There is a pattern in working with AI that deserves more attention. We send a request, wait for a response, check the result and ask another question. A lot can happen in that time, but the work is broken into short stretches of action and anticipation.
I have noticed this in my own development work, alongside the practical benefits of using AI. It raises a wider question for anyone working this way: what happens to our concentration when waiting for the next answer becomes a regular part of the day?
The uncertainty matters too. Will the response be right? Will we understand it? How much checking will it need? That can leave us alert for the next result, carrying a low level of tension even while the tool is supposed to be taking work off our hands. It is a potential problem worth considering across workplaces, well beyond one developer’s experience.
Waiting without really switching off
Being absorbed in a piece of work means having room to hold the problem in mind and follow a thought through. That might involve code, a design, a piece of writing or a difficult decision. The work can be demanding and still feel satisfying. That is the experience of flow being discussed here.
An AI response introduces something less predictable into that process. It might be a small suggestion that fits immediately. It might take the work in a different direction, leaving us to understand and assess an approach we did not choose ourselves.
The waiting is awkward. Starting another demanding task means putting the first one down before it is resolved. Staying with the response means watching progress instead of making it. Even when we look elsewhere, some attention may remain with the answer that is on its way.
There is also the temptation to fill every gap. Another tab, another request, another small job. Each one looks harmless on its own. Together, they can leave us with several unfinished thoughts and very little time spent properly inside any of them.
This does not mean that every AI interaction breaks concentration. A well-timed suggestion could help us continue. The concern is the repeated cycle of waiting and checking, especially when several requests are running at once.
The result arrives before the understanding
The response arriving does not necessarily bring concentration back. It brings a review job. We need to understand what the tool has actually done beneath the explanation, including the assumptions or changes it has left out of its summary.
In a Laravel application, that could mean following validation, permissions and database relationships. In WordPress, it could mean checking when an asset loads or whether a change still works in Gutenberg. A page can look right while something less visible has become wrong.
The same issue applies to a generated report, a proposal or a design. A finished-looking result can arrive before we have worked through the decisions behind it. There may be facts to verify, assumptions to question and consequences to consider.
Reviewing work is normal. What can change is the amount of context we need to rebuild. When we make something ourselves, we follow its decisions as we go. With generated work, we may have to reconstruct that reasoning afterwards.
That leaves a choice: spend time examining the result properly, or accept a convincing explanation and hope it covers everything. Asking AI to explain helps us get started, but its account still needs checking against the work itself.
The uncertainty can be part of the strain. Responsibility remains with the person using the result. Delegating its production does not remove the need to understand it.
Interruptions have a cost beyond lost time
Research on interrupted work gives this concern some context. In a 2008 experiment by Gloria Mark, Daniela Gudith and Ulrich Klocke, people completed an interrupted email task faster, but reported more stress, frustration, time pressure and effort. It was a laboratory study of 48 participants, mostly students, rather than a study of AI or developers.
The important distinction is that completing work quickly and finding the work comfortable were different things. If a workplace measures only what gets finished, it could miss the effort and tension involved in keeping up.
Sophie Leroy’s 2009 research on attention residue examined another part of switching tasks. Across two experiments, attention was difficult to move away from unfinished work, and performance on the following task suffered. Applying that to waiting for AI is an interpretation, rather than something that paper tested. It offers a useful way to think about why opening another task may not provide a clean break.
There is evidence pointing in another direction too. In GitHub’s 2022 survey of more than 2,000 Copilot users, 73% reported that the tool helped them stay in flow. Those were self-reports from users of a different kind of assistance, published by the company behind the product. They are a reason to be careful with the word “never”. The task, the tool and the way it is used matter.
Finishing something and learning it are different
There is a second question here: what happens to understanding when the whole day becomes a sequence of requests? We still need to practise the thinking that allows us to judge whether a result is good.
A 2025 study from Microsoft Research and Carnegie Mellon researchers surveyed 319 knowledge workers. Greater confidence in generative AI was associated with less reported critical-thinking effort. The researchers also described a shift towards checking information, integrating responses and overseeing the task. It was a survey of people’s accounts of their work, not a measurement of their brains or proof that their abilities had deteriorated.
Checking is valuable work, but it asks something different of us from developing an idea and following it through. If producing an answer becomes almost entirely delegated, where do we leave time to practise working out an answer ourselves?
A 2026 coding study published by Anthropic looked at 52 mostly junior developers learning an unfamiliar Python library. The AI-assisted group averaged 50% on a quiz afterwards, compared with 67% for the group coding without AI. The quiz happened shortly after the task, so it cannot tell us whether those differences lasted.
The researchers also observed better comprehension among some participants who used AI to ask conceptual questions or understand generated code. Those small groups do not establish which habits caused the difference. They do suggest that using AI to learn deserves separate attention from using it to finish.
What if distraction becomes the habit?
The longer-term concern is whether a day built around short bursts of prompting and checking could make sustained work feel harder. If every sticking point leads straight to another request, we get fewer chances to stay with the uncertainty and work through it.
For someone learning a skill, that could mean fewer opportunities to practise independently. For someone experienced, it could mean a working day increasingly spent supervising outputs while carrying several unresolved tasks. Neither possibility should be treated as an inevitable outcome, but both deserve attention when we change how people work.
The research linked here does not establish that AI permanently damages attention, prevents flow or puts us into a lasting state of anxiety. Interruptions, learning, critical thinking and sustained attention are related questions, but evidence about one does not settle all the others.
We need longer studies of everyday AI-assisted work: how people concentrate, how they feel afterwards, and whether their ability to work independently changes. We also need to consider what happens when saved time becomes an expectation to manage more tasks at once.
A tool might save time on an individual job while contributing to a more fragmented day. That is a possibility organisations should examine alongside productivity. Repeated interruption or low-level stress matters even without evidence of permanent harm.
Protecting time to think
A useful starting point is to decide where AI belongs in a task before opening it. Writing down the problem, the likely cause and the expected result gives us something of our own to compare the response against.
Smaller, clearer requests can also make review more manageable. A result we can examine in one sitting is easier to follow than a large batch of work across several parts of a project. Running more requests in parallel may increase output, but it also creates more things to keep track of.
For work that needs sustained concentration, there is value in setting aside time with suggestions and chat closed. We can read, trace a problem and try an approach without immediately reaching for another answer. Getting stuck is sometimes part of finding out how something works.
Waiting can have a boundary too. Noting where a task was left and choosing one activity while it runs may be more manageable than repeatedly checking the response and starting additional requests. These are approaches worth trying, rather than proven solutions to the wider problem.
Teams have a part to play. Time to review, understand and learn needs to count as work. A faster first draft should not automatically become an expectation that someone can safely supervise twice as many tasks.
The design of the tools matters as well. Clear completion signals, control over suggestions and an understandable record of changes could make AI easier to fit around concentrated work. As with any product experience, the way an interface asks for our attention deserves as much thought as what it can produce.
AI could help us spend more time on difficult, worthwhile thinking. It could also encourage a working day in which we rarely get the chance to settle into it. How we use the tools, what employers expect and whether we protect time for understanding will help shape which experience we have.