Large language models agreed with tumour board decisions in 71% of thyroid cancer cases, with discordance concentrated in complex ones

In a prospective single-centre study, anonymised clinical data of 59 consecutive thyroid cancer patients discussed by a multidisciplinary tumour board were submitted to two general-purpose language models. ChatGPT 5.2 matched the board's decision in 71.2% of cases (42/59, kappa 0.623) and Gemini 3.0 in 64.4% (38/59, kappa 0.527). Discordance increased in complex scenarios involving lateral neck dissection and radioiodine therapy. The authors conclude that such models may serve as supporting tools but not as substitutes for expert multidisciplinary assessment; the study is single-centre with a small sample and was not designed to assess patient outcomes.