Clinical AI Notes Risks: 20% of Severe Errors Slipped by Top Checkers

AI Engineer · youtube · 2026-08-23

Sebastian Fox's research reveals that while AI ambient scribes are used in about a third of US practices, they pose significant risks: roughly 5% of notes contain errors serious enough to cause harm, nearly 20% have important omissions, and over 10% contain hallucinations.

Even a sophisticated 'checker' using a frontier model and a faithfulness rubric failed to catch 20% of these severe errors.

The Core Difficulty: Verification can easily spot surface differences between transcripts and notes, but determining which differences matter is tacit, contextual, and constantly evolving. For instance, omitting a holiday detail about France might be noise, while omitting Lake Malawi could be a diagnostic clue.

The Solution: Instead of rigid specifications, use 'standard examples' discovered from real outputs, assembled dynamically per note.

Original post →

More from Apps

Apps channel →