A new study from MIT complicates the narrative of AI as a simple force-multiplier in medicine, finding that the benefits of LLM-based diagnostic assistance are not evenly distributed across users. The technology's promise, the research suggests, is inseparable from the expertise of the people wielding it.
The gap between novices and seasoned clinicians proved stark. Non-experts tended to defer to the model's recommendations, accepting its output even when it was incorrect. Clinicians, by contrast, were more likely to spot and flag AI errors, suggesting that domain expertise acts as a critical safeguard against the blind spots of automated reasoning.
The findings carry serious implications for healthcare deployment: medical institutions must carefully consider who uses these systems and how they are trained to use them. As AI diagnostic tools become more prevalent, understanding these human factors may prove as important as the underlying technology itself, marking a critical distinction between experts who treat AI as a checked tool and novices who may treat it as an unchallenged authority.