Divided Attention: The Real Cost of Watching AI Work

Yesterday a developer I know described a normal Tuesday. An AI tool handed back a chunk of work, and while reading it in one pane, they were trying to keep hold of the approach they had started forming in their own head. By the time they finished reading, the thread they were holding had gone quiet. Nothing broke. It just slipped. That small slip has a name in the research, and it is worth knowing.

What a tracking-and-touching study found about doing two things at once

Start with the plain idea. Divided attention is the cost of doing two things at the same time. It is different from switching, where you put one task down and pick another up. Here both tasks are live at once. The old and well-supported picture is that attention runs on one shared, limited pool, so two simultaneous tasks draw from the same tank and each one gets a little worse.

A study out earlier this year put a clean number on it. Mallory Terry and Lana Trick, writing in a peer-reviewed experimental-psychology journal, had 79 people track moving objects on a screen while also touching items that changed color. Tracking on its own was fine. Add the second live task, and the tracking dropped.

Six percent sounds small. The shape is what matters. A second task running quietly in the background taxed the first one, and the people doing it were not choosing to drop performance. That is the same shape as reading an AI tool’s output while holding your own approach in mind. The watching is a real second task, and it competes with the thinking. It is a close cousin of the mental fatigue that shows up when a day turns into mostly oversight and not delegation, and it feeds the same quiet question about whether heavy AI reliance slowly erodes the felt sense of being capable.

Here is the part that makes it easy to miss. There is a well-known finding that people know, in general, that multitasking costs something, and often overestimate the average cost. But when it comes to their own performance, the link between what they predict and what actually happens is close to zero. We are poor judges of how much divided attention costs us specifically.

The limits: a lab task, no AI, and a number from earlier this year

Be honest about what this is. The study tracked dots and touched colored squares. It was not about AI, and it was not about knowledge work. The bridge to sitting with an AI tool is mine, not the researchers’. No fresh study landed this week on this exact question, so the number here is from earlier this year, not this morning. The effect was modest, the sample was ordinary adults, and a single lab task is not a law. What it gives us is a clean, measured version of a thing most working people can feel but rarely name.

Where your own thread goes thin while the tool runs

Next time an AI tool hands work back and the reading starts while your own idea is still forming, notice that these are two tasks running at once, not one smooth handoff. Watch whether that thread gets thinner while your eyes are on the tool’s version. That thinning is the divided-attention cost, showing up in real time. None of it has to be fought. Naming it is enough to change what happens in the next minute, maybe finishing the half-formed thought on paper before reading the tool’s, then pointing the recovered attention at the work where coding-agent returns show up first. The cost is quiet and the research says we underrate it. Noticing it is the small, steady move.