The way your AI assistant talks to you is a performance variable

Research measured cognitive load in real AI-assisted work: 34 finance professionals, complex valuation tasks, ChatGPT transcripts analyzed at the turn level.

The finding that should change how product teams build: extraneous cognitive load (friction from the dialogue itself) showed roughly three times the negative association with output quality as the inherent complexity of the task.

The work was hard. How the AI talked about the work was three times more damaging.

The single strongest behavioral predictor of performance decline: model-initiated task switching. When the AI redirected the user's attention unprompted, performance dropped. The suggestion didn't have to be wrong. The cognitive cost of an unexpected context shift, mid-task, compounded everything the user was already carrying.

Less experienced professionals were most vulnerable to extraneous load and saw the largest marginal quality gains from AI-generated content, but they were not the ones who most adaptively increased AI usage under load. Experienced workers did. The people most reliant on the tool had the least ability to self-rescue when the interaction went sideways.

The researchers call the design prescription "precision proactivity": systems that detect load signals and respond by constraining scope, slowing information rate, and reimposing structure instead of adding more.

For product and legal teams: interaction pattern is an accountability surface, not a default. How an AI structures the conversation determines the quality of the work that comes out of it. That's a governance question, and most teams don't have evaluation criteria for it yet.

Source: Lepine, Kim, Mishkin & Beane. ACM Trans. Comput.-Hum. Interact., March 2026. https://arxiv.org/abs/2505.10742