- AI brain fry is mental fatigue from using or overseeing AI tools beyond what your attention can hold. The term comes from a March 2026 Harvard Business Review article by Boston Consulting Group researchers, who surveyed 1,488 full-time US workers and found roughly one in seven reported it.
- Reported symptoms include mental fog, trouble concentrating, headaches and slower decisions. In the BCG data, workers doing heavy AI oversight reported 12% more mental fatigue and were 19% more likely to report information overload, and those who felt they used too much AI reported 39% more errors.
- Supervising AI is tiring because monitoring is hard work. Joel Warm, Raja Parasuraman and Gerald Matthews (Human Factors, 2008) reviewed decades of vigilance research and concluded that watching for problems is mentally demanding and stressful, not restful.
- Running many tools at once multiplies switching costs and attention residue. Sophie Leroy (Organizational Behavior and Human Decision Processes, 2009) showed that unfinished tasks keep pulling attention away from the next one, and ten parallel agents means ten unfinished tasks.
- The fix is structural: cap concurrent tools, review in batches instead of watching continuously, set a hard stop on monitoring, and leave a short ready-to-resume note when you switch away, which Leroy and Theresa Glomb (Organization Science, 2018) found reduces attention residue.
In This Article
AI brain fry is the mental fatigue that comes from using or supervising AI tools beyond what your attention can actually hold. It is not a sign that you are bad at the tools. It is what happens when your job quietly changes from doing the work to watching the work, and watching turns out to be one of the most expensive things you can ask of a human brain.
The term comes from a March 2026 Harvard Business Review article by Boston Consulting Group researchers Julie Bedard, Matthew Kropp, Megan Hsu and Gabriella Rosen Kellerman, working with Olivia Karaman and Jason Hawes of UC Riverside. Seven months later it found its plainest description in a post on X. On October 14, 2026, the founder Tibo (@tibo_maker) wrote: "AI was supposed to make us work less. I've never been this tired in my life." He described having up to ten agents running at once, one fixing a bug, one writing pages, one refactoring something he had already forgotten asking for. The post spread widely because so many people recognized the day it described.
This essay explains what the research actually found, why supervision drains you, and what a saner setup looks like.
What is AI brain fry?
The BCG team surveyed 1,488 full-time US workers and defined AI brain fry as mental fatigue from excessive use or oversight of AI tools beyond one's cognitive capacity. About 14 percent of respondents said they had experienced it. They described a mental haze, trouble concentrating, headaches and slower decisions. Some said they had to physically step away from their screens to reset.
Three findings matter most.
- Oversight is the heavy part. Workers who kept a high degree of oversight over AI reported 14 percent more mental effort, 12 percent more mental fatigue, and were 19 percent more likely to report information overload.
- More is not always more. Perceived productivity rose when people went from one AI tool to two, and rose again, less steeply, with a third. After three tools, productivity scores dipped.
- Errors climb. Workers who felt they were using too much AI reported error rates 39 percent higher than others.
There was also good news in the same data. Workers who used AI to offload repetitive, routine tasks reported about 15 percent less burnout. The authors are careful to separate brain fry from burnout, which carries physical and emotional exhaustion built over months. Brain fry is narrower and faster. It is the feeling of a mind that has been asked to evaluate more than it can process.
A caveat before going further: this is one self-report survey, published as a management article rather than a peer-reviewed paper. Treat the percentages as an early signal. The mechanisms underneath them, though, have decades of research behind them.
Why does supervising AI feel so exhausting?
For most of the twentieth century, monitoring was considered the easy job. You sat, you watched, you intervened if something went wrong. Researchers assumed the tiredness people felt came from boredom and understimulation.
That turned out to be wrong. In a 2008 review in Human Factors titled, bluntly, "Vigilance requires hard mental work and is stressful," Joel Warm, Raja Parasuraman and Gerald Matthews pulled together behavioral, neural and self-report evidence. Watching for problems produced high subjective workload, measurable drops in the brain's available resources, and rising distress. The authors noted the finding applies to most human-machine systems that require human monitoring, particularly automated ones.
The psychologist Lisanne Bainbridge saw the shape of this in 1983, in a short paper in Automatica called "Ironies of Automation." When you automate most of a job, she argued, the human is left with two things: the monitoring, which is draining, and the rare, hard exceptions the machine cannot handle, which require exactly the skill the human no longer practices. Automation does not remove the human burden. It concentrates it.
Then there is trust. Raja Parasuraman and Dietrich Manzey (Human Factors, 2010) reviewed the research on automation complacency and automation bias and found that both appear in novices and experts alike, that complacency shows up under multiple-task load, and that automation bias could not be prevented by training or instructions. In plain terms: when you are juggling several tools, you will sometimes stop checking carefully, and you cannot simply decide your way out of that. So you either check everything, which is exhausting, or you check less, which is how the errors in the BCG data happen. Neither option is restful.
The switching tax
What does ten agents running at once actually mean for your attention? It means ten unfinished tasks.
Joshua Rubinstein, David Meyer and Jeffrey Evans (Journal of Experimental Psychology: Human Perception and Performance, 2001) measured the time cost of switching between tasks and found it grew with the complexity of the rules involved. Reviewing a code change, then a marketing draft, then a data summary is not one task repeated. It is three rule sets loaded and unloaded in sequence.
Sophie Leroy (Organizational Behavior and Human Decision Processes, 2009) named the cost that lingers after the switch: attention residue. When you leave a task unfinished, part of your attention stays with it, and your performance on the next task suffers. Most agent workflows are built from unfinished tasks. Each agent is waiting for your review, your next instruction, your judgment. Each one leaves residue.
This is why brain fry often arrives as fog rather than tiredness. You are not depleted by any single task. You are fragmented across all of them. We describe the same pattern in burnout brain fog, although there it builds over months rather than hours.
Cheap generation, expensive decisions
Here is the part the productivity pitch leaves out. When generation becomes cheap, evaluation becomes the bottleneck, and evaluation is decision-making.
Every output is a small verdict. Accept, reject, revise, re-prompt, check the source, run the test. A tool that produces ten drafts an hour hands you ten decisions an hour. Decisions are where human cognitive cost concentrates, which is why we have argued that the final-say tax often costs more than the work itself.
This pattern showed up before the agent era. The Upwork Research Institute surveyed workers and executives in 2024 and found that 77 percent of employees using AI said the tools had added to their workload, with reviewing and moderating AI-generated content among the leading reasons. Meanwhile 96 percent of C-suite leaders said they expected AI to raise productivity. Both groups were looking at the same tools. One was counting output. The other was doing the reviewing.
Why faster tools raise the baseline
Why does every efficiency tool seem to end with more work, not less?
The historian Ruth Schwartz Cowan documented a version of this in More Work for Mother (1983). Washing machines, vacuums and electric stoves did not free households from housework. They raised the standards of cleanliness, shifted tasks that servants, men and children had shared onto one person, and left the total time roughly where it started. The technology was real. The relief was absorbed.
AI follows the same logic. When a report takes ten minutes instead of a day, the expectation does not become "one report, nine fewer hours." It becomes more reports, more variants, faster turnaround, and a new baseline that everyone else is also racing to meet. This is the optimization paradox in its newest form, and it sits alongside the separate fear of replacement we covered in AI anxiety and the nervous system.
Something else disappears too: the gaps. Work used to contain natural pauses. Waiting for a colleague, for a build, for the printer. Those minutes were not wasted. They were micro-recovery the body took without asking permission. Agents fill the gaps. There is always something ready to review. Recovery does not get scheduled out. It gets compressed out.
How to recover from AI brain fry
The fix is structural, not motivational. You cannot willpower your way through a vigilance load, for the same reason a pilot cannot will away fatigue on a long flight. You redesign the load.
| What you notice | What is likely happening | What helps | |---|---|---| | Fog by mid-afternoon | Sustained monitoring load | Fewer concurrent tools, scheduled review windows | | Forgetting what you asked an agent | Too many open loops | Ready-to-resume notes, fewer parallel tasks | | Approving things you did not really read | Complacency under load | Smaller batches, review when fresh | | Unable to stop checking at night | No boundary on monitoring | A hard stop time, notifications off |
Cap concurrency. The BCG productivity curve peaked around three tools. That is a hint, not a law, but it points the right way. Run fewer agents in parallel and finish more of them.
Batch the review. Continuous monitoring is the vigilance condition the research warns about. Check output at set times, two or three windows a day, instead of watching it stream.
Set a hard stop on monitoring. Agents can keep running. You do not have to keep watching. Pick a time after which you do not check status, and protect it the way you would protect a meeting.
Leave a ready-to-resume note. Sophie Leroy and Theresa Glomb (Organization Science, 2018) found that writing a quick plan for where you left a task and what you will do next reduced attention residue and improved performance on the next task. One line per agent is enough.
Delegate the routine, keep the judgment small. The BCG data suggests AI lowers burnout when it absorbs repetitive work, and raises fatigue when it multiplies the judgment calls you have to make. Aim it at the first.
Then actually rest your attention. Not a different screen. Mental rest means time when nothing is asking you to evaluate it.
A design problem, not a stamina problem
Gabriella Rosen Kellerman, one of the BCG authors, called AI brain fry "a leadership challenge, not just one for individual contributors." That is the right frame. If a tool makes the people using it foggy, error-prone and tired, the problem is the system the tool sits in, not the people.
The promise was that AI would give us time back. It still might. But only if someone decides that the time saved belongs to the human, not to the next batch of output. Until then, the most useful thing you can do is notice when you have stopped working and started watching, and close a few of the windows.
Get about one essay a week on the science of rest and the cost of optimization culture. Subscribe to the newsletter. Free, about one essay a week, no upsell.
Related reading: Attention Residue · AI Fatigue · AI Anxiety and the Nervous System · The Final-Say Tax
The research is actionable. So is the newsletter. About one essay a week, no fluff.
No spam. Unsubscribe anytime.
Frequently Asked Questions
What is AI brain fry?
AI brain fry is mental fatigue that comes from excessive use or oversight of AI tools beyond your cognitive capacity. The term was introduced by Boston Consulting Group researchers in Harvard Business Review in March 2026. People describe it as a buzzing, foggy feeling with difficulty concentrating, headaches and slower decision-making, especially after long stretches of supervising AI output.
How common is AI brain fry?
In the BCG survey of 1,488 full-time US workers, about 14% reported experiencing it. Rates varied by function, with marketing reporting some of the highest levels. Because it was a single self-report survey, the figure is best read as an early signal rather than a precise prevalence estimate.
Is AI brain fry the same as burnout?
No. The BCG authors distinguish the two. Burnout includes physical and emotional exhaustion that builds over months, while AI brain fry describes acute cognitive overload from AI use and oversight. In the same data, workers who used AI to offload repetitive, routine tasks actually reported about 15% less burnout, so how you use AI matters more than whether you use it.
Why does managing AI agents feel so exhausting?
Because the job changes from doing to monitoring, and monitoring is costly. Vigilance research shows that watching for errors is demanding and stressful. Each agent also creates an open task that competes for attention, and every output becomes a decision to accept, reject or revise. Faster generation means more decisions per hour, not fewer.
How do you recover from AI brain fry?
Reduce the number of tools or agents running at once, review output in scheduled batches instead of watching it continuously, stop monitoring at a fixed time each day, and write a one-line note on where you left each task before switching. Then give your attention genuine time off: a walk, a meal, or anything that does not involve evaluating output.