AI's Bigger Risk: Confident Refusals and Silent Omissions, Not Hallucinations

While AI hallucination gets most of the attention, researchers and practitioners are flagging two more dangerous failure modes that largely go unaddressed. Linus Torvalds recently documented how Google's Gemini repeatedly declared a Linux kernel bug 'impossible to solve' during a debugging session, yet the fix turned out to be a single character change after 24 patches and 18 reboots. Physician and Composo founder Sebastian Fox analyzed 847 production clinical AI notes — all passing standard evaluations — and found 34 omitted findings, including 11 rated severity-critical. In one case, an AI note recorded a patient's headache but silently dropped jaw pain, a key red flag for giant cell arteritis, a condition that can cause permanent blindness if untreated. Unlike hallucinations, these failures — false surrender and critical omission — produce no detectable errors, making them far harder to catch with existing safeguards.