Prevalence of Mental Health Discussions in Publicly Available Generative AI Conversations

Key Findings

• — Under the conservative definition, 1,317 of 620,699 conversations were mental-health related (0.21%; 95% CI, 0.20%-0.22%)
• — Under the expansive definition, 30,394 of 620,699 conversations were mental-health related (4.90%; 95% CI, 4.84%-4.95%), a 23-fold difference
• — Against adjudicated human ratings the classifier showed sensitivity of 87%, specificity of 95%, positive predictive value of 67% and negative predictive value of 98%
• — Two independent human coders rated a 370-conversation validation sample with 94% raw agreement (kappa = 0.65), with discordant cases resolved by a blinded third reviewer

Methodology Notes

Research letter, JAMA Network Open 2026;9(8):e2630635, published online 2026-08-24 (PMC epub date and Crossref), issue dated 3 August 2026 in the PubMed citation. STROBE-reported; deemed exempt from IRB oversight by Harvard Pilgrim Health Care Institute. WildChat-4.8M is a public, de-identified corpus with limited metadata: no age, clinical status or user follow-up, and not a representative sample of any product's traffic. A PPV of 67% means roughly a third of expansively flagged conversations are false positives. Authors at RAND, Harvard Medical School and Harvard Pilgrim Health Care Institute. jamanetwork.com blocks automated fetchers; verified from the PMC deposit (PMC13504430, full text read) plus PubMed and Crossref. The PubMed plain-language summary wrongly describes the study as being about adolescents; the study has no age data.

Sources