When researchers tested the system, 53% of reports by the surgeons contained discrepancies, compared with 29% of AI reports.