Simple Summary
Artificial intelligence chatbots are being studied as tools that could help oncologists answer consultation requests from other hospital departments. Because expert review is slow, many studies now use another chatbot to grade these answers. We tested whether this shortcut gives the same results as an oncologist. For 108 real consultation requests concerning hospitalised patients with cancer, we compared the original oncologist’s note with answers written by two chatbots. An experienced oncologist and a chatbot grader scored all answers without being told their source. The oncologist found only small differences between the three sources, whereas the chatbot grader rated the oncologists’ notes far lower than the chatbot answers, gave higher scores to longer answers, and flagged different answers, at different rates, as potentially unsafe. The choice of grader therefore changed the conclusions of the study. Oncologists conducting or reading such studies should check whether the grader, human or chatbot, has been compared against independent expert judgement.
Abstract
Large language models (LLMs) are increasingly tested as aids for oncology consultation and, increasingly, as graders of such outputs (“LLM-as-a-judge”). We assessed whether an LLM judge and an oncologist reach comparable conclusions on real oncology documentation. For 108 consecutive inpatient oncology consultations at a tertiary centre, the oncologist’s original note and responses generated by Gemini and ChatGPT were scored by a medical oncologist and an LLM judge (Claude Opus 5), both blinded to source labels, using a nine-domain rubric (0–45). The oncologist separated the three sources by ≤1.6 points (Kendall’s W = 0.082), whereas the LLM judge scored the two models 9.8 and 13.4 points above the oncologists’ notes (W = 0.772). Agreement was poor overall (ICC(2,1) = 0.167) and within each source (0.003–0.235). Between-evaluator differences were 10.5 points for oncologist notes versus 1.1 and −1.2 points for model responses (distribution-free interaction p < 0.001). LLM-judge scores, but not oncologist scores, increased with response length, including among oncologists’ notes (ρ = 0.56). Safety flags were assigned at different rates and overlapped no more than expected by chance. In oncology consultation research, evaluator choice can determine study conclusions; no single evaluator, human or LLM, should serve as sole assessor without calibration against independent expert judgement.