AI fatigue and cognitive debt

What is AI fatigue and cognitive debt?

AI fatigue is the tiredness and the shrug that follow a year of AI at work: a new tool every quarter, a mandate to use it, and a working day that looks exactly like it did before. People stop reading the mails about it. They open the new assistant once, decide it does not fit their work, and go back to what they were doing. The word describes the mood, not the technology.

Cognitive debt is the other half, and it moves the other way. It builds up when you hand the thinking to a tool often enough that you lose the practice of doing it yourself. The output keeps arriving. What quietly drops is your ability to produce it unaided, and above all your ability to check it. The name is borrowed from technical debt, and the borrowing is fair: both are cheap now and expensive later, and neither shows up until something needs fixing.

Three neighbouring terms in this dictionary sit close to this without being the same. Workslop is about the output: unreviewed work pushed onto a colleague who now has to sort it out. Automation bias is about trust: taking the machine's answer over your own judgement. Approval fatigue is about volume: so many prompts that clicking yes stops being a decision. AI fatigue and cognitive debt sit one level above all three, and they are what makes all three more likely.

How you recognise both on the floor

Neither one announces itself. Both show up as small things that are easy to explain away one at a time.

The signs of fatigue are behavioural. A licence count that does not match the usage report: forty seats bought, eleven people who opened the tool in the last month. And the sign that costs the most: people stop saying what does not work, because the last three times they did, the answer was a training session.

The signs of debt are about capability, and they take longer to appear. A colleague who can produce an answer to any customer question and cannot say why that answer is right. A month in which nobody on the team sent a single draft back to be redone. The practical test is simple: take a task your team does with a tool every day and ask whether the person could still do it, slowly and clumsily but correctly, without it. If the honest answer is no, and the task matters, you are carrying debt on it.

What the research shows, and what it does not

This is a field with more opinion than measurement, so be precise about who measured what.

The phrase cognitive debt comes from a study at the MIT Media Lab, put on the arXiv preprint server in June 2025 by Nataliya Kosmyna and colleagues. Fifty-four people, recruited near a handful of universities in one part of the United States, wrote essays under three conditions, with a language model, with a search engine, or with nothing, while an EEG cap recorded their brain activity. Brain connectivity scaled down with the amount of help: strongest in the group writing unaided, weakest in the group writing with the model. That last group also had trouble quoting its own work accurately. In a fourth session eighteen of them switched conditions, and the ones who had been writing with the model and now wrote unaided showed reduced connectivity, which the authors read as under-engagement.

That result went round the world in a week, and it carries less far than what has been read into it. The authors say so themselves. Their own limitations page names the narrow geography, the single model tested, the single task, and the fact that the work had not been through peer review, and they asked journalists in writing not to use words like brain rot, dumb or damage about it. In December 2025 four researchers, one of them at the University of Vienna, published a comment on arXiv with five objections: the sample size, the reproducibility of the analysis, the EEG method, inconsistencies in how the results were reported, and too little transparency. Their tone is friendly and their point is not that the finding is wrong. It is that it should be read more conservatively than it has been. Fifty-four people writing essays for a study is not evidence about your bookkeeping team.

Three other pieces of work sit closer to a real workplace.

Microsoft Research and Carnegie Mellon, CHI 2025. They surveyed 319 knowledge workers who use AI at work weekly and collected 936 real examples of tasks. The pattern: the more confidence someone had in the AI for a task, the less critical thinking they reported doing on it, and the more confidence they had in themselves, the more. They also describe the work changing shape rather than disappearing, from making the thing to verifying, combining and stewarding what the model produced. It is self-reported, which is a real limit: it covers what people say they do, not what a measurement showed.

METR, July 2025. A randomised trial with 16 experienced open-source developers on 246 real tasks in their own large repositories. With AI tools allowed, they took 19 percent longer. Asked afterwards, they estimated the tools had made them 20 percent faster. METR writes plainly that this does not show AI fails to speed up developers in general: it is sixteen people, on mature codebases they know inside out, with the tools of early 2025. What it does show is that a team's own sense of the gain is not a measurement.

The Lancet Gastroenterology and Hepatology, August 2025. The one result that is genuinely about skill in real work. Four Polish endoscopy centres introduced AI polyp detection at the end of 2021. Across 19 endoscopists who had each done more than 2,000 colonoscopies, the detection rate on procedures performed without the AI fell from 28.4 percent before the tool arrived to 22.4 percent after it. Six points in absolute terms, about a fifth in relative terms, among experienced people. The design is observational, so other things could explain part of it, and the authors write that themselves. It remains the closest thing anyone has to a measurement of a skill going down after a tool went in.

And the honest note. A good part of what gets called AI fatigue is ordinary change fatigue with a new label on it. Teams were tired of ERP rollouts, of the move to the cloud and of the third intranet in five years long before any of this. What is new is the pace and the number of tools, not the reaction. That matters practically: what helped then, fewer simultaneous changes and a reason attached to each one, helps here too.

The causes that sit with management

Little of this is about bad technology and most of it is about how the technology was brought in. Four causes come back again and again, and all four are decisions somebody made.

Adoption with no problem attached. A target of so many percent of staff using AI weekly is a usage number, and usage numbers get met. In the survey behind Gallup's June 2025 report, 22 percent of US employees said their organisation had communicated a clear plan for integrating AI and 30 percent said guidelines or a policy existed, while only 16 percent strongly agreed that the AI tools their employer provided were useful for their work. The use went up. The plan did not.

Too many overlapping tools. A chat assistant, a meeting summariser, a copilot in the office suite, something in the CRM and a coding assistant add up to five interfaces with five prompting styles and five sets of rules about which data may go in. Each one is defensible on its own. Together they cost more attention than the work they replace.

No time given to learn. The Upwork Research Institute surveyed 2,500 working people across four countries in the spring of 2024, 1,250 of them in the C-suite. Nearly all the leaders, 96 percent, expected AI to raise productivity. Among the employees using AI, 77 percent said the tools had added to their workload in at least one way, and 39 percent were spending more time reviewing or correcting AI output. Almost half, 47 percent, said they did not know how to reach the gains their employer expected. Nobody had given them the afternoon.

A metric that rewards volume. If the dashboard counts mails answered, tickets closed or documents produced, a tool that multiplies volume will look like a success whatever happened to the quality. This is where fatigue and debt meet: the people producing the volume know it is thinner than it looks, and saying so puts them behind on the number.

A support team where the numbers went up

A technical wholesaler with forty staff puts an assistant in the mailbox of its six-person support team in January. Customers ask which part fits which machine and when it can be delivered, the assistant drafts an answer from the catalogue and the order history, and the person sends it after a look.

The dashboard moves straight away. Mails handled per person per day go from about 30 to about 45, so the team goes from 180 a day to 270. Time to first response drops from four hours to under one.

What nobody counted was the second-contact rate, the share of mails where the customer writes back because the first answer did not solve it. Pulled out of the ticket system in June, it had gone from 12 percent to 21 percent. On 180 mails a day that is around 22 follow-ups. On 270 it is around 57. Of the 90 extra mails the team handles, about 35 are its own second rounds. Count only the questions resolved on the first answer and the team went from 158 a day to 213, a rise of about 35 percent rather than the 50 percent on the dashboard. The gain is real. It is a third smaller than the number.

The second thing that got worse appears in no system at all. Two people joined in February and have never answered a question without the assistant. Ask them why a particular coupling does not fit a particular pump and they cannot say, because they have never had to work it out from the catalogue. The team lead, who used to spot a wrong article number at a glance, now reads 45 drafts a day instead of 30 and finds fewer errors than she used to, which she read for months as the drafts getting better.

AI fatigue versus cognitive debt

The two get named together because they arrive together, and they behave differently on the dimension that decides what you do about them: whether the cost lands now or later.

AI fatigue is a cost you are paying today and can see today. People are disengaged this quarter, the tool sits unused this month. That makes it unpleasant and it makes it fixable. Stop three of the five rollouts, attach a real problem to the two that stay, and the mood turns inside a quarter, because nothing has to be rebuilt.

Cognitive debt is a cost you are not paying yet. Everything works while the tool is right, and the tool is right most of the time. The bill arrives on the day it is wrong and the person in front of it no longer has the practice to notice. That is the failure mode to plan for in skilled work: not the model making an error, but the reviewer who has lost the reflex to catch it. There is no quick repayment either, because the only way to get the skill back is to do the work by hand again.

They also ask for opposite responses. Fatigue is treated by asking less of people. Debt is treated by deliberately giving back a piece of what you took away: work that has to be done without help so the ability stays warm. A company that treats only the fatigue ends up with a calm team that can no longer check its own output.

What actually helps

None of this is wellbeing advice. Every measure changes a decision somebody at the top can make this month.

1. Cut the number of tools on purpose. Pick one assistant per kind of work and switch the rest off, including the ones that came free with a licence. Each extra tool costs learning time and adds a place where company data can end up. If people go around your choice, the chosen tool does not fit, and that is not a reason to add a sixth. Shadow AI describes what happens then.

2. Write down which work is done by hand. One page, specific: the year-end figures are calculated, not drafted; a complaint above a certain amount is answered by a person; a migration script is read line by line. This belongs in the same document as your AI usage policy. It works because it is decided in advance instead of in the moment.

3. Keep the ability to work without the tool where it matters. For the handful of skills your business genuinely runs on, schedule the unaided version: a morning a week for new staff answering without the assistant, a quarterly close done by hand by the person who will have to defend the figures. This is the only measure on the list that pays down debt, and on a productivity report it looks like waste, which is exactly why it needs a decision rather than good intentions.

4. Measure the outcome, not the usage. Seats used and prompts sent say nothing about whether the work got better. Pick the number that describes the result: second-contact rate, errors that surface further down the chain, hours of rework, days to close. Take a reading before the tool goes in, or you will have nothing to compare against.