ChatGPT Logs Reveal Behavioral Fingerprints of Depression, But Fall Short as a Screening Tool

A dataset of 187,093 ChatGPT conversations from 766 users is offering researchers a rare, granular look at how depressive symptoms shape everyday interactions with a chatbot, without ever asking the model to diagnose anyone.

The study, submitted to arXiv on July 6, 2026 and accepted as a companion paper for CSCW 2026 (the ACM Conference on Computer-Supported Cooperative Work and Social Computing), had participants complete the PHQ-8, a standard eight-item depression screening questionnaire. Researchers then split the group at a PHQ-8 score of 10, the conventional threshold for moderate symptoms, and compared how the two cohorts actually used ChatGPT.

The differences were consistent and, in places, striking. Participants scoring at or above the threshold turned to ChatGPT more frequently for conversations touching on mental health, loneliness, interpersonal conflict, and general emotional support. Their sessions skewed heavily toward late-night hours and showed recurring patterns that persisted month over month, suggesting the chatbot had become a fixture of ongoing coping routines rather than a one-off tool. Linguistically, this group's messages contained more first-person singular pronouns ("I," "me," "my") and more absolutist language, terms like "always," "never," and "completely," a pattern long associated in psycholinguistics research with depressive rumination.

Higher-PHQ users were also more likely to disclose sensitive personal information within their chats, deepening the sense that some are using the model as a confidant during difficult moments. That raises an obvious follow-up question: does ChatGPT respond differently when someone appears to be struggling? According to the study, not in any protective way. Rates at which the model redirected users toward professional mental-health resources did not increase for the higher-PHQ group, meaning the chatbot was no more likely to point at-risk users toward real-world help despite the clearer behavioral signals of distress in their conversations.

The researchers also tested whether these linguistic and behavioral markers could be used to flag depression automatically. A language-based prediction model reached an AUROC of 0.591, a metric where 0.5 represents pure chance and 1.0 represents perfect classification. The authors were blunt in their own assessment: that performance is too weak to serve as a screening mechanism, despite the clear statistical differences observed between groups.

That distinction matters for how this research should be read. The patterns are real and measurable at a population level, late-night use, rumination-heavy phrasing, higher disclosure, but they are not yet reliable enough to identify an individual's mental state from a chat log alone. For an industry increasingly interested in AI-driven wellness features and crisis detection, the findings offer a caution: aggregate behavioral signals do not automatically translate into individual-level accuracy, and current systems show no evidence of adapting their responses to protect users who may need it most.