Test Finds 16 Major AI Models All Miss Dosage Errors on Slightly Blurred Pages

A test using the RxHandBD handwritten prescription dataset and FDA drug labels found that 16 leading open-source models all read 300mg correctly on clear pages, but 7 misread it as 100mg on fax-quality blurred pages, with none flagging the error.

2026-11-06 ~ 2026-11-06 · 2 related posts