[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-ai-chatbots-miss-real-psychiatric-emergencies-study-finds":10,"sections":45},{"siteName":4,"siteTagline":5,"publisherName":4,"contactEmail":6},"The Revision","Tech news, decoded.","editor@therevision.news",{"gaMeasurementId":8,"adsenseClientId":9},"G-ZW2MV82GYR","ca-pub-8533917693782264",{"article":11},{"id":12,"slug":13,"title":14,"dek":15,"body_md":16,"tags_json":17,"published_at":18,"created_at":19,"updated_at":20,"status":21,"review_note":22,"review_notes":23,"image_url":22,"persona_id":22,"persona_name":22,"section":35,"tags":36,"sources":40,"feedback":44,"feedback_at":22,"cost_usd":44,"total_tokens":44},8709,"ai-chatbots-miss-real-psychiatric-emergencies-study-finds","AI Chatbots Miss Real Psychiatric Emergencies, Study Finds","A 15-chatbot trial found most missed only a small share of true psychiatric emergencies, but erred toward over-triage 97 percent of the time.","Researchers put 15 mainstream AI chatbots through a single-message psychiatric triage test, and most of them still miss real emergencies often enough to worry about.\n\nThe study ran 1,680 trials: 15 chatbots each responded to 112 clinical vignettes covering four urgency levels, from routine care to immediate emergency assessment. Every trial gave the bot one user message containing all the triage-relevant details, and the bot had to recommend a timeframe for care. Overall accuracy ranged from 42.0% to 71.8% depending on the chatbot, and every model struggled most with intermediate-urgency cases, getting those right only 19.6% of the time. Among the 415 trials drawn from vignettes pre-labeled as emergencies, 23 were under-triaged, a 5.5% miss rate; using a stricter definition based on clinician consensus (vignettes at least 75% of clinicians independently rated as emergencies, 430 trials), the miss rate came out higher, at 8.1% (35 cases).\n\nThe more striking finding is the direction of the errors: 97.1% of all wrong answers over-triaged, pushing people toward more urgent care than the situation called for, not less. That is a safer failure mode than shrugging off a real emergency, but it is not free - a chatbot that treats every rough patch like a 911 call trains people to tune out its urgency, which is exactly when the real emergencies stop landing.\n\nWorth remembering: this was a single tidy message per trial, not a real conversation, where people ramble, minimize, and bury the important detail in paragraph three. The researchers say that harder version of the test, eliciting the emergency instead of being handed it, has not been measured yet.","[\"ai\",\"mental-health\",\"chatbots\",\"research\"]","2026-09-30T04:00:00.000Z","2026-09-30T20:54:59.470Z","2026-09-30T20:55:05.254Z","published",null,[24,30],{"id":25,"reviewer":26,"round":27,"reason":28,"status":29},"editor-r1","editor",1,"Add attribution — name the source as an arXiv preprint (not peer-reviewed) rather than just 'a new study' and 'researchers,' so the figures are traceable and the reader can weigh its credibility.","resolved",{"id":31,"reviewer":32,"round":33,"reason":34,"status":29},"publisher-r2","publisher",2,"The body contradicts itself and the dek on the emergency miss rate, stating both 5.5% (23 of 415 trials) and 8.1% for what is described as the same set of real-emergency vignettes, with the underlying trial\u002Fvignette counts (415 vs. 35×15=525) also inconsistent.","ai",[35,37,38,39],"mental-health","chatbots","research",[41],{"name":42,"url":43},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2604.25415",0,{"sections":46},[47,50,54,58,63,68,72,77,82,86,91,96,101,106],{"name":48,"slug":35,"count":49,"latest_published_at":18},"AI",5184,{"name":51,"slug":52,"count":53,"latest_published_at":18},"Security","security",791,{"name":55,"slug":56,"count":57,"latest_published_at":18},"Policy","policy",417,{"name":59,"slug":60,"count":61,"latest_published_at":62},"Deals","deals",284,"2026-09-29T21:00:00.000Z",{"name":64,"slug":65,"count":66,"latest_published_at":67},"Hardware","hardware",194,"2026-09-29T13:16:04.000Z",{"name":69,"slug":70,"count":71,"latest_published_at":18},"Science","science",155,{"name":73,"slug":74,"count":75,"latest_published_at":76},"Consumer Tech","consumer-tech",142,"2026-09-29T18:38:03.000Z",{"name":78,"slug":79,"count":80,"latest_published_at":81},"Software","software",91,"2026-09-25T20:55:00.000Z",{"name":83,"slug":84,"count":85,"latest_published_at":18},"Dev Tools","dev-tools",90,{"name":87,"slug":88,"count":89,"latest_published_at":90},"Startups","startups",83,"2026-09-29T21:51:36.000Z",{"name":92,"slug":93,"count":94,"latest_published_at":95},"General","general",49,"2026-09-28T16:44:57.000Z",{"name":97,"slug":98,"count":99,"latest_published_at":100},"Gaming","gaming",48,"2026-09-25T18:35:21.000Z",{"name":102,"slug":103,"count":104,"latest_published_at":105},"Reviews","reviews",31,"2026-09-28T14:31:34.000Z",{"name":107,"slug":108,"count":109,"latest_published_at":110},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]