A New England Journal of Medicine article highlights a critical gap in large language models: they struggle to express uncertainty appropriately, especially in medicine where clinicians are expected to acknowledge the limits of their knowledge. The piece uses the metaphor of passing Turing tests (demonstrating human-like conversation) versus passing a more demanding modern test—knowing when not to know. This matters because AI tools deployed in clinical settings that confidently assert incorrect answers pose real patient safety risks. The broader pattern documented across multiple sources shows LLMs routinely hallucinate facts and overstate confidence, a limitation that persists despite their impressive performance on many benchmarks.