This paper explores active inference for user interfaces. We implement an active inference approach for 1-of-N selection, a fundamental building block of interactive systems. In this setup, users provide noisy discrete inputs and the interface sequentially identifies an intended target. This problem has an optimal solution (Horstein’s algorithm) where the channel noise is iid and known a priori, but is an open problem where the noise is unknown or varying. We reformulate the problem as free energy minimisation and derive a practical active inference implementation. Active inference with a flat noise prior performs comparably to Horstein with conservative noise assumption in the first interaction sequence and as well as Horstein with perfectly calibrated noise thereafter, demonstrating fast adaptation. We also show that active inference can infer the input polarity, offering an extra degree of freedom to users, and adapt to non-stationary noise. The application of active inference to interaction is novel, and we hope this example establishes the groundwork for the community to explore active inference in human-computer interaction. Code available at https://github.com/drsstein/iwai2024