up to 36.7%
of high-risk late-life depression questions under-triaged by a consumer chatbot
What the research found
Researchers put 90 questions about depression in later life, written from the view of older adults and family carers and split evenly by risk level, to three consumer chatbots (ChatGPT, Gemini and Doubao), and two psychiatrists rated the 270 answers blind. Overall 78.9%, 72.2% and 60.0% of answers were clinically acceptable, falling to 63.3%, 50.0% and 36.7% on high-risk questions, where 16.7% to 36.7% were under-triaged. The question set was deliberately weighted toward risk, so the figures are a stress test rather than real-world rates.
Why it matters for providers
Older adults and their carers already ask chatbots about low mood, and the questions most likely to be under-triaged are the ones where a delay matters most.
Original source
Frontiers in Psychiatry (Guilin Medical University and others, China)
Peer-reviewed or official source, the most reliable tier.