Inside 847 Production Clinical AI Notes
Seb gave this talk at the AI Engineer World’s Fair in San Francisco: a teardown of production clinical notes from deployed AI scribes, and why the dangerous failures are the ones that read completely fine.
In the video: a note that looks complete until you know what the patient actually said, the failure patterns across three commercial scribes and why they happen even with a perfect transcript, the reveal that a carefully engineered LLM judge passed the failing notes too, and the Discover, Capture, Calibrate loop that catches what a rubric can’t.
The full verified study is now on arXiv as two papers: One note in three, our census of three deployed scribes, and LLM Judges Verify Presence, Not Absence, on why judges miss omissions and what recovers detection. Both are accepted at NeurIPS 2026 workshops (GenAI for Health and Trust-AI-Eval, Sydney, December 2026). The benchmark behind them is open on Hugging Face, with code on GitHub.
The figures in the talk come from an earlier audit of 847 notes. The published census reports its own method and numbers: across 565 notes from three deployed scribes, every finding was adversarially verified and samples were adjudicated blind by two clinicians. One note in three carries a verified failure, and the paper shows the measured rate moving between 28% and 97% of notes depending on the review standard.