After you change your mind

Imagine planning an event with an AI assistant. It produces a venue shortlist, a schedule and an invitation. Then you change one premise: children will be coming. The assistant warmly acknowledges the update and rewrites the invitation. The late finish and the unsuitable venue remain.

The answer was responsive. The shared plan was not repaired. That difference offers a useful test for whether working with AI has become collaboration: when one fact changes, does the system help work out what else must change, or does the person still have to carry the whole dependency map?

A record of the conversation would not be enough. It could faithfully preserve both “adults only” and “children welcome.” What the work needs is a current agreement about which assumption governs, which decisions depend on it and which decisions remain open. Remembering every instruction and understanding their present relationship are different achievements.

In their account of communication, Herbert Clark and Susan Brennan distinguish presenting an utterance from establishing that it has been understood well enough for the current purpose. They call that process grounding. Their proposed measure of conversational economy is the effort participants expend together, from a contribution to its mutual acceptance.¹

Applied to an assistant, this shifts attention away from how little the user has to type. A short prompt followed by twenty minutes of repairs can be an expensive interaction. A clarifying question can be cheap if it prevents work on the wrong premise. The relevant cost includes noticing that the misunderstanding happened at all.

The event example also explains why a friendly acknowledgement can be misleading. It supplies the social appearance of agreement without demonstrating that the plan changed. A better response might identify the affected decisions, revise the reversible ones and surface the unresolved choice about the venue. The user would have evidence of what the system took the correction to mean.

Initiative complicates the test. An assistant that changes the venue without asking has propagated the correction, but perhaps too far. The original venue might matter more than the finish time. Guessing at that trade-off would create another task for the user: undoing a decision the assistant was never entitled to make.

Eric Horvitz described this problem in a 1999 paper on mixed-initiative interfaces. His principles include considering uncertainty about the user's goals, the cost of interruption, easy termination and doing less when uncertainty is high. His Microsoft prototype LookOut could infer appointment details from email, show its guesses and leave them editable; if the time was unclear, it could display a relevant calendar span instead.²

The example matters because useful initiative can stop short of completing the task. Showing the right week helps without pretending to know the right hour. The paper is a design account of the author's own system, not evidence that more autonomous assistance generally improves performance.

There is empirical reason to resist that assumption. Michelle Vaccaro and colleagues synthesized 106 experiments published between January 2020 and June 2023. On the studied tasks and performance measures, human-AI combinations on average improved on people working alone but performed worse than whichever standalone partner was better.³ Help and synergy were different results.

The distinction does not make collaboration pointless. The human may supply goals or hold responsibility that the measured task does not represent. But it prevents us from treating the presence of two participants as evidence of a better system. The review also required separate human-only and AI-only results, so it excluded tasks that could only be done jointly. Its authors warn that laboratory configurations may not represent practical use.³

Neither the older interface design nor those experiments establish that maintaining a revisable plan will solve the problem. That is a proposal derived from the coordination failure. It should be tested against a simpler assistant, using changed requirements rather than only clean requests, and counting how much work the person must do to discover and repair the consequences.

Whether it asks more or less, the system should reduce the total burden of keeping the work coherent. A count of actions completed between messages would miss the cost of correcting them.

The next demonstration of an AI colleague could therefore begin halfway through a task. Change a premise. Leave the old plan visible. See whether the system knows what is now unsettled, and whether it can help settle it without quietly taking ownership of the decision.