A study published in the Journal of Clinical Oncology on June 1, 2026, evaluated the performance of ChatGPT-5 as a clinical decision support tool in gynecologic oncology. Researchers analyzed 97 cancer cases, including 34 ovarian, 41 endometrial, 16 cervical, and 6 rare tumors, comparing AI recommendations against those from a multidisciplinary tumor board (MDT) at Cukurova University. Cases were processed using standardized clinical summaries, and reproducibility was tested by querying the model at three distinct time points.

Evaluation by two blinded oncologists showed mean performance scores of 89.8% to 90.1% for ChatGPT-5, compared to 93.8% to 94.2% for the MDT (p<0.001). While the MDT and AI both showed high inter-rater reliability, concordance between the two was only fair (Cohen’s kappa κ=0.267 to 0.341). ChatGPT-5 maintained full consistency across all three queries in only 38% (37/97) of cases.

Subgroup analysis indicated that the AI performed significantly better in early-stage disease (p=0.024) but struggled with complex scenarios. Specifically, the model showed inferior performance in recommending fertility-sparing approaches (p=0.045), genetic testing (p=0.019), and novel therapeutics (p=0.012). The authors concluded that the model lacks the reliability needed for independent decision-making, emphasizing that human expertise remains essential for clinical safety.

A study published in the Journal of Clinical Oncology on June 1, 2026, evaluated the performance of ChatGPT-5 as a clinical decision support tool in gynecologic oncology. Researchers analyzed 97 cancer cases, including 34 ovarian, 41 endometrial, 16 cervical, and 6 rare tumors, comparing AI recommendations against those from a multidisciplinary tumor board (MDT) at Cukurova University. Cases were processed using standardized clinical summaries, and reproducibility was tested by querying the model at three distinct time points.

Evaluation by two blinded oncologists showed mean performance scores of 89.8% to 90.1% for ChatGPT-5, compared to 93.8% to 94.2% for the MDT (p<0.001). While the MDT and AI both showed high inter-rater reliability, concordance between the two was only fair (Cohen’s kappa κ=0.267 to 0.341). ChatGPT-5 maintained full consistency across all three queries in only 38% (37/97) of cases.

Subgroup analysis indicated that the AI performed significantly better in early-stage disease (p=0.024) but struggled with complex scenarios. Specifically, the model showed inferior performance in recommending fertility-sparing approaches (p=0.045), genetic testing (p=0.019), and novel therapeutics (p=0.012). The authors concluded that the model lacks the reliability needed for independent decision-making, emphasizing that human expertise remains essential for clinical safety.