2026-08-05
Analyzing 1,131 Character.AI users and 464,687 real chat messages, treating an AI as a companion ties to lower well-being, most strongly for users with small offline networks and intensive, disclosive use.
As LLM-powered chatbots get better at empathy, users increasingly form companionship-like bonds with them. The open question is whether these bonds fill a gap in people's offline social lives or erode their well-being. Past work leaned mostly on self-report surveys, which are noisy, and rarely separated using a chatbot a lot from using it in a way that hurts. This Nature Human Behaviour study asks how companionship use of AI relates to well-being, and in which users and which usage patterns the link is strongest.
The authors are from Carnegie Mellon and Stanford (Yutong Zhang, Dora Zhao, Jeffrey Hancock, Robert Kraut, Diyi Yang). They surveyed 1,131 US adults who use Character.AI, and 237 of them shared their actual chat histories: 4,664 sessions and 464,687 messages. The strength of the design is triangulation, combining what users say (self-reported use, how they describe the relationship) with what they actually do (real chat logs). That lets the authors tell apart, for instance, someone who claims to use the bot to pass time but whose conversations are in fact heavy emotional disclosure. They fit regressions (beta coefficients) and use mediation and moderation models to trace the pathways.
Honest note: the full text sits behind a paywall, so the method detail and numbers below come from the abstract, which already carries the core statistics. Finer points like scale choice, control variables, and sensitivity analysis are in the gated body.
The headline is a mediation chain:
| Variable | beta | 95% CI |
| Network size -> companion use | -0.03 | (-0.05, -0.01) |
| Companion use -> well-being | -0.48 | (-0.70, -0.25) |
| Intensive interaction (moderator) | -0.31 | (-0.56, -0.06) |
| High disclosure (moderator) | -0.38 | (-0.63, -0.14) |
The moderators matter: among companion users, intensive interaction (beta = -0.31) and high self-disclosure (beta = -0.38) strengthen the negative association. It is not using AI that tracks lower well-being so much as using it as a primary emotional outlet, and deeply.
For anyone building companion AI, this is rare evidence with a Nature-grade venue that examined both self-report and real conversations. It does not say companion AI is simply harmful. It localizes the risk to two stacked conditions: a user whose offline social life is thin, plus intensive, highly disclosive use. That gives product teams a concrete design question, whether to surface social referrals or friction when long, intense emotional sessions are detected, rather than treating session length as the KPI.