AI Should Learn to Work with You.

The position in brief

As AI systems evolve into persistent agents capable of planning, using tools, and acting on people's behalf, effective collaboration depends on more than task completion. Agents repeatedly decide when to ask, act, explain, seek confirmation, or return control. Different people may want different interaction processes even when they want the same outcome.

We argue that AI systems should learn not only what users prefer, but how to work with them. We formulate this challenge as User-Specific Interaction Learning (USIL): learning from sparse, heterogeneous, and evolving evidence to adapt interaction policies across tasks and over time.

The paper connects existing traditions in mixed-initiative interaction, adjustable autonomy, personalization, and human–AI collaboration. It motivates the costs of shared interaction thresholds under stated assumptions and proposes four open research challenges spanning evidence, representation, decisions, and evaluation.

Why it matters

An agent can complete the correct task and still interact in a way that feels intrusive or unauthorized. Learning how to collaborate responsibly requires attention to individual differences—without relaxing safety, privacy, or authorization boundaries.

Availability

The public preprint, PDF, and LaTeX source will be linked here once released. This page is a preview and does not represent a published or accepted conference paper.