Bengio explains that Scientist AI will output probabilities with confidence intervals, reflecting epistemic humility. Unlike current models with excessive confidence, this AI will indicate uncertainty when data is insufficient, a crucial trait for handling serious safety questions.
Impact: High. This feature directly addresses a critical safety concern: AI overconfidence. By explicitly signaling uncertainty, the AI becomes more trustworthy and less prone to catastrophic errors in high-stakes situations.
In the source video, this keypoint occurs from 01:38:20 to 01:39:20.
Sources in support: Rob Wiblin (Host)

