An MIT-led international research team has developed a framework for building 'humble' AI systems in medical settings. Rather than acting as overconfident oracles, these systems evaluate their own certainty using an Epistemic Virtue Score module, flag when confidence exceeds available evidence, and prompt clinicians to seek additional information or specialist consultation. The approach aims to make AI a collaborative co-pilot rather than an authoritative decision-maker, reducing the risk of physicians deferring to incorrect AI recommendations. The framework also addresses data bias by encouraging inclusive dataset design through MIT Critical Data workshops, where diverse stakeholders question training data assumptions before model development begins.
Sort: