Universality and Hidden Information in Concept Bottleneck Models — AI Alignment Forum