How to create “humble” AI | MIT News

Source Domain: news.mit.edu

Researchers led by MIT stress that AI systems can overconfidently make incorrect diagnostic decisions, posing risks for medical practice.
To tackle this issue, they propose creating AI systems that are “humble” and exhibit curiosity, revealing when they are uncertain about diagnoses or recommendations.
The researchers have introduced a framework that includes modules like the Epistemic Virtue Score to instill self-awareness within AI systems, ensuring their confidence is balanced by clinical uncertainty.
This collaborative AI model aims to work as a co-pilot with doctors, enhancing creativity and problem-solving through human-AI partnership rather than isolated AI decision-making.
The team emphasizes the importance of inclusivity in AI development to avoid biases and structural inequities, recommending that diverse perspectives are brought into AI system design processes.
They highlight the need for AI systems to be trained on comprehensive, well-conceptualized datasets that do not exclude patients due to factors like geographic location or access issues.