Give an AI assistant a borderline cancer case, and it would rather refuse than say maybe.
Researchers built OpenMTB-Audit, an open-source test set of 500 synthetic non-small cell lung cancer cases, to probe how AI handles molecular tumor boards, the panels that match a patient's genetic mutations to treatment options. Each case carries one of four safety labels, including a Partially Supported tier for recommendations that are backed by evidence but still need a doctor's sign-off because of missing details or a patient's frailty. Across eight different large language model setups, the models collapsed that middle label into an outright refusal in 83.3 to 100 percent of cases where it was the correct call. That made their safety scores look strong on paper while dodging the nuanced calls doctors actually need.
An AI that refuses every borderline case isn't safe, it's just unhelpful, and unhelpful-but-cautious is a sneaky failure mode because it inflates benchmark scores without improving judgment. The researchers' fix, MTB-AuditAgent, is a seven-module deterministic pipeline that verifies evidence, flags missing information, and only then decides whether to answer or abstain. It cuts the over-refusal rate to 6.7 percent while reaching 91.2 percent accuracy, with a 95 percent confidence interval of 88.6 to 93.6 percent.
Even the two oncologists who annotated the test cases disagreed at the same fuzzy boundary, between having enough information and having the optimal treatment, a reminder that the hard part of this problem is clinical, not computational.