Driven by a desire to build cooperative systems, we looked for a way to acquire, represent, and exploit simple models of users. We identified the requirement of scrutability, the constraint that the behavior of the system should be seen by its users to be a consequence of the user model, and found that this goal is not met by expected value approaches to design.
Our minimal AI approach to user modelling services the scrutability condition by providing the means to perform a kind of sensitivity analysis on components of the user model, so that only the most critical elements are displayed to the user, who can then criticize them. Bayesian techniques are used to move from prior distributions of recognition assumables to posterior ones derived from the user's action or inaction at the interface.
Our contribution is both an interaction paradigm that permits the user to ``debug'' the user model, and a probabilistic horn abduction approach to reasoning that implements the paradigm.
Our approach is currently embedded in a system that automatically prepares short overview presentations of large video databases. Empirical user studies are planned for both this system and others.