The Muscle We Stopped Using: Humanity, Cognition and the Quiet Art of Outsourcing Thought

I. The White Flag

The most dangerous thing about AI in the workplace is not that it will be wrong. It is that we will stop noticing when it is.

This spring, The Economist’s Bartleby column ran a piece titled AI and the danger of cognitive surrender. Paul Blow’s illustration says more than the headline. A man in a dark suit stares, slack-jawed, at his phone. A small hand emerges from the open top of his skull, waving a white flag. The mind has not been conquered. It has volunteered.

The column draws on a Wharton paper by Steven Shaw and Gideon Nave, Thinking: Fast, Slow, and Artificial, posted to SSRN in February 2026. It is now among the most downloaded papers on the platform, with more than 54,000 downloads at the time of writing. That popularity is itself a signal. Something in the paper names a feeling that a great many professionals already had, and had no word for.

The word is surrender. It is deliberately stronger than the polite vocabulary we usually reach for: reliance, assistance, augmentation, co-pilot. Across three pre-registered experiments and roughly 10,000 individual trials, Shaw and Nave found that the mere availability of an AI was enough for people to hand over their reasoning. Nobody forced them. Nobody needed to.

I have built and deployed AI systems for more than 25 years, since the expert systems and neural networks of the late 1980s. I have watched every generation of the technology promise to free the human mind. Rule engines would free us from routine decisions. Search would free us from memorisation. Analytics would free us from guesswork. Each delivered, in part. Generative AI is different in one crucial respect. It is the first generation that can plausibly retire the human mind rather than free it, and do so with our enthusiastic consent.

That outcome is not inevitable. But it is the default, and in organisations defaults are destiny unless someone designs against them. This essay is about that design. It sets out what the evidence now says, what history has already taught us, what regulators in Brussels and Singapore have begun to require, and what leaders can build on Monday morning.

My argument, in one sentence: cognitive surrender is not a skills gap that training will close; it is a governance gap that only architecture can close.

II. Offloading Is Not Surrender

Every conversation about AI and thinking eventually reaches for two comforting precedents. Calculators did not make us innumerate. Satellite navigation made driving easier. Both are examples of cognitive offloading: a deliberate decision to hand a defined task to a tool while human reasoning stays in charge.

The precedents are less comforting than they look. In 2011, Betsy Sparrow, Jenny Liu and Daniel Wegner showed in Science that when people expect information to remain accessible, they remember less of the information itself and more about where to find it. Wegner called this transactive memory: the internet as a colleague who remembers on our behalf. Nine years later, Louisa Dahmani and Véronique Bohbot found in Scientific Reports that drivers with greater lifetime GPS use had worse spatial memory when forced to navigate unaided. In a small longitudinal follow-up, heavier GPS use over three years tracked a steeper decline in hippocampus-dependent spatial memory. Crucially, the heavy users had not turned to GPS because they had a poor sense of direction to begin with. The tool came first; the decline followed.

So offloading is not free. But it has a defining feature: the human still knows what the tool is for, still knows the destination, and can still tell when the turn-by-turn voice has sent them into a lake. The task moves; the judgement stays.

Shaw and Nave’s contribution is to name the thing that is not offloading. In their framing, cognitive surrender is adopting AI outputs with minimal scrutiny, overriding both intuition and deliberation. Gideon Nave draws the line with precision: with offloading, a person hands a specific task to a tool while human reasoning stays in charge; with surrender, the AI makes the decision and the person adopts it as their own without recognising that a transfer took place. Offloading delegates the how and keeps the what. Surrender gives away both, and keeps only the signature.

The distinction is ancient, even if the name is new. In the Phaedrus, Plato has Socrates worry that writing will give people the appearance of wisdom without its substance, because readers will rely on marks outside themselves rather than memory within. Socrates was wrong about writing, or at least the trade proved worth it, partly because a book does not answer back. A language model does. It answers fluently, instantly and with total composure, on any subject, in any register, at three in the morning. The calculator never pretended to understand the question. The model performs understanding so well that we stop checking whether ours is still in the room.

This is why Shaw and Nave reach for a structural metaphor. Building on Daniel Kahneman’s famous distinction between fast, intuitive System 1 and slow, deliberate System 2, they propose a Tri-System Theory in which AI is System 3: artificial cognition that operates outside the brain and can either supplement or supplant the other two. Critics have argued that System 3 is really System 2 moved partly outside the head. I think that debate matters less than the behaviour it predicts.

Put bluntly: System 3 is the colleague who always has an answer. The danger is not that we consult it. The danger is that we quietly promote it to System 1, so that its answer becomes our instinct before our own instinct has had a chance to speak.

III. The Asymmetry

The number that should keep every board awake is not an average. It is a spread.

In the core Wharton experiment, participants solved problems from the Cognitive Reflection Test, a set of questions designed so that the intuitive answer is wrong and the right answer requires a moment of deliberate thought. Some had access to an AI assistant; unknown to them, the assistant had been set to be right on some items and wrong on others. Relative to people with no AI at all, accuracy rose by 25 percentage points when the AI was right. When the AI was wrong, accuracy fell by 15 points below the no-AI baseline. The authors call this +25/−15 pattern the behavioural signature of cognitive surrender.

Read that twice. When it errs, the tool does not merely fail to help. It actively makes people worse than if it had never been switched on. A good AI system, Nave argues, is one that helps you when it is right. The uncomfortable corollary is that a system which is usually right can still be a net liability, if the moments it is wrong are the moments that matter. In most businesses, they are. Routine decisions are forgiving. The rare, novel, high-stakes decision is precisely where models are least reliable and where surrender is most expensive.

Three further findings turn a statistical curiosity into an organisational risk.

Confidence inflates. People who surrendered were not hesitant about it. They were more confident in their answers, including the wrong ones. In the Wharton framing, fluent, confident, frictionless output is exactly the condition under which deference rises. We are mistaking fluency for correctness, and the model is very, very fluent.

Doubt goes quiet. The researchers report that the internal cue which normally prompts deeper deliberation, the small sense that something does not add up, appears to be suppressed when surrender takes hold. Surrender does not feel like surrender. It feels like efficiency. It feels, if anything, like competence.

The vulnerable are predictable. Across the studies, participants with higher trust in AI, and lower need for cognition and fluid intelligence, surrendered more. That is an awkward finding for any organisation whose AI change programme has spent two years telling staff to trust the tools. We have been actively cultivating the single trait most associated with the failure mode.

This is what makes the phenomenon so hard to manage. A drop in productivity shows up on a dashboard by Friday. An erosion of judgement stays invisible until the day it is needed: the client meeting with no laptop open, the board question nobody rehearsed, the regulator’s follow-up, the novel crisis no model was trained on. I have written before about the Measurement Gap, the distance between what organisations can count and what actually creates value. Cognitive surrender lives entirely inside that gap. What we can count, throughput, improves. What we cannot count, judgement, decays. And because the decay is silent, it compounds.

IV. The Evidence Stack

One paper is a provocation. Seven pointing the same way is a pattern. Over the past three years, researchers in psychology, management, neuroscience, software engineering and medicine have approached the same question from different angles, with different methods, on different populations. Their findings converge.

Study	Population and method	Key finding
Shaw and Nave, Wharton (2026)	1,372 participants, three pre-registered experiments, Cognitive Reflection Test	Accuracy +25 points when AI right, −15 points below baseline when AI wrong
Dell’Acqua, Mollick et al., HBS and BCG (2023; Organization Science 2026)	758 BCG consultants, randomised field experiment	Inside the frontier, large quality gains; on a task just outside it, AI users 19 points less likely to be correct
Lee et al., Microsoft Research and CMU (CHI 2025)	319 knowledge workers, 936 real task examples, survey	Confidence in AI associated with less critical thinking; self-confidence with more
Gerlich, SBS Swiss Business School (Societies, 2025)	666 participants, survey plus interviews	Heavier AI use correlated with lower critical thinking, mediated by offloading; strongest among 17 to 25 year olds
Kosmyna et al., MIT Media Lab (2025)	54 students, EEG during essay writing	LLM users showed weakest neural connectivity and poor recall of their own essays: “cognitive debt”
Shen and Tamkin, Anthropic (2026)	52 developers learning a new library, randomised trial	AI group scored 50% on comprehension quiz versus 67% for hand-coders; biggest gap in debugging
Budzyń et al., Lancet Gastroenterology & Hepatology (2025)	19 experienced endoscopists, four Polish centres, observational	Unassisted adenoma detection fell from 28.4% to 22.4% after routine AI exposure

Three threads run through the table.

The first is the frontier problem. The BCG study gave us the phrase “jagged technological frontier”: tasks that look equally hard to a human can fall on opposite sides of what the model does well. Inside the frontier, AI lifts performance dramatically. Just outside it, the output still looks excellent, and that is precisely the trap. The consultants who got it wrong were not careless. They were persuaded by polish. The frontier is invisible from the inside, and it moves with every model release.

The second is the confidence problem. The Microsoft and Carnegie Mellon survey found a clean duality: workers confident in the AI thought less critically; workers confident in themselves thought more. Many respondents also reported they could not verify outputs in areas where they lacked expertise, which is to say exactly the areas where they most needed to. Confidence in the tool and competence in the domain are not substitutes. The first without the second is surrender with a smile.

The third, and to me the most alarming, is the residue problem. The MIT, Anthropic and Lancet studies all measured what happens after the AI is taken away. In each case, something was missing. MIT’s students who had relied on ChatGPT struggled to recall what they had just written. Anthropic’s developers were least able to debug, the exact skill needed to supervise AI-written code. And the Polish endoscopists, with an average of 28 years’ experience, detected fewer precancerous growths on their own after only a few months of routine AI use. The study’s authors admitted they had assumed experienced clinicians would be less affected. They were not.

Rigour demands caveats, and I want to be explicit about them. Gerlich’s study is correlational. The MIT study is small, with only 18 participants in its crucial final session, and its authors have warned against reading it as proof that AI damages brains. The Lancet study is observational, and some gastroenterologists have questioned whether it can isolate deskilling from other effects. The Anthropic trial measured comprehension immediately after a task, not months later, and was published by a model developer. No single study here is decisive.

But that is the point of a stack. Different designs, different weaknesses, the same direction. When a survey, a field experiment, three lab experiments, an EEG study, a randomised trial and a clinical audit all point the same way, the burden of proof shifts. The question is no longer whether sustained AI use can erode independent judgement. It is how fast, for whom, and what we intend to do about it.

V. We Have Been Here Before

None of this should surprise anyone who has spent time with the human factors literature. We have run this experiment before, in control rooms and cockpits, and we know how it ends.

In 1983, the cognitive psychologist Lisanne Bainbridge published a five-page paper in the journal Automatica titled “Ironies of Automation”. It has since become one of the most cited works in its field. Her argument was deceptively simple. The classic aim of automation is to replace human manual control, planning and problem solving. But automation is never total. Designers automate what they can and leave the human operator with whatever is left: monitoring the system, and taking over when it fails.

Herein lie the ironies. The first is that the humans who design automation are as fallible as the operators they replace, so error is relocated rather than removed. The second, and the one that matters most here, is that the more reliable the automation, the less practice the operator gets, and so the less able they are to take over in precisely the rare, abnormal situation where they are needed. Skills that are not used decay. Knowledge that is not exercised becomes inaccessible. The operator is left as a monitor of a process they no longer understand well enough to rescue. Bainbridge’s conclusion was that the more advanced a control system becomes, the more crucial, not less, the contribution of the human operator.

Read that paper today and replace “process plant” with “knowledge work” and “automatic controller” with “agent”. It reads like a briefing note written last week.

In 2010, Raja Parasuraman and Dietrich Manzey published what remains the definitive review of the phenomenon, “Complacency and Bias in Human Use of Automation”, in Human Factors. They distinguished automation complacency, the failure to monitor a system adequately, from automation bias, the tendency to follow automated advice even when it is wrong, producing both errors of omission (missing what the system missed) and errors of commission (doing what the system wrongly said). They showed both are rooted in attention, both affect experts as well as novices, and, in a finding every Chief Learning Officer should pin above their desk, that automation bias cannot be prevented by training or instructions alone.

And then there is Air France Flight 447. On 1 June 2009, an Airbus A330 flying from Rio de Janeiro to Paris entered high-altitude ice crystals over the Atlantic. The pitot probes that measure airspeed iced over, the autopilot disconnected, and the crew found themselves hand-flying an aircraft at cruise altitude, at night, in turbulence, with unreliable instruments. Within about four minutes the aircraft had stalled and fallen into the ocean. All 228 people on board died. France’s investigation agency, the BEA, found that the co-pilots had not been trained in manual handling at high altitude or in the unreliable airspeed procedure, and its final report issued 25 safety recommendations, many concerning training and manual flying.

I do not raise AF447 to be lurid. I raise it because it is the purest case of Bainbridge’s irony in the historical record. The automation worked so well, for so long, that the moment it handed control back was the moment the humans were least prepared to take it. The aircraft was salvageable for part of those final minutes. The skills that would have saved it had simply not been exercised.

Aviation learned from this. Upset prevention and recovery training became mandatory. Airlines rebuilt manual flying into recurrent training. Regulators began to treat pilot proficiency without automation as a safety-critical metric in its own right, not a nostalgic nicety. In other words, aviation stopped treating automation complacency as a personal failing and started treating it as a system property to be engineered against.

Knowledge work is about to learn the same lesson. The only open question is whether it learns from aviation, or from its own accidents.

VI. Why This Is Not a Training Problem

Bartleby’s prescriptions are sensible, and I want to give them their due. Hire people with a high “need for cognition”, who enjoy thinking for its own sake and are somewhat, though not entirely, protected against surrender. Teach people techniques that prompt them to override the model when it is wrong. Engineer AI-free periods. The Wharton team itself points in the same direction: organisational training that helps people recognise when they are surrendering and when they should push back.

I agree with all of it. I also think, on its own, it will fail. Parasuraman and Manzey told us why sixteen years ago: automation bias is not reliably cured by instruction. The deeper reason is that surrender is not a flaw in the individual. It is an equilibrium produced by the system the individual works in.

Consider what most organisations actually measure in 2026. Tickets closed. Decks produced. Lines of code merged. Proposals out of the door. Utilisation. Every one of those metrics rewards the person who accepts the model’s first answer and moves on. The person who stops, forms an independent view, checks a source and then disagrees with the machine produces fewer countable artefacts in the same week. At year-end calibration, the visible worker wins the argument. The careful one is told, kindly, to “lean into AI more”.

Then add the liability structure. The standard disclaimer says, in effect, that AI can make mistakes and so the user should check the output. What it actually means is that the AI is permitted to make mistakes and the human is accountable whether they checked or not. We have built workplaces where the machine has authority without responsibility, and the human has responsibility without the time to exercise it. That is not a recipe for vigilance. It is a recipe for the rubber stamp.

Then add the economics. In his 2022 essay The Turing Trap in Dædalus, Stanford’s Erik Brynjolfsson argued that technologists, executives and policymakers all face excess incentives to pursue AI that automates human work rather than AI that augments it, even though augmentation tends to create more value over time. Cognitive surrender is the Turing Trap operating inside a single mind. Each individual act of surrender is locally rational: it saves time today. Summed across an organisation, and across a decade, it converts augmentation into automation by stealth, with none of the governance that a deliberate automation decision would have attracted.

In that environment, telling people to “think harder” is like telling a factory worker to slow down while paying piece rates. No amount of personal discipline survives an incentive system pointed the other way. And there is a final twist. The researchers found that higher trust in AI predicts greater surrender, and the BCG experiment found AI’s polish could mask poor answers from the very supervisors meant to catch them. Most corporate AI literacy programmes are designed, quite explicitly, to raise trust and adoption. Without a counterweight, we are training people into the failure mode and calling it transformation.

This is why I describe cognitive surrender as a governance gap rather than a skills gap. Skills live in people. Governance lives in systems: in metrics, in workflows, in interfaces, in accountability, in what gets measured and what gets rewarded. If the system pays for surrender, it will get surrender, however many e-learning modules it assigns.

VII. The Regulators Have Noticed

One reason to take cognitive surrender seriously now, rather than as a philosophical worry, is that it is quietly becoming a compliance obligation.

The EU AI Act is unusual among technology statutes in naming a cognitive bias in its operative text. Article 14(4)(b) requires that high-risk AI systems be designed so that the people assigned to oversee them can remain aware of the tendency to automatically rely, or over-rely, on the system’s output, which the Act itself labels automation bias, particularly where the system provides information or recommendations for human decisions. Article 26(2) then requires deployers to give that oversight role to people with the necessary competence, training and authority. Those deployer obligations apply from August 2026.

The drafting matters. It is not enough to write a procedure saying a human will review the output. As several practitioners have pointed out, a reviewer approving hundreds of AI decisions an hour, accepting virtually all of them, is not providing meaningful oversight at all; the oversight is performative. Legal scholars have gone further. Melanie Fink at Leiden argues that the empirical evidence on automation bias means Article 14 will only succeed if it is implemented with those cognitive limits in mind, rather than treating a human in the loop as a standalone safeguard.

Singapore has been more practical still. In January 2026, at the World Economic Forum in Davos, Minister Josephine Teo launched IMDA’s Model AI Governance Framework for Agentic AI, the world’s first governance framework written specifically for AI agents. I was privileged to co-author it. Its second core dimension is meaningful human accountability, and it names the risk directly: greater autonomy brings greater automation bias, the tendency to over-trust a system that has performed reliably in the past. The framework calls for defined checkpoints requiring human approval for high-stakes or irreversible actions, and for regular audits to confirm that human oversight remains effective over time.

The May 2026 update went a step further, adding practical guidance on guarding against automation bias, including monitoring human override rates and response times. That is a small phrase with large consequences. It moves the conversation from whether a human is in the loop to whether the human in the loop is still thinking, and it makes that measurable.

Taken together, Brussels and Singapore are converging on a principle that boards should internalise before their auditors do: a human who does not exercise judgement is not a control. Oversight is not a role on an organisation chart. It is a capability that must be designed, resourced, measured and evidenced. Organisations that treat human review as a box-ticking formality are, increasingly, carrying a regulatory liability as well as a strategic one.

VIII. Designing for Dissent

If surrender is a governance gap, it needs governance controls: things you can build, measure and enforce. This is where my work on TrustOS, a seven-layer architecture for agentic governance mapped to the EU AI Act, ISO 42001, NIST’s AI RMF, MAS FEAT and IMDA’s frameworks, becomes directly relevant. Most AI governance today asks whether the model is safe, fair and robust. Very little asks whether the human is still functioning as a check. TrustOS treats the human overseer as a safety-critical component of the system, with its own failure modes, its own telemetry and its own controls.

My test for any control is the enforceability test: if you cannot show me where it runs, who owns it, what evidence it produces and what happens when it fails, it is a slogan, not a control. Applied to cognitive surrender, the test yields a short and practical set.

Control	What it measures or enforces	Warning signal	Owner
Override rate	How often humans reject or materially amend AI output, by role and decision type	Rate drifting towards zero on decisions that genuinely need judgement	Risk, with business line
Time-to-accept	Seconds between output and approval on consequential decisions	Approval faster than a human could read the output	Operations
Justification depth	Whether a reviewer can explain, unaided, why the output is right	Shorter, generic, copy-pasted rationales over time	Line managers
Seeded errors	Deliberately wrong outputs inserted into review queues, as auditors and simulator instructors already do	Planted errors passing review unflagged	Internal audit
Prediction before reveal	On high-stakes calls, the human records a view before seeing the AI’s	Human view converging on AI’s before it is shown, or never recorded	Process owners
Confidence calibration	Model shows uncertainty; human confidence tracked against outcomes	Human confidence rising while accuracy falls	Data science
Friction by tier	Higher-impact decisions require an independent human position and a second reviewer	One-click acceptance on irreversible actions	Governance office
Unassisted proficiency	Scheduled AI-free work, with performance measured	Teams unable to perform core tasks without the tool	HR and L&D

Three design principles sit underneath the table.

1. Friction is a feature, not a bug. Product teams have spent twenty years removing friction from interfaces. For consequential decisions, a measured amount of friction is the point. Requiring a reviewer to type a one-line reason before accepting a recommendation, or to state their own estimate before the model reveals its own, costs seconds and buys judgement. The Microsoft and CMU researchers reached the same conclusion from the design side: AI tools should be built to provoke reflection, not merely to deliver answers.
2. Measure the human, not just the model. We already run evaluations on models before deployment and monitor them afterwards. We almost never run equivalent evaluations on the humans who supervise them. An override rate of zero is not evidence that the AI is perfect. It is evidence that nobody is looking. Equally, a sudden spike in overrides may mean the model has drifted. Either way, the signal is only visible if you collect it.
3. Tier by consequence, not by technology. Not every AI-assisted decision deserves friction. Drafting an internal email does not need a second reviewer. Approving a loan, a diagnosis, a sanction or a payment does. TrustOS tiers controls by the reversibility and impact of the decision, mirroring IMDA’s emphasis on human checkpoints for high-stakes or irreversible actions.

None of this is exotic. Aviation has done a version of it for decades. Autopilot flies most of the journey, yet pilots must hand-fly regularly, are drilled on failures in simulators, and are assessed on unassisted competence. Banks already plant test transactions to check that controls fire. Auditors already sample work to confirm that reviewers reviewed. The only novelty is pointing these well-understood techniques at a new kind of reviewer: the knowledge worker supervising an AI.

This is also the next step in a progression I have described elsewhere: Prompts → Loops → Loop Governance. In the prompt era, the human wrote every instruction and read every answer. In the loop era, agents plan, act, check and retry across long chains of work, and the human steps back to set goals and approve outcomes. In the loop-governance era, the human’s job is to govern the loops themselves. A governor who has surrendered is not a governor. It is a rubber stamp with a salary, and a regulator will eventually ask to see the evidence that it was ever anything more.

There is a mirror-image problem, and it deserves its own essay. Everything above concerns the human half of oversight: keeping the reviewer thinking. The other half is keeping the agent honest. Researchers such as Princeton’s Elad Hazan and MIT’s Stephen Morris are now applying mechanism design, which MIT Technology Review has called “reverse game theory”, to AI agents: building incentives so that following the rules becomes the agent’s rational choice, with solvers checked by auditors and disputes escalated to a human expert. It is promising work. But notice where those designs end: with a human who settles the appeal. Incentive-compatible agents do not remove the need for an unsurrendered human; they concentrate it at the point of judgement. The two problems compound rather than substitute. I take up the agent’s side in the companion essay, Governing the Governed.

IX. The Apprentice Problem

For experienced professionals, surrender is a risk to judgement they already possess. For the young, it is a risk to judgement they have not yet formed. The second is far more serious, because it is harder to see and slower to reverse.

Expertise is built from scaffolding: years of doing the dull, foundational work badly, then less badly, until pattern recognition becomes instinct. The junior analyst who builds a model by hand learns where models break. The junior lawyer who drafts the clause learns why the clause exists. The junior engineer who spends an afternoon chasing a bug learns how systems actually fail. Once that scaffolding is in place, AI is an extraordinary accelerant. Before it is in place, AI can prevent it from forming at all.

Anthropic’s randomised trial is the cleanest evidence I have seen. Developers learning an unfamiliar Python library with an AI assistant scored 50% on a follow-up comprehension quiz; those who coded by hand scored 67%. The AI group was not even significantly faster. The largest gap was in debugging, which the researchers flagged as the very skill needed to notice when AI-generated code is wrong. Their conclusion was blunt: AI-enhanced productivity is not a shortcut to competence.

But the same study contains the hopeful finding, and it is the one I would put in front of every graduate programme director. How people used the assistant mattered more than whether they used it. Participants who delegated wholesale, or who repeatedly pasted errors back into the chat, averaged below 40%. Those who used the AI to ask conceptual questions, request explanations and interrogate the code they were given scored between 65% and 86%. Same tool, opposite outcomes. The difference was posture: offloading versus surrender, inquiry versus delegation.

Gerlich’s survey points the same way from a different angle: the heaviest AI dependence and the lowest critical thinking scores were among 17 to 25 year olds, while higher educational attainment was associated with stronger critical thinking regardless of AI use. Foundations protect. Their absence exposes.

This is the shadow side of what I have called the Frozen Workforce. Many organisations are already slowing graduate hiring because AI can do much entry-level work. If they also allow the juniors they do hire to surrender rather than learn, they are not saving money. They are borrowing it from the future at a ruinous rate of interest. In ten years they will have a generation of senior titles without senior judgement behind them, supervising systems they cannot audit, advised by a model that will never tell them so.

The Wharton researchers single out healthcare, law, education and management consulting as the fields most exposed, because accountability and critical thinking are the product. I spent much of my career in professional services, and I would put it more starkly. In those professions, judgement is the only thing the client is really paying for. The hours, the decks and the memos are packaging. Surrender the judgement and you have commoditised yourself, at exactly the moment the client discovers they can buy the packaging directly.

The practical response is not to ban AI from the apprenticeship. It is to redesign the apprenticeship around it. That means AI tools configured as tutors rather than ghostwriters for those still learning, assessment of the reasoning as well as the result, protected time for unassisted work, and senior people who explain why rather than simply approving what. Some educators are already doing this: Harvard’s CS50 permits its own tutoring assistant, designed to guide rather than solve, and restricts tools that simply complete the work. Corporate academies should be at least as thoughtful as a first-year computer science course.

X. In Defence of Surrender

An essay that only marshals evidence for its own side is advocacy, not thought leadership. So let me make the strongest case against my argument, as its best defenders would.

In 1911, the philosopher Alfred North Whitehead wrote that “civilization advances by extending the number of important operations which we can perform without thinking about them.” Nobody today checks the arithmetic of their spreadsheet, recalculates their mortgage by hand, or verifies the physics of the lift before stepping in. We surrender to systems constantly, and we are right to. Trust in reliable systems is not a pathology. It is the foundation of every complex society. A world in which every professional re-derived every answer from first principles would be a world of very slow, very tired professionals.

Shaw and Nave themselves are careful on this point. They note that surrender is not always harmful, and their deeper question is when it is adaptive. If a model is more accurate than a person on a given class of decision, and reliably so, then deferring to it improves outcomes. In the Wharton data, accuracy rose by 25 points when the AI was right. For a well-characterised task where the model is right 99% of the time, surrender might be the rational policy, and insisting on independent human review might simply add cost and error.

There is a second argument, about distribution. The early field studies of generative AI at work, including Brynjolfsson’s own work with customer service agents, found that the least experienced workers gained the most. Surrender, on this view, is a great leveller: it lets a novice perform like an expert on day one. Insisting that everyone first earn their judgement the slow way could entrench incumbents and slow social mobility.

A third argument is historical. Every cognitive technology has provoked this panic. Socrates on writing. Moralists on the printing press. Teachers on calculators. Commentators on Google making us stupid. Each time, we lost a skill we no longer needed and gained capacities we could not previously imagine. Perhaps the ability to unaidedly solve a Cognitive Reflection Test item is the twenty-first-century equivalent of mental long division: impressive, and increasingly beside the point.

I take each of these seriously, and I think each contains a truth. But each also contains an assumption that does not survive contact with the evidence.

Whitehead’s operations become safe to perform without thinking once they are stable, well-characterised and independently verified: arithmetic, lift engineering, mortgage maths. Generative AI fails all three conditions. Its frontier is jagged and moving; its errors are fluent rather than obvious; and its outputs are, by design, different every time. The rational case for surrender depends on knowing which side of the frontier you are on. The BCG study showed that highly capable people, given a task deliberately chosen to sit just outside the frontier, could not reliably tell.

The levelling argument is real in the short term and dangerous in the long term. A novice who performs like an expert on day one, but never becomes one, is levelled up only as long as the tool is right and present. The Anthropic and Lancet data suggest the levelling can quietly become a ceiling.

And the historical argument is precisely why this time deserves a governance response rather than a moral panic. Writing, printing and calculators were offloading technologies: they externalised memory or computation while leaving judgement in human hands. Generative AI is the first widely deployed technology whose core function is to produce judgement-shaped outputs. Losing long division was cheap. Losing the capacity to tell a good argument from a bad one is not.

So my position is not that surrender is always wrong. It is that surrender must be chosen, bounded and governed, never drifted into. That is the difference between a pilot engaging the autopilot and a pilot falling asleep.

XI. My Usual Panel, Expanded

How the argument fares against the positions these voices have taken publicly. Their views are characterised from their published work and statements, not quoted, except where linked. Where a voice is no longer with us, the panel speaks from the work.

The builders

Dario Amodei, chief executive of Anthropic, would broadly agree, with a caveat. He has said that using AI the wrong way could make people “become stupider”, but that this is a choice: even when a model is better than you at something, you can still learn it and enrich yourself by doing so. His challenge to this essay would be that it underplays agency. Surrender is a decision, not a sentence. He would also point, fairly, to his own company publishing the coding study that cuts against its commercial interest, and to learning modes that deliberately make the user write the code.

Demis Hassabis would reframe the question around science. AlphaFold did not make structural biologists stop thinking; it gave them a starting point and moved their judgement upstream, to which proteins matter and which predictions to trust. His likely test: does the tool expand the frontier of what humans can investigate, or merely replace the investigation? That is a sharper version of Brynjolfsson’s augmentation question, applied to the lab.

Andrej Karpathy, who coined the phrase “vibe coding” in 2025, is a useful provocateur here. He described it as giving in to the vibes and forgetting the code exists, and was explicit that it suits throwaway weekend projects. In effect, Karpathy has already drawn the line this essay argues for: surrender is acceptable where the stakes are low and the output disposable. The problem is that organisations have taken a technique labelled for weekends and deployed it on Monday.

Mustafa Suleyman has warned against building AI that appears conscious, arguing that the illusion of a person behind the screen will distort how people relate to these systems. That bears directly on surrender. We defer more readily to something that sounds like a confident colleague than to something that looks like a tool. Design choices about persona, voice and certainty are, in his framing, choices about how much people will defer.

The scientists

Yann LeCun would supply the technical sting. He has argued for years that today’s language models produce fluent text without a grounded model of the physical world, and lack the persistent memory and planning that real understanding requires. On this view, the confidence a model projects is not evidence of comprehension, and surrendering to fluency is a category error: mistaking the map’s handwriting for the territory.

Geoffrey Hinton would take it somewhere darker. He has warned that advanced AI could become better than us at persuasion and manipulation. If people already defer this readily to systems that are merely helpful, what happens when those systems are deployed by actors who want our deference? Cognitive surrender at work, he might suggest, is a rehearsal for something much larger in politics and society.

Fei-Fei Li would reframe the whole debate around human-centred AI, the cause she has championed at Stanford. A system that leaves its user less capable has failed a design test, whatever its benchmark scores. She would want the controls in section VIII built into the product by its makers, not bolted on afterwards by HR, and she would insist that the measure of success is human flourishing, not task completion.

Daniel Kahneman, speaking from the work, completes the lineage. His research showed that System 1 jumps to coherent conclusions from whatever is in front of it, and that confidence reflects the coherence of the story rather than the quality of the evidence. A fluent AI answer is the most coherent story available, delivered instantly. Kahneman would not be surprised that people surrender. He would be surprised that anyone expected otherwise.

The economists

Ethan Mollick, co-author of the BCG study and of Co-Intelligence, would push back hardest. He urges people to bring AI into everything they do, partly because the only way to learn where the jagged frontier lies is to probe it constantly. Avoidance, he would argue, is the bigger risk: the professional who refuses AI falls further behind than the one who over-trusts it. Fair. My answer is Long-AND, not Short-OR. Use it everywhere and govern how it is used. The choice is not between adoption and judgement. Mollick’s own data show the frontier is invisible without deliberate inspection.

Erik Brynjolfsson would see this essay as a chapter of his own. The Turing Trap argues that our incentives tilt towards AI that substitutes for people rather than AI that complements them. Cognitive surrender is substitution happening one decision at a time, beneath the level at which anyone decided to automate. His remedy would be structural: change what firms are rewarded for, so that augmentation pays.

Daron Acemoglu, with Simon Johnson in Power and Progress, would warn about “so-so automation”: technology that displaces human effort without delivering much productivity in return. Much surrender may be exactly that. The work looks faster, the quality is quietly worse, and the long-run capability of the workforce erodes. His point that the direction of technology is a choice, not a destiny, is the economic foundation of everything in section VIII.

The humanists

Lisanne Bainbridge, speaking from the 1983 paper, would note drily that none of this is new. Automate the easy parts, leave the human the hard parts, deprive them of the practice the hard parts require, and then blame them when they fail. Her prescription then is ours now: keep the human genuinely engaged, give them the practice and the information to take over, and design the collaboration rather than the replacement.

Nicholas Carr, whose book The Glass Cage traced automation complacency from cockpits to medicine a decade ago, would argue that the deepest cost is not error but meaning. Skill is not only instrumental; mastering a craft is part of how people find purpose in work. A workforce that has surrendered its thinking may still be productive, but it will be less fulfilled, and less fulfilled people make poorer stewards of the systems they supervise.

Sherry Turkle would focus on the friction we are engineering away. Her work on technology and conversation argues that capacities such as empathy, reflection and the tolerance of difficulty are built in effortful, sometimes uncomfortable exchanges. A system that always has an answer, always agrees and never makes us wait removes the very resistance through which judgement grows.

Yuval Noah Harari would make the civilisational point. In Nexus, he argues that AI is not a tool but an agent: the first technology that can make decisions and generate ideas by itself. If that is right, then surrender is not merely a productivity issue but a question of who, or what, is actually doing the thinking in a society. Boards that allow it to happen by default are making a political choice without noticing.

The sharpest critique

And the critique I take most seriously comes from no single panellist. The executive class has always practised a version of this. Analysts write the reports, consultants write the strategy, speechwriters write the speeches, and the leader signs. Perhaps AI has not created cognitive surrender at all. Perhaps it has simply democratised it, giving every employee the same option senior leaders have long enjoyed: to approve rather than to think.

If that is true, it is not a comfort. It is an indictment. The organisations that will thrive are the ones whose leaders can still do the thinking themselves, and who insist that their people, and their machines, show their working.

Where the panel lands

Strip away the disagreements and the panel converges on three things. First, surrender is real and measurable. Second, it is shaped far more by design and incentives than by individual virtue. Third, the remedy is neither abstinence nor blind adoption, but deliberate use: inside the frontier, delegate freely; at the edge of it, think first; beyond it, keep your own hands on the controls.

XII. The Fork

I often describe our moment as a fork in the road. One path leads to Star Trek: abundant machines, and humans freed to explore, create and decide. The other leads to Mad Max: abundant machines, and humans scrabbling for relevance among them. Most of the forces pushing us down one path or the other are loud: geopolitics, energy, capital, regulation. Cognitive surrender is a quiet one. It pushes us towards the second road not through catastrophe but through convenience, one accepted answer at a time.

The crew of the Enterprise did not stop thinking because the ship’s computer could answer almost anything. They asked it better questions, argued with its conclusions, and overruled it when the situation demanded. The computer was extraordinary; the bridge was still staffed by people who could fly the ship. That is the posture we need to design for. Not less AI, but AI that strengthens the human who uses it, inside organisations that reward the person who disagrees with the machine when the machine is wrong.

If I were advising a board tomorrow, I would ask for five things.

1. A surrender audit. Map the decisions in the business where AI now shapes the outcome, and tier them by reversibility and impact.
2. Human telemetry. For every high-tier decision, start measuring override rates, time-to-accept and justification depth, as IMDA’s framework now recommends.
3. Designed friction. Introduce prediction-before-reveal and reasoned approval on the decisions that matter, and remove friction everywhere else.
4. Protected apprenticeship. Configure AI as tutor, not ghostwriter, for those still forming their judgement, and assess reasoning as well as results.
5. Incentives that pay for judgement. Reward the person who caught the error, not only the person who shipped the most.

None of this slows adoption. It makes adoption safe to accelerate. Organisations that can prove their humans are still thinking will be trusted with more autonomy, by regulators, clients and their own boards. Those that cannot will discover, at the worst possible moment, that their human in the loop was a white flag all along.

The goal, as Steven Shaw puts it, is to gain when the tool is right and lose nothing when it is not. That will not happen by default. It will happen only if we measure it, govern it and build it in.

Short-term turbulence, long-term abundance. But only if we keep our hands off the white flag.

Sources

• Bartleby, AI and the danger of cognitive surrender, The Economist, 30 April 2026 (print edition, 2 May 2026)
• Shaw, S. D. and Nave, G., Thinking: Fast, Slow, and Artificial, Wharton School Research Paper, SSRN, 2026
• Wharton researchers explain how cognitive surrender takes hold, HR Executive
• Thinking Fast, Slow, Artificially, Wharton at Work, May 2026
• Dell’Acqua, F. et al., Navigating the Jagged Technological Frontier, Organization Science, 2026
• Lee, H.-P. et al., The Impact of Generative AI on Critical Thinking, CHI 2025 (coverage)
• Gerlich, M., AI Tools in Society, Societies, 2025
• Kosmyna, N. et al., Your Brain on ChatGPT, MIT Media Lab, 2025 (coverage)
• Shen, J. H. and Tamkin, A., How AI assistance impacts the formation of coding skills, Anthropic, January 2026
• Budzyń, K. et al., Endoscopist deskilling risk after exposure to AI in colonoscopy, Lancet Gastroenterology & Hepatology, 2025 (summary)
• Sparrow, B., Liu, J. and Wegner, D. M., Google effects on memory, Science, 2011 (Columbia summary)
• Dahmani, L. and Bohbot, V. D., Habitual use of GPS negatively impacts spatial memory, Scientific Reports, 2020
• Bainbridge, L., Ironies of Automation, Automatica, 1983
• Parasuraman, R. and Manzey, D., Complacency and Bias in Human Use of Automation, Human Factors, 2010
• BEA findings on Air France 447 (Irish Times)
• Brynjolfsson, E., The Turing Trap, Dædalus, 2022
• EU AI Act, Article 14; Fink, M., Human Oversight under Article 14
• IMDA, Model AI Governance Framework for Agentic AI (January 2026 launch; May 2026 update)
• The cost of letting AI think for you, The Deep View