Despite the power of generative AI, massive investments have yet to yield proportional productivity gains. The issue is not capability but coordination: humans struggle to control AI outputs and hesitate to delegate meaningful autonomy to systems they do not fully trust.

This talk presents an approach to building control and trust through shared representations -- explicit structures that capture the deep organization of a solution and are interpretable by both humans and AI. When people and AI co-develop these structures, they can iterate toward a common goal with greater alignment, transparency, and confidence. I will demonstrate how deep structure can be extracted from examples and used to scaffold human-AI systems that make interaction both more controllable and more trustworthy.