AI chatbots struggle with colorectal cancer questions

Clinical takeaway: If you use general AI chatbots in clinical practice, treat them as drafting or information-gathering tools—not sources of definitive clinical guidance. Verify cancer-related recommendations against current guidelines and clinical judgment, particularly for diagnosis and treatment decisions.

AI chatbots are becoming increasingly common in healthcare, but their performance varies widely depending on the clinical task. A new pilot study suggests that even the latest generation of large language models remains unreliable when answering specialized questions about colorectal cancer, underscoring the need for expert oversight before these tools are used in patient care.

Researchers tested six advanced AI chatbots using 137 guideline-based multiple-choice questions adapted from the 2023 Chinese colorectal cancer guidelines. The models were evaluated under standardized conditions without prompting them to explain their reasoning.

Performance was poor across all models. The highest overall score was achieved by Kimi K2 Thinking at 27.7%, followed by Claude Opus 4.5 at 26.3%. Gemini 3 Pro Preview, DeepSeek V3.2, GPT-5.1, and Qwen3-Max each answered fewer than one in five questions correctly, with GPT-5.1 scoring 14.6%.

Accuracy also varied substantially by topic. For example, Kimi K2 correctly answered 37.0% of questions on endoscopic imaging, whereas Qwen3-Max answered 7.4% correctly, suggesting that performance may differ markedly across clinical domains.

Qualitative analysis identified recurring error patterns, including selecting answers based on keyword associations rather than clinical meaning, failures in multistep clinical reasoning, incorrect retrieval of medical facts, and fabricated information ("hallucinations"). The researchers found no relationship between question difficulty and chatbot accuracy, indicating that errors were not limited to the most challenging topics.

"Under constrained prompts, next-generation AI chatbots demonstrate unsatisfactory colorectal cancer performance, often relying on keyword matching rather than physiological simulation. This leads to dangerous clinical errors, highlighting the critical need for chain-of-thought prompting, expert oversight, and domain-specific fine-tuning before unsupervised use," the study authors concluded.

Source: Chen H, et al. (2026 Jul 9) Updates Surg. Performance of next-generation AI chatbots in colorectal cancer knowledge assessment: a comparative pilot study of ChatGPT-5.1, Gemini-3Pro Preview, DeepSeek-V3.2, Kimi K2 Thinking, Qwen3-Max and Claude Opus 4.5