Stanford researchers identify 'mirage reasoning' in medical AI models
Stanford researchers just found a big flaw in some medical AI models: they can confidently generate detailed image descriptions and clinical findings even when they haven't seen any images at all.
This "mirage reasoning" is kind of like the AI "hallucinations" people have worried about since ChatGPT's early days, and it could be risky if left unchecked in healthcare.
Mohammad Asadi warns, Stanford proposes 'B-Clean'
The team noticed that even some advanced models were giving out strong medical opinions without real data, which could put patients at risk.
researcher Mohammad Asadi stressed how important it is to fully test these systems before using them with actual patients.
To help fix the problem, Stanford proposed a new safety check called "B-Clean," designed to make sure AIs are only trusted for tasks where they really need image input.