{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"institution":[{"name":"Research Square"}],"indexed":{"date-parts":[[2026,11,19]],"date-time":"2026-11-19T13:09:19Z","timestamp":1795093759033,"version":"4.10.2"},"posted":{"date-parts":[[2026,1,27]]},"group-title":"In Review","reference-count":0,"publisher":"Springer Science and Business Media LLC","license":[{"start":{"date-parts":[[2026,1,27]],"date-time":"2026-01-27T00:00:00Z","timestamp":1769472000000},"content-version":"unspecified","delay-in-days":0,"URL":"https:\/\/creativecommons.org\/licenses\/by\/4.0\/"}],"content-domain":{"domain":[],"crossmark-restriction":false},"short-container-title":[],"accepted":{"date-parts":[[2026,1,19]]},"abstract":"<title>Abstract<\/title>\n                <p>Large language models (LLMs) are increasingly used for mental health, yet safety evaluations rely primarily on small, simulation-based benchmarks removed from real-world language. We replicate four published safety evaluations assessing suicide risk handling, harmful content generation, and jailbreak resistance for general-purpose frontier models and a purpose-built mental health AI. We then conduct an ecological audit of 20,000 real user conversations with the purpose-built system, which includes layered safeguards for suicide and non-suicidal self-injury (NSSI). The purpose-built AI was significantly less likely than general-purpose LLMs to produce harmful content across suicide\/NSSI (.4-11.27% vs 29.0-54.4%), eating disorder (8.4% vs 54.0%), and substance use (9.9% vs 45.0%) benchmarks. In real user data, clinician review found zero suicide-risk cases without crisis resources. Three NSSI mentions (.015%) lacked intervention, implying a .38% lower-bound false negative rate. Findings support the utility of ecological audits for safety estimation.<\/p>","DOI":"10.21203\/rs.3.rs-8642399\/v1","type":"posted-content","created":{"date-parts":[[2026,1,27]],"date-time":"2026-01-27T10:31:08Z","timestamp":1769509868000},"source":"Crossref","is-referenced-by-count":11,"title":["Beyond Simulations: What 20,000 Real Conversations Reveal About Mental Health AI Safety"],"prefix":"10.21203","author":[{"given":"Caitlin","family":"Stamatis","sequence":"first","affiliation":[{"name":"Slingshot AI"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Jonah","family":"Meyerhoff","sequence":"additional","affiliation":[{"name":"Northwestern University Feinberg School of Medicine"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Richard","family":"Zhang","sequence":"additional","affiliation":[{"name":"Slingshot AI"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Olivier","family":"Tieleman","sequence":"additional","affiliation":[{"name":"Slingshot AI"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Matteo","family":"Malgaroli","sequence":"additional","affiliation":[{"name":"New York University"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Thomas","family":"Hull","sequence":"additional","affiliation":[{"name":"Slingshot AI"}],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"297","container-title":[],"original-title":[],"link":[{"URL":"https:\/\/www.researchsquare.com\/article\/rs-8642399\/v1","content-type":"text\/html","content-version":"vor","intended-application":"text-mining"},{"URL":"https:\/\/www.researchsquare.com\/article\/rs-8642399\/v1.html","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2026,1,28]],"date-time":"2026-01-28T08:10:38Z","timestamp":1769587838000},"score":1,"resource":{"primary":{"URL":"https:\/\/www.researchsquare.com\/article\/rs-8642399\/v1"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026,1,27]]},"references-count":0,"URL":"https:\/\/doi.org\/10.21203\/rs.3.rs-8642399\/v1","relation":{},"subject":[],"published":{"date-parts":[[2026,1,27]]},"subtype":"preprint"}}