Brain Fry and the Cognitive Budget of the Agentic Age – Why Better AI Makes Worse Humans!

An opinion essay by Luke Soon. AI Futurist, Ethicist and Philosopher; author of Genesis and Synthesis.

I. Prologue: The Buzzing

On New Year’s Day 2026, the programmer Steve Yegge released Gas Town, an open-source rig for orchestrating swarms of coding agents in parallel. Software assembled itself at a speed that would have been science fiction three years earlier. One early user described watching it with a palpable sense of stress, and confessed that the system was simply moving too fast for him.

I want to begin with that sentence, because I think it is the most honest thing anyone said about artificial intelligence this year. Not a benchmark. Not a valuation. A human being, standing in front of the most capable machinery our species has built, reporting that his nervous system could not keep up.

Two months later, a team at Boston Consulting Group gave the feeling a name. Surveying 1,488 American workers, they called it AI brain fry: mental fatigue from excessive use or oversight of AI tools beyond one’s cognitive capacity. Participants described a buzzing, a fog, headaches, slower decisions. One in seven had felt it. And this week, the World Economic Forum turned the finding into a call for “brain-healthy thinking” at work.

It would be easy to file this under wellness. Another listicle about breaks and breathing. That would be a serious mistake.

My argument in this essay is simple and, I hope, unsettling. Human attention has become the last bottleneck of the agentic economy. We have spent three years and well over a trillion dollars removing every constraint on machine cognition: compute, data, context length, tool use, autonomy. We have spent almost nothing on the one constraint we cannot scale, the roughly 1.4 kilograms of tissue that must still read, judge, approve and carry the liability for what the machines produce.

We are not reducing cognitive work. We are relocating it, from production to oversight. And oversight, as half a century of human factors research, a decade of neuroscience and two and a half thousand years of philosophy all warn us, is the most expensive, most fragile and most easily degraded form of human thought.

This matters far beyond individual wellbeing. Every serious governance framework for agentic AI, including the one I helped write with IMDA, places a human somewhere in the loop. Every enterprise risk register assumes that human is awake. Brain fry is what happens when the human in the loop runs out of budget. It is a governance failure that presents as a headache.

In what follows I convene an unusually wide panel. The evidence from 2025 and 2026. A 1983 paper from a cognitive psychologist that predicted all of this. Neuroscientists who have measured fatigue as a molecule. Psychologists of attention and complacency. Philosophers from Plato to Byung-Chul Han. And my usual panel of AI leaders, who, as you will see, do not agree. I will end where I always end: with the conviction that this is short-term turbulence on the way to long-term abundance, but only if we choose to design for the human, rather than around them.

II. The Evidence: What 2025 and 2026 Actually Told Us

Strip away the headlines and a remarkably consistent picture emerges from the research of the past eighteen months. It has five strands.

First, oversight is the toxin, not usage. The BCG study (Bedard, Kellerman and colleagues, Harvard Business Review, March 2026) is careful on this point. Workers whose AI use demanded high oversight, reading, checking and correcting generated output, reported 14% more mental effort, 12% more fatigue and 19% more information overload than low-oversight users. Those with brain fry showed 33% more decision fatigue, made 39% more major errors, and were markedly more likely to intend to quit (34% against 25%). Yet the same study found that AI used to replace routine, repetitive work was associated with lower burnout. The machine is not the problem. The supervisory relationship is.

Second, there is a ceiling. Self-reported productivity rose as workers went from one AI tool to two and three, then fell once they juggled four or more. Three is not a magic number, but it is suspiciously close to the limits of working memory that cognitive psychologists have documented for decades.

Third, AI intensifies work rather than lightening it. Aruna Ranganathan and Xingqi Maggie Ye of Berkeley Haas followed roughly 200 employees at a US technology firm for eight months. Nobody mandated AI. Yet people worked faster, took on broader scope and stretched work into more hours of the day, on their own initiative, because AI made doing more feel cheap and even pleasurable. The freed capacity was not banked as slack. It was spent as volume. Simon Willison, one of the most thoughtful practitioners in the field, summarised it in a line I have not been able to forget: the productivity boost these things provide is exhausting.

Fourth, reliance erodes the faculty oversight requires. In January 2026 Anthropic, a company with every commercial incentive to report otherwise, published a randomised trial of 52 developers learning a new Python library. The AI-assisted group finished barely faster but scored 50% on a follow-up comprehension quiz against 67% for those who coded by hand, with the widest gap on debugging. Developers who delegated wholesale scored below 40%; those who used AI to ask conceptual questions scored between 65% and 86%. MIT Media Lab’s EEG study, Your Brain on ChatGPT, found the weakest neural connectivity, lowest recall and lowest sense of ownership among essay writers who leaned on the model, and a residue of under-engagement when the tool was taken away. Microsoft Research and Carnegie Mellon found, across 319 knowledge workers and 936 tasks, that confidence in AI predicted less critical thinking while confidence in oneself predicted more.

Fifth, we cannot trust our own sense of how we are doing. In METR’s randomised trial, experienced open-source developers were 19% slower with AI on their own codebases, yet believed afterwards that they had been about 20% faster. Upwork found in 2024 that 96% of executives expected AI to raise productivity while 77% of employees said it had added to their workload. The gap between perceived and actual productivity is itself a symptom: a fatigued mind is a poor auditor of its own fatigue.

Put these together and the shape is unmistakable. AI shifts the human role from maker to checker. Checking is costly. The cost compounds with every additional tool or agent. The saved time is reinvested in more work, not recovery. And the very act of checking AI, done passively, erodes the skill needed to check it well.

That last loop is the one that should keep boards awake. It is not new. A woman named Lisanne Bainbridge described it with uncanny precision more than forty years ago.

III. The Ironies of Automation, Revisited

In 1983 the journal Automatica published a five-page paper by Lisanne Bainbridge, a cognitive psychologist at University College London. It was about chemical plants, power stations and aircraft. It is, I would argue, the single most important document an executive deploying agentic AI can read in 2026.

Bainbridge’s central irony is this: the more advanced the automated system, the more crucial, and the more difficult, the contribution of the human operator becomes. Designers automate what they can and leave the human what they cannot. That residue is, by definition, the hardest part: the anomalies, the edge cases, the moments when the system fails in a way nobody anticipated. We then ask the human to handle those rare, high-stakes moments while giving them nothing to do the rest of the time except watch.

And humans are very bad at watching. Vigilance research going back to Norman Mackworth’s radar operators in the Second World War shows that sustained monitoring of a reliable system degrades within half an hour. Worse, Bainbridge observed, skills that are not exercised decay. The operator who once flew the aircraft by hand, or ran the plant from first principles, gradually loses the very competence they would need on the day the automation hands control back. A monitor who has never done the job cannot tell when the job is being done badly.

Read those two paragraphs again and substitute “agent” for “automation”. The Anthropic trial’s collapse in debugging skill is Bainbridge’s skill-decay irony, measured in a randomised experiment. BCG’s oversight fatigue is her vigilance irony. METR’s developers who felt faster while being slower are her irony of the operator who can no longer judge the system’s performance.

In 2024 a team from Microsoft Research, Edinburgh and UCL led by Auste Simkute made the connection explicit in a paper titled Ironies of Generative AI, later published in the International Journal of Human-Computer Interaction. They identified four mechanisms by which generative AI loses productivity: a shift of the user’s role from production to evaluation; unhelpful restructuring of workflows; constant interruptions; and a tendency for automation to make easy tasks easier and hard tasks harder. Each has a forty-year paper trail in aviation and process control.

The most chilling evidence comes from Fabrizio Dell’Acqua at Harvard Business School. In a pre-registered field experiment, 181 professional recruiters screened CVs with AI assistance of varying quality. Those given better AI made less accurate decisions than those given worse AI. With a reliable assistant, they stopped exerting effort and deferred. With an unreliable one, they stayed alert and learnt to work with it. Dell’Acqua called this falling asleep at the wheel.

Sit with the implication. As frontier models improve, and they will, the temptation to stop checking grows, while the cost of the rare unchecked error does not shrink. Agentic AI adds a twist Bainbridge’s factories never faced: the system does not merely execute a fixed routine, it decides how to act, generating novel failure modes that no checklist anticipated. We are building systems that are reliable enough to lull us and fallible enough to hurt us.

This is what I call the Overseer’s Paradox. The better the agent, the worse the human overseer becomes, unless we deliberately design the oversight role to keep the human cognitively engaged. Human-in-the-loop is not a control by declaration. It is a control only if the human in that loop is awake, skilled and resourced. Apply my enforceability test to most enterprise AI policies today and the human checkpoint fails it: it cannot be sustained at the volumes the agents generate.

IV. The Body Keeps the Ledger: What Medicine Knows About a Tired Mind

For most of the twentieth century, mental fatigue was treated as something close to a mood: real to the sufferer, invisible to the instrument. That changed in August 2022, when Antonius Wiehler, Mathias Pessiglione and colleagues at the Paris Brain Institute published a study in Current Biology that I consider foundational for the AI era.

They asked forty volunteers to spend a working day, more than six hours, on either demanding or easy cognitive tasks, and used magnetic resonance spectroscopy to measure brain chemistry as they went. In the high-demand group, glutamate accumulated in the lateral prefrontal cortex, the seat of cognitive control, by around 8% relative to the low-demand group. No such build-up appeared in the visual cortex. And as glutamate rose, choices shifted: participants increasingly preferred small immediate rewards and low-effort options.

The authors propose a neuro-metabolic model in which glutamate accumulation makes further activation of the control network more costly. Fatigue, in other words, is the brain raising the price of the very faculty oversight depends on: inhibiting the impulse, holding several things in mind, refusing the easy answer. The prescription they drew was disarmingly simple. The only known way to clear it is rest, and above all sleep.

That is the bridge to the most important organ system in this debate: the night. Maiken Nedergaard’s work at Rochester, published in Science in 2013, showed that the brain’s glymphatic system flushes metabolic waste far more efficiently during sleep. And Hans Van Dongen’s classic 2003 study in Sleep, cited in this week’s WEF piece, found that fourteen nights of six hours’ sleep produced cognitive deficits equivalent to up to two nights of total sleep deprivation, while participants rated themselves as only slightly impaired. The Berkeley finding that AI pulls work into evenings is therefore not a lifestyle footnote. It attacks the recovery mechanism directly.

Medicine also offers the clearest natural experiments in decision fatigue. Jeffrey Linder and colleagues showed in JAMA Internal Medicine (2014) that primary-care physicians grew steadily more likely to prescribe unnecessary antibiotics as their clinic sessions wore on. The easy, defensive choice wins when the control network is spent. Every AI-supervising knowledge worker faces the same gradient: the easiest choice on an overloaded afternoon is to click approve.

And medicine is now producing the first hard evidence of AI-induced deskilling in expert practitioners. A multicentre study in The Lancet Gastroenterology & Hepatology (2025) reported that endoscopists who had grown used to AI-assisted polyp detection saw their unassisted adenoma detection rate fall, from roughly 28% to 22%, within months. Radiology studies of automation bias have shown readers of every experience level drifting towards an AI’s incorrect suggestion in mammography. These are not junior clerks. They are highly trained physicians, and their skills moved in months, not years.

The chronic end of this spectrum has a formal name. Since 2019 the World Health Organization’s ICD-11 has classified burnout as an occupational phenomenon resulting from chronic workplace stress that has not been successfully managed, marked by exhaustion, cynicism and reduced efficacy. Bruce McEwen’s concept of allostatic load, the cumulative wear of repeated stress responses on cardiovascular, metabolic and immune systems, tells us why acute strain repeated daily is not harmless. Brain fry is acute. But acute strain, repeated without recovery, is precisely how chronic disease is manufactured.

The medical lesson for leaders is stark. Cognitive fatigue is not weakness of character. It is chemistry, measurable and predictable, and it degrades the precise faculties we are now asking humans to contribute to the human-machine partnership: judgement, inhibition, scepticism. You cannot train your way out of glutamate. You can only design work that leaves room to clear it.

V. The Psychology of the Overseer

If medicine tells us that the overseeing mind tires, psychology tells us how it fails. Five ideas, each well established, explain almost everything in the brain fry data.

Effortful thought is a scarce resource. Daniel Kahneman’s distinction between fast, automatic System 1 and slow, deliberate System 2 is now so familiar that we forget its sting: System 2 is lazy by design, because it is expensive. Checking an AI’s output is pure System 2 work. Accepting it is System 1. Every agent we deploy is, in effect, a machine for generating System 2 demands. We should be honest that the once-popular “ego depletion” model of willpower as a muscle has fared poorly in large replication efforts. But the more defensible successor, Robert Kurzban’s opportunity-cost account, in which effort feels costly because the brain is constantly pricing what else it could be doing, fits the AI workplace even better. When six agents are each clamouring for review, every one of them carries the opportunity cost of the other five.

Working memory is small, and switching is not free. John Sweller’s cognitive load theory, built for education, distinguishes the load intrinsic to a task from the extraneous load imposed by poor design. Most AI oversight today is extraneous load: reconstructing what an agent did, in what order, and why, from scattered logs and chat histories. Sophie Leroy’s research on attention residue shows that when we switch from one task to another, part of our attention stays behind on the first, degrading performance on the second. Gloria Mark at UC Irvine has tracked our average focus on a single screen falling from around two and a half minutes in 2004 to under a minute today. Multi-agent workflows are attention residue at industrial scale. BCG’s three-tool ceiling is what that residue looks like in a survey.

Trust decays into complacency. Raja Parasuraman and Dietrich Manzey’s landmark 2010 review defined two related failures: automation complacency, reduced monitoring of a system believed to be reliable, and automation bias, using the system’s output as a heuristic substitute for one’s own search and judgement. Both are found in novices and experts, and neither is fixed by simply telling people to be careful. That is precisely why the WEF authors are right that policy knowledge does not change behaviour. A manager can know the rule and still accept the plausible hallucination at 6pm.

Offloading changes what we remember. Evan Risko and Sam Gilbert’s 2016 review of cognitive offloading shows that we readily externalise memory and reasoning to tools, and that this is often rational. But Betsy Sparrow’s “Google effect” studies showed we remember where information lives rather than the information itself. With generative AI we are no longer offloading facts. We are offloading the reasoning, and the MIT EEG data suggest the circuits that would have done the reasoning simply engage less. The researchers’ term, cognitive debt, is exactly right. Like technical debt, it accrues silently and is called in at the worst possible moment.

Struggle is how expertise is built. Robert Bjork’s concept of desirable difficulties holds that conditions which make learning feel harder, such as retrieval, spacing and generating answers before seeing them, produce more durable learning. The Anthropic trial’s control group encountered three times as many errors and learnt more precisely because of it. AI removes difficulty indiscriminately, the desirable along with the undesirable. Anders Ericsson’s deliberate practice, which the WEF piece invokes, depends on effortful engagement at the edge of one’s ability. You cannot deliberately practise a skill you have delegated.

The flip side is hopeful. Mihaly Csikszentmihalyi’s flow describes the state in which challenge and skill are matched and attention is fully absorbed: the antithesis of brain fry. AI can create flow, by stripping away the tedious and leaving the meaningful, or destroy it, by fragmenting attention across a swarm of half-supervised threads. Which one we get is not a property of the model. It is a property of the work design.

VI. The Philosophers’ Warning: Attention, Achievement and the Self That Exploits Itself

The data tell us what is happening. Philosophy tells us what it means, and, unusually, the philosophers got here first.

Plato and the gift that forgets. In the Phaedrus, Socrates recounts the myth of Theuth, the god who presents writing to King Thamus as a remedy for memory. Thamus refuses the gift. Writing, he says, will implant forgetfulness; its users will rely on external marks rather than their own understanding, and, in Jowett’s translation, “they will appear to be omniscient and will generally know nothing.” For two millennia we have rightly laughed at Thamus, since writing made civilisation. Yet his precise objection, that a tool can supply the appearance of wisdom without its substance, is the exact phenomenon the BetterUp and Stanford team named workslop, and the exact failure the MIT essay writers displayed when they could not quote what they had just “written”.

Byung-Chul Han and the achievement-subject. The Korean-German philosopher’s slim book The Burnout Society (2010) is, I believe, the essential text for this moment. Han argues that the disciplinary society of “you should” has given way to an achievement society of “you can”. Its exhaustion is not imposed by a master. The achievement-subject exploits itself, voluntarily and without limit, mistaking self-exploitation for freedom. Now reread the Berkeley study: nobody mandated the extra hours; workers did more because AI made more feel possible and intrinsically rewarding. That is Han’s thesis confirmed in a field experiment. Generative AI is the ultimate “you can” machine. Its infinite affordance is precisely what makes it exhausting. As the BCG authors put it in their closing line, the fact that workers can keep iterating with AI does not mean they should.

Simone Weil and attention as moral act. Weil held that attention, patient, receptive and unhurried, is the rarest and purest form of generosity, and that prayer itself consists of attention. For Weil, attention was not a productivity resource but the substance of care, of justice, of truly seeing another person. What happens to a civilisation that spends its finest attention auditing machine text? We should at least ask whether the most valuable thing humans bring to the partnership is not throughput but this: the capacity to attend, fully, to a patient, a client, a student, a problem. Brain fry is attention spent on the wrong object.

Heidegger and the standing-reserve. In The Question Concerning Technology (1954), Heidegger warned that modern technology enframes the world as Bestand, standing-reserve, resources on call to be optimised. His darkest worry was that humans themselves would come to be seen this way. When firms measure performance by tokens consumed, as BCG reports some now do, the human has become standing-reserve for the machine: a supervisory resource to be maximised, rather than a person whose judgement the machine exists to serve.

Stiegler and the pharmakon. The French philosopher Bernard Stiegler, reviving Plato’s word, insisted that every technology is a pharmakon: simultaneously poison and cure. His concept of the proletarianisation of knowledge describes how workers lose know-how when it is externalised into machines, as artisans did into the factory. Generative AI threatens a proletarianisation of thinking. The cure is not to refuse the pharmakon, but to learn its dosage.

Arendt, Aristotle and Pieper. Hannah Arendt distinguished cognition, which machines can increasingly perform, from thinking, the inner dialogue through which we judge; she located the banality of evil partly in thoughtlessness, the failure to stop and think. A rubber-stamping overseer is a small, procedural thoughtlessness, multiplied by a million approvals. Aristotle reminds us that practical wisdom, phronesis, is formed by habit and practice; you become judicious by judging. And the Catholic philosopher Josef Pieper, in Leisure, the Basis of Culture (1948), argued that genuine leisure is not idleness but the receptive, contemplative state in which insight arrives. Recovery is not the opposite of thought. It is where the best thought comes from.

There is, in fairness, a counter-tradition. Andy Clark and David Chalmers’s extended mind thesis (1998) holds that tools like notebooks can become genuine parts of our cognitive system. On this view, a well-integrated AI is not a crutch but a new cognitive organ. I find the argument compelling, with one condition its authors also stressed: the tool must be reliably available, trusted to the right degree and integrated into the agent’s own reasoning. An extended mind is a coupled system. Brain fry is what happens when the coupling is badly designed and the human end is left carrying the load.

VII. The Panel Convenes

As is my habit, I convened my panel of AI leaders. None of them commented on this essay; what follows is my reading of their public positions, set against the question at hand. Where I quote, the words are theirs and on the record. Where I paraphrase, the judgement of how their views bear on brain fry is mine. They do not agree, and the disagreement is the point.

The Practitioner: Andrej Karpathy. If anyone should be immune to AI overwhelm it is the man who coined “vibe coding”. Yet on 26 December 2025 he posted: “I’ve never felt this much behind as a programmer.” He went on to list the new layer one must now master, agents, subagents, prompts, memory, permissions, hooks, MCP and more, and to describe the need for a mental model of “fundamentally stochastic, fallible, unintelligible and changing entities”. Sixteen million people read it. Karpathy’s framing is energetic, roll up your sleeves; but read it as a cognitive scientist would and it is a near-perfect description of extraneous load. His testimony matters because it shows brain fry is not a deficiency of the unskilled. It is most acute at the frontier.

The Educator: Ethan Mollick. Mollick’s Co-Intelligence urges us to always invite AI to the table, and his 2023 “jagged frontier” experiment with BCG consultants showed large gains on tasks inside the models’ competence and worse outcomes on tasks just outside it. Mollick also popularised Dell’Acqua’s “falling asleep at the wheel”. His synthesis, as I read it, is the most useful for practitioners: the danger lies not in using AI but in not knowing which side of the frontier you are standing on. That judgement is itself System 2 work, and the first thing to go when the battery runs low.

The Builders: Dario Amodei and Sam Altman. Amodei’s essay Machines of Loving Grace imagines AI compressing a century of scientific progress into a decade; he has also warned, unusually for a chief executive, that AI could eliminate a large share of entry-level white-collar jobs within a few years. Altman’s The Gentle Singularity (2025) argues the transition will feel more gradual than we fear, because humans adapt quickly. The tension between them is instructive. If Amodei is right about speed, Altman’s assumption of human adaptability is precisely what brain fry puts under strain: adaptation is itself cognitive work, and it is being demanded of the same people who are supervising the agents.

The Economist: Erik Brynjolfsson. His 2022 essay The Turing Trap argued that we over-invest in AI that imitates and replaces humans and under-invest in AI that augments them. His “Canaries in the Coal Mine?” research with Bharat Chandar and Ruyu Chen found a 13% relative employment decline for 22- to 25-year-olds in the most AI-exposed jobs, concentrated where AI automates rather than augments; the August 2026 update puts the group 19% below trend. His explanation goes straight to our theme: older workers hold tacit knowledge, learnt through experience, that “may never be written down anywhere”. Now connect it to Bainbridge. If juniors no longer do the work through which tacit knowledge forms, who will be competent to oversee the agents in 2035? This is my Frozen Workforce made literal: a generation that never thaws into expertise.

The Scientists of Risk: Geoffrey Hinton, Yoshua Bengio and Max Tegmark. Hinton has warned that mundane intellectual labour is most exposed and suggested, only half in jest, that plumbing is a safer career. Bengio launched LawZero in 2025 to pursue non-agentic “Scientist AI”, partly from concern about systems that act autonomously in the world. Tegmark has long argued for controllable tool AI over unconstrained agents. Brain fry adds a human-factors argument to their safety case: the more autonomous and numerous the agents, the more the residual human oversight becomes a fiction, signed but not read.

The Humanists: Fei-Fei Li and Andrew Ng. Li has made human-centred AI the founding principle of Stanford HAI: technology should enhance human dignity and agency, not erode them. Ng, ever the pragmatist, has long argued that AI makes skills like coding more valuable, because those who understand the work can direct the machines far better. Together they offer the constructive answer: keep the human competent and keep the human central, and the tools become amplifiers rather than anaesthetics.

The Long View: Demis Hassabis, Mo Gawdat and Elon Musk. Hassabis speaks of an era of radical abundance, and has said that we will need new philosophers to help us understand what it means to be human in it. Gawdat, in his Diary of a CEO appearances, has predicted a painful transitional period, measured in years, before any utopia arrives. Musk has suggested that in a world of AI abundance, work will become optional. If he is right, the endgame is not brain fry at all but a crisis of meaning. Yet we must cross the turbulent middle first, and that middle is where today’s workers live.

What does the panel converge on, despite itself? That the capability curve is steep and will steepen. That the human role is changing faster than human institutions. And, implicitly, that nobody has yet designed the human side of the system with anything like the care lavished on the machine side.

VIII. Against Panic: The Case for the Optimists

I have made a dark case. Intellectual honesty requires me to argue against it as hard as I can, and the counter-case is genuinely strong.

Every tool shift has had an overload phase. Email, the BlackBerry, open-plan offices and Slack each produced a wave of exhaustion before norms, etiquette and design caught up. The “always on” anxieties of 2010 now read as a transitional period, not a permanent condition. It is entirely plausible that brain fry is the agentic equivalent: a symptom of immature interfaces and absent norms, not of AI itself.

The data are early and mostly self-reported. BCG, Upwork, BetterUp and Microsoft are surveys. METR studied sixteen developers with early-2025 tools. The MIT EEG study involved fifty-four students writing essays, and remains a preprint. The Anthropic trial measured comprehension after a thirty-five-minute tutorial, not long-term skill. Several of these sources come from firms with products to sell, including, in the WEF case, a brain-coaching company. A careful reader should hold the findings as strong signals, not settled science.

The best evidence actually points to a design solution. Look again at the strongest results and a pattern emerges that is profoundly optimistic. AI that takes over routine work reduces burnout. Developers who used AI to ask conceptual questions scored nearly as well as those who coded by hand. MIT participants who wrote first and then used AI showed the strongest neural engagement of any group. Recruiters with appropriately calibrated trust learnt to work better with their AI. The problem is never “AI”. It is a specific, identifiable pattern of use: passive delegation plus high-volume oversight plus no recovery.

Oversight itself can be automated, carefully. Much of today’s brain fry comes from humans reviewing output at the wrong level of abstraction: reading two hundred lines of code instead of the decision behind them. Verifier models, test harnesses, structured logs and agents that explain their reasoning can lift human review from the artefact to the intent. In my own framing of Prompts to Loops to Loop Governance, this is exactly the maturity curve: humans stop babysitting individual outputs and start governing the loop.

Humans are remarkably good at co-evolving with tools. The extended-mind philosophers are right that we have always thought with things: the abacus, the notebook, the map, the search engine. Each was feared; each became part of us. The Flynn effect, the long rise in measured abstract reasoning through the twentieth century, is often attributed partly to a richer cognitive environment. AI could, in principle, be the richest yet: a Socratic tutor, an infinitely patient sparring partner, a scaffold that lets more people reach expert reasoning sooner.

I accept most of this. Where I part company with the optimists is on one point: timing. Norms took a decade to form around email, and email did not supervise itself, write our code or negotiate with other agents on our behalf. The capability curve is now steeper than our institutional learning curve. Brain fry is what lives in that gap. Optimism is warranted about the destination. Complacency is not warranted about the crossing.

This is my Long-AND, not Short-OR principle. The answer is not “use less AI” OR “push adoption harder”. It is more AI on drudgery AND less concurrent oversight per human; faster machines AND deliberate human practice; ambitious automation AND a protected cognitive budget.

IX. A Cognitive Budget for the Agentic Enterprise

If human attention is the last bottleneck, it must be managed like every other scarce resource: explicitly, measurably and at board level. I propose that every organisation deploying agents at scale adopt a Cognitive Budget, a governed allocation of human oversight capacity, as a first-class element of AI governance. Here is what it contains.

1. Treat oversight capacity as a control, not a perk. In TrustOS, my seven-layer architecture for agentic governance, the human checkpoint is a control layer like any other. A control that degrades invisibly under load is not a control. Each human-in-the-loop requirement should therefore carry a stated capacity: how many reviews, of what complexity, per person per day, before quality falls. When volume exceeds capacity, the system must escalate, throttle or route, never silently pile up approvals.

2. Cap concurrency per person, not per team. BCG’s ceiling suggests three concurrent AI tools or supervised agents is a sensible default. A department can run fifty agents. An individual should not be supervising fifty agents.

3. Move oversight up the abstraction ladder. Review plans, thresholds, exceptions and outcomes, not every artefact. Require agents to state their intent and confidence, and use automated verifiers to filter the routine. Humans should spend their scarce System 2 on the decisions that genuinely need judgement.

4. Stop paying people to over-use. Retire metrics that reward token consumption, prompt counts or AI “adoption” for its own sake. Reward verified outcomes, error rates caught and quality sustained. You get the behaviour you measure, and right now many firms are measuring their way into brain fry.

5. Design for desirable difficulty. Adopt “think first, then delegate” as a working habit, following the MIT ordering effect. In learning and early-career roles, protect hands-on work deliberately, as aviation protects manual flying hours. The pilots of 2035 are the juniors of 2026; do not let Bainbridge’s skill-decay irony hollow them out.

6. Batch AI-intensive work and protect recovery. The Berkeley researchers recommend confining intensive AI work to defined blocks. Pair that with genuine recovery between blocks and firm boundaries on evening work. The glutamate data are unambiguous: the only clearance mechanism is rest.

7. Measure brain fry separately from burnout. Because brain fry is acute and cognitive, engagement and burnout surveys miss it. Add specific items to pulse surveys and track error rates and attrition among your heaviest AI users: they are your canaries.

8. Ring-fence any cognitive telemetry. The WEF authors propose monitoring fatigue signals. Done carelessly, this becomes neuro-surveillance, an intrusion that would itself raise stress. If it is used at all, it must be opt-in, purpose-limited to wellbeing and workload design, aggregated, and forbidden from any role in performance assessment.

This is also an extension of a formula I have used for years: HX = CX + EX, human experience as the sum of customer and employee experience. In the agentic age, employee experience must explicitly include a cognitive term. An enterprise that optimises agent throughput while exhausting the humans who govern it is not efficient. It is running down an asset it has not even put on the balance sheet, a form of the Ghost GDP I have written about before: value that appears in the dashboards while its human foundations quietly erode.

For governments and regulators, including our own in Singapore, the implication is that agentic AI governance frameworks should move beyond simply requiring human oversight to specifying that it must be meaningful and sustainable. A requirement that cannot be met by a rested human at realistic volumes fails the enforceability test. Better to say so honestly, and design accordingly, than to write fictions into our risk registers.

X. Coda: Short-term Turbulence, Long-term Abundance

I often describe the AI transition as a fork in the road. Down one path lies Star Trek: abundance, curiosity, technology in the service of human flourishing. Down the other lies Mad Max: scarcity, exhaustion, a scramble for whatever the machines leave behind. Brain fry is a small, early signpost at that fork, and it points both ways.

It points towards Mad Max if we keep doing what many organisations are doing now: adding agents as if human attention were free, measuring adoption instead of outcomes, letting the saved hours dissolve into more work, and leaving a generation of juniors to supervise systems they were never allowed to learn. In that world the human in the loop becomes a ceremonial signature, the risk register becomes a work of fiction, and the first great agentic failure will arrive, as Bainbridge predicted of the chemical plants, at the precise moment a tired, deskilled overseer is asked to take back control.

It points towards Star Trek if we recognise brain fry for what it is: not an argument against AI, but the clearest signal yet of where the design work remains. Every piece of evidence in this essay contains its own remedy. Routine automation reduces burnout. Engaged, inquisitive use preserves skill. Thinking first and delegating second strengthens the brain. Fewer concurrent tools raise productivity. Rest clears the chemistry. Calibrated trust beats blind trust. None of this requires a breakthrough. It requires that we take the human half of the human-machine system as seriously as we take the machine half.

There is something almost poetic in where we have arrived. For seventy years we built machines to think like us. Now that they can, the scarcest thing in the economy turns out to be the thing Simone Weil prized above all: undivided human attention. Not our labour, which machines can increasingly supply. Not our knowledge, which they can retrieve. Our attention, our judgement and the care that flows from them.

The philosophers would not be surprised. Thamus warned that a tool could counterfeit wisdom. Han warned that a society of infinite possibility would exhaust itself voluntarily. Pieper reminded us that insight is born in leisure, not in frenzy. Aristotle told us we become wise by practising wisdom. Each, in a different idiom, said the same thing: the human mind is not a resource to be maximised. It is the point of the whole enterprise.

So here is my challenge to every chief executive, every board and every policymaker reading this. Before you approve the next hundred agents, ask a simple question: whose attention will govern them, and how much of it do they have left? Put the answer in your plan, your budget and your risk register. Fund the human side of the loop. Protect the night.

The machines will keep getting smarter. That is the turbulence. Whether we stay wise enough to steer them, rested enough to notice when they drift, and skilled enough to take the wheel when it matters, that is the abundance. It is not given to us by the technology. It is ours to design.

Luke Soon is an AI Futurist, Ethicist and Philosopher with more than 25 years in artificial intelligence, co-author of the world’s first Agentic AI Model Governance Framework with IMDA Singapore, and author of Genesis: Human Experience in the Age of Artificial Intelligence and Synthesis: SuperIntelligence Protocol. He writes at GenesisHumanExperience.com.

References

AI and work, 2024 to 2026

• Bedard, J., Kropp, M., Hsu, M., Karaman, O. T., Hawes, J. and Kellerman, G. R. (2026). When Using AI Leads to “Brain Fry”. Harvard Business Review, March.
• Ranganathan, A. and Ye, X. M. (2026). AI Doesn’t Reduce Work, It Intensifies It. Harvard Business Review, February.
• Leopold, N. and Campbell, P. (2026). Why brain-healthy thinking can prevent AI burn-out. World Economic Forum, 12 October.
• World Economic Forum and McKinsey Health Institute (2026). The Human Advantage: Stronger Brains in the Age of AI.
• Anthropic (2026). How AI assistance impacts the formation of coding skills. 29 January.
• Brynjolfsson, E., Chandar, B. and Chen, R. (2025, updated 2026). Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. Stanford Digital Economy Lab.
• Kosmyna, N. et al. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. MIT Media Lab, arXiv:2506.08872.
• METR (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity.
• Lee, H.-P. et al. (2025). The Impact of Generative AI on Critical Thinking. CHI ’25, Microsoft Research and Carnegie Mellon University.
• BetterUp Labs and Stanford Social Media Lab (2025). The hidden costs of workslop.
• Upwork Research Institute (2024). AI-Enhanced Work Models.
• Simkute, A., Tankelevitch, L., Kewenig, V., Scott, A. E., Sellen, A. and Rintel, S. (2024). Ironies of Generative AI. International Journal of Human-Computer Interaction.
• Dell’Acqua, F. (2022). Falling Asleep at the Wheel: Human/AI Collaboration in a Field Experiment on HR Recruiters. Harvard Business School working paper.
• Karpathy, A. (2025). Post on X, 26 December.
• Willison, S. (2026). AI Doesn’t Reduce Work, It Intensifies It. Link blog, 9 February.
• Mollick, E. (2024). Co-Intelligence: Living and Working with AI. Portfolio.

Neuroscience and medicine

• Wiehler, A., Branzoli, F., Adanyeguh, I., Mochel, F. and Pessiglione, M. (2022). A neuro-metabolic account of why daylong cognitive work alters the control of economic decisions. Current Biology, 32(16), 3564 to 3575.
• Van Dongen, H. P. A. et al. (2003). The cumulative cost of additional wakefulness. Sleep, 26(2), 117 to 126.
• Xie, L., Kang, H., Nedergaard, M. et al. (2013). Sleep drives metabolite clearance from the adult brain. Science, 342, 373 to 377.
• Linder, J. A. et al. (2014). Time of day and the decision to prescribe antibiotics. JAMA Internal Medicine, 174(12).
• Budzyń, K. et al. (2025). Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy. The Lancet Gastroenterology & Hepatology.
• Dratsch, T. et al. (2023). Automation bias in mammography: the impact of artificial intelligence BI-RADS suggestions on reader performance. Radiology.
• McEwen, B. S. (1998). Protective and damaging effects of stress mediators. New England Journal of Medicine, 338, 171 to 179.
• World Health Organization (2019). Burn-out an “occupational phenomenon”: ICD-11 (QD85).

Psychology and human factors

• Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775 to 779.
• Parasuraman, R. and Manzey, D. (2010). Complacency and bias in human use of automation. Human Factors, 52(3).
• Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux.
• Kurzban, R. et al. (2013). An opportunity cost model of subjective effort and task performance. Behavioral and Brain Sciences, 36.
• Sweller, J. (1988). Cognitive load during problem solving. Cognitive Science, 12.
• Leroy, S. (2009). Why is it so hard to do my work? Attention residue. Organizational Behavior and Human Decision Processes, 109.
• Mark, G. (2023). Attention Span. Hanover Square Press.
• Risko, E. F. and Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9).
• Sparrow, B., Liu, J. and Wegner, D. M. (2011). Google effects on memory. Science, 333.
• Bjork, R. A. (1994). Memory and metamemory considerations in the training of human beings.
• Ericsson, K. A., Prietula, M. J. and Cokely, E. T. (2007). The making of an expert. Harvard Business Review.
• Csikszentmihalyi, M. (1990). Flow. Harper & Row.
• Hagger, M. S. et al. (2016). A multilab preregistered replication of the ego-depletion effect. Perspectives on Psychological Science, 11(4).

Philosophy

• Plato. Phaedrus, trans. B. Jowett (1892).
• Han, B.-C. (2010; English 2015). The Burnout Society. Stanford University Press.
• Weil, S. (1951). Waiting for God.
• Heidegger, M. (1954). The Question Concerning Technology.
• Stiegler, B. (2010). For a New Critique of Political Economy. Polity.
• Arendt, H. (1978). The Life of the Mind.
• Aristotle. Nicomachean Ethics, Book VI.
• Pieper, J. (1948). Leisure, the Basis of Culture.
• Clark, A. and Chalmers, D. (1998). The extended mind. Analysis, 58(1).

AI leaders cited

• Amodei, D. (2024). Machines of Loving Grace.
• Altman, S. (2025). The Gentle Singularity.
• Brynjolfsson, E. (2022). The Turing Trap. Daedalus, 151(2).