How can adaptive Human-AI systems be designed to understand individual users and provide personalized assistance while maintaining human control and trust?

This question grows out of the argument in Arbour, Bojinov, Feller, and Ni (2026) that AI systems should be judged by their causal effects on people in the field, not only by laboratory or benchmark performance.

If evaluation must include how humans understand, trust, and use a system, then design cannot treat the model as an independent technical object. Adaptive assistance has to remain inspectable and controllable by the person who is supposed to benefit from it.