Google DeepMind trained an AI for psychiatric intake interviews, in simulation only

A DeepMind paper fine-tuned Gemini 3.1 Flash-Lite to run psychiatric intake interviews entirely in simulation, then tested it with 32 psychiatrists playing patients. Blind raters preferred it over the base model in 76.7% of matched cases, appropriate possible diagnoses rose from 33.3% to 68.3%, and missed acute risk in tricky scenarios fell from 30.5% to 12.9%, which is still about one case in eight. The authors say it isn't ready to be used without supervision.

Why it mattersA good model for testing any AI that talks to customers: role-play the hard cases with experts before real people see it, and measure what it misses, not only what it gets right.

Curated by Sara Simeone as part of our human editorial coverage. The original sources are linked below; curation is distinct from a Hype or Fact? assessment.

SHARE THISPost on X ↗Share on LinkedIn ↗Copy link

How we knowCompany figures are labelled as their own claims; anything we couldn't verify is labelled as a report or a claim.