Correct but Incomplete: Limitations of AI-Assisted Decision Support in Rectal Cancer

Area of Science:

• Oncology
• Medical Informatics
• Artificial Intelligence in Medicine

Background:

• Artificial intelligence (AI) is increasingly integrated into clinical decision-making processes.
• Physicians frequently utilize AI tools such as ChatGPT-5 and OpenEvidence.
• The reliability limits of these AI platforms in clinical decision support are not well-defined.

Purpose of the Study:

• To evaluate the performance of ChatGPT-5 and OpenEvidence against the National Comprehensive Cancer Network (NCCN) Guidelines for Rectal Cancer.
• To identify areas where AI tools excel and where they demonstrate limitations in supporting rectal cancer management decisions.

Main Methods:

• The NCCN Guidelines for Rectal Cancer (Version 2.2025) were reviewed.
• 138 clinical scenarios across workup/diagnosis, treatment, and surveillance were generated.
• AI platforms were queried, and responses were scored by two physicians on a 5-point Likert scale for correctness and accuracy.

Main Results:

• Both AI platforms exhibited high overall guideline concordance.
• ChatGPT-5 achieved 98.6% Correctness and 84.1% Accuracy; OpenEvidence achieved 92.8% Correctness and 81.2% Accuracy.
• ChatGPT-5 showed statistically significant higher Correctness (P=0.035), while Accuracy was comparable (P=0.634).

Conclusions:

• AI tools demonstrate reliable identification of the general direction of care in rectal cancer management.
• Limitations exist in AI's completeness, nuance, and handling of preference-sensitive decisions.
• AI-assisted decision support should augment, not replace, multidisciplinary review in rectal cancer care.