Comment

Comment by dsr_

Pre-AI voice recognition (2018), followed by 2 human reviews

https://jamanetwork.com/journals/jamanetworkopen/fullarticle...

=> the error rate was 7.4% in the version generated by speech recognition software, 0.4% after transcriptionist review, and 0.3% in the final version signed by physicians. Among the errors at each stage, 15.8%, 26.9%, and 25.9% involved clinical information, and 5.7%, 8.9%, and 6.4% were clinically significant, respectively.

AI "scribes" in a perfectly replicable best-of-all-worlds scenario (2025): https://bmjdigitalhealth.bmj.com/content/1/1/e000092

=> Omissions dominated error counts (83.8%, p<<0.001), with CAISs varying widely in error frequency and severity, and a median of 1–6 omissions per consultation (depending on CAIS). Although less frequent, hallucinations and factual inaccuracies were more often clinically serious. No tested CAIS produced error-free summaries.

On the gripping hand, people who work in the management end of the US healthcare industry can't be trusted with healthcare or information security to begin with.

parent dsr_ 6 months ago

Discussion

Live from the official Hacker News API

A modern reader for Hacker News · SvelteKit + Skeleton