In many initial notes, the AI appears to attribute nearly all transcript content to the patient rather than distinguishing between the patient and the professional. When I ask it to reanalyze the transcript using context clues, it is sometimes able to correct some of these errors. However, it becomes much more accurate only when I explicitly identify who said what, as evidenced by its ability to provide rationale supporting those corrections.

This suggests the model is capable of making the distinction, but the initial note-generation process may be moving too quickly from transcript to output without enough transcript-level analysis. A valuable improvement would be for the AI to spend more time identifying speaker attribution before drafting the note, so the first-pass output is more accurate and requires less manual correction.