“this means developing AI systems that can assess their own confidence levels and recognize when they’re operating at the edges of their capabilities.”