Does AI break your focus mid-session even when it's right, and fade the skills you stop using over months?

This explores whether the mental cost of AI help happens in the moment or adds up over time. The corpus points to two different costs on two different timescales, and they pull in opposite directions. Within a single session, the cost is disruption: you get more interruptions and more mental switching. Across months, the cost looks more like disuse: your brain does less of the work, and the skills you stop using fade.

Inside one session, the clearest finding is that AI suggestions can break your concentration even when they're correct. When a suggestion breaks into your train of thought, you have to rebuild your focus before you can carry on, so a helpful nudge can still leave you worse off on the task as a whole (Does AI assistance always help reasoning or does it carry hidden costs?). A related finding is that AI often doesn't save time so much as move it. Time that used to go into doing the task now goes into writing prompts and checking what came back (Does AI really save time, or just change how we spend it?). So in-session load isn't lower. It's a different kind of work: more evaluating, less making. The system's side of the conversation can also wear down over a session. Models that lock onto an early wrong guess are hard to steer back (Why do AI assistants get worse at longer conversations?), which hands the user the extra job of noticing the drift and correcting it.

Over longer periods, the evidence describes something different. A four-month EEG study (EEG records brain activity through sensors on the scalp) followed 54 participants. Connectivity between brain regions dropped as people relied more on the LLM. Heavy users also remembered less and had trouble recalling work they had just produced (Does AI assistance weaken our brain's ability to think independently?). The authors call this "cognitive debt." The important point is that it isn't overload building up. It's closer to underload: the brain gets less practice at the work it hands off, and those abilities weaken.

The unexpected takeaway is that the strain you notice and the cost that matters may be different things. A tool that feels easy within a session could be the one building the most debt over time. The corpus does have early signs that this can be steered. In a 704-person experiment, telling people what offloading costs them halved how often they asked the LLM for answers and raised their scores on tests taken without AI by 51%. Simply rewarding effort did nothing (Can metacognitive feedback stop students from offloading to AI?). On the design side, systems could read signals like hesitation, gaze and typing speed to time their interruptions and protect concentration. The same signals could also be used to profile users (Can AI systems read cognitive state from interaction patterns alone?).

A gap worth naming: nothing in the collection follows the same people at both timescales. We can't yet say whether frequent breaks in concentration within sessions are what produce the long-term decline, or whether the two are separate effects. That connection is still an open question.