Teaching AI To Work With Us

Smart AI reads the room and saves the rewrites

Businesses have swiftly brought AI into everyday work. But the value of an AI depends on its entire exchange with a user: how well it understands an employee’s needs, how much correction it requires, and whether, together, they eventually get the job done. A polished first answer can still lead to a frustrating collaboration. A useful AI assistant must be able to adjust its approach when something goes wrong. My research asks: How can we teach AI to work more effectively with people? And how can it help without making users spend so much time explaining, correcting, and starting over?

In a project with several co-authors, we train AI assistants to anticipate the feedback a user might give. Before responding, the AI assistant drafts an answer, predicts how the user would react, and revises its answer using that prediction. Suppose an AI assistant drafts a technically correct answer that leaves the user unsure how to proceed. By anticipating that reaction, the AI can revise the answer to make the next step clearer, resembling a thoughtful colleague considering what someone will need before sharing a draft. We teach AI to practice this process by explicitly training it to predict user feedback. This is not the training AI models normally receive, which focuses on producing a good answer to a task. What we study steeps an AI in its relationship with a user.

Our experiments across mathematical reasoning, writing, and coding show that AI assistants trained this way perform better collaboratively and reduce user effort. Surprisingly, we also found his training helps assistants make better use of the advice users do provide. After an incorrect first attempt, an AI that has undergone our training is more likely to incorporate the user’s feedback and reach a correct solution, shortening the several rounds of correction and refinement in real-world workplace tasks.

Businesses are evaluating AI by what employees can accomplish with it and how much guidance it requires along the way. Choosing a value-adding AI assistant means asking how well it adapts as a task evolves, whether it recognizes when clarification is needed, and how well it works with users to correct mistakes and reach a solution. We should deliberately develop and test these abilities rather than expect them to be ready “off the shelf” as AI takes on more of what is done at the workplace.

“Collaborative Theory of Mind: Learning To Self-Revise by Anticipating User Responses.” Joint work with Fanzhi Zeng, Anubrata Das, Maytal Saar-Tsechansky. Under Review.