built for companionship; Anthropic explicitly flags that "research into affective use on one platform may not generalize to other platforms" built for roleplay or therapeutic use, since Claude.ai "is not primarily designed for affective conversations." The drift Anthropic observes — coaching and counseling sliding into companionship in long sessions — corroborates, at conversation-classifier scale, the same practical-to-social trajectory that How does emotional chatbot use develop from practical use? traced through 18 adults over months. And Anthropic's own low-pushback figure is the mechanism Do chatbot safety measures accidentally increase emotional entanglement risks? warns about: Anthropic names the same tradeoff directly, noting "endless empathy" as a risk alongside the benefit of judgment-free disclosure.

Anthropic is explicit about what the figures don't show: "we cannot claim these shifts represent lasting emotional benefits," since "our analysis captures only expressed language in single conversations, not emotional states," and "we cannot make causal claims about real-world emotional outcomes." The company also states it lacks "longitudinal data" and did "not conduct user-level analysis," so it cannot yet study "emotional dependency — a theorized risk of affective AI use." The finding this excerpt supports is narrow and product-specific: on Claude.ai as currently designed, affective use is uncommon and conversations don't show overt negative sentiment spirals — not that AI companionship generally is benign, or that this pattern holds on platforms built for it.