Critical Thinking and Decision-Making

In a lab at MIT's Media Lab, a group of volunteers sat down to write essays. Some wrote with the help of ChatGPT. Some used a search engine. Some used nothing but their own minds. All the while, electrodes tracked the electrical conversation happening across their brains. When the researchers, led by Nataliya Kosmyna, published their findings in mid-2025 under the arresting title Your Brain on ChatGPT, one detail stood out from the tangle of neural data. The people who had leaned on the chatbot showed the weakest connectivity between brain regions — the measurable signature of engaged, effortful thinking. And when asked, minutes later, to quote a sentence from the essay they had just "written," many of them could not. They had produced a document. They had not, in any meaningful sense, produced a thought.

That gap — between generating an output and doing the cognitive work the output is supposed to represent — is the heart of a growing unease. It has a name now, or several names, in a fast-multiplying research literature: cognitive offloading, decision-making deskilling, the illusion of understanding. What they describe is not that AI makes us stupid. It is subtler and more troubling than that. AI makes it possible to skip the parts of thinking that build the capacity to think, and to do so without noticing anything has been lost.

What the research actually shows

The alarm did not come from one study but from a cluster of them arriving in close succession, each looking at a different population and finding a version of the same thing.

The most cited is the work of Michael Gerlich, published in the journal Societies in January 2025. Gerlich surveyed 666 people across age groups, measuring their use of AI tools against their performance on a standard critical-thinking assessment. He found a strong negative correlation: heavier AI users scored lower on critical thinking, and the relationship ran through a measurable mediator — cognitive offloading, the habit of handing mental tasks to an external system rather than performing them internally. The effect was sharpest among the youngest participants, who both offloaded the most and scored the lowest.

A parallel finding came from inside the technology industry itself. Researchers at Microsoft and Carnegie Mellon, led by Hao-Ping Lee, surveyed 319 knowledge workers in early 2025 about how generative AI shaped their thinking at work. The pattern they uncovered was quietly damning. The more a worker trusted the AI, the less critical thinking they reported applying to its output. The more they trusted their own expertise, the more they scrutinized it. Confidence, in other words, was a zero-sum resource: faith in the machine came directly out of the account of one's own judgment. Workers described shifting from doing the task to verifying the AI's version of it — and verification, done under time pressure with a plausible draft in hand, tends to collapse into a nod.

Then there are the education studies, which supply the chapter's most quoted figure. In controlled coursework experiments, students who used ChatGPT to work through problems consistently scored higher on immediate task completion but lower on later tests of conceptual understanding — in one widely reported chemistry-course study, roughly 17 percent lower on questions probing whether they actually grasped the material rather than merely produced correct answers. The signature is consistent: the AI lifts the visible output and depresses the invisible learning. A larger and better-controlled experiment by Hamsa Bastani and colleagues at Wharton, run in a Turkish high school, made the mechanism unmistakable. Students given a GPT-4 tutor sailed through practice problems — and then, when the tutor was taken away for the exam, performed worse than classmates who had never had it. The crutch had prevented the leg from strengthening.

The scale of this, honestly assessed, is not civilizational collapse. These are effect sizes in surveys and classroom experiments, not the fall of reason. But the nature of the finding is what unsettles researchers. It is not that people using AI know fewer facts. It is that the practice through which judgment is built — struggling with a hard problem, generating wrong answers, noticing why they are wrong — is precisely the practice AI removes. What atrophies is not knowledge but the capacity to generate it.

The pull toward the machine: algorithm appreciation

To understand why this deference sets in so easily, start with a counterintuitive experiment from 2019. Jennifer Logg, Julia Minson, and Don Moore ran a series of studies asking people to make estimates — forecasting a person's weight, predicting a song's popularity, judging romantic compatibility — and then offered them advice. Sometimes the advice was labeled as coming from a human expert. Sometimes it was labeled as coming from an algorithm. The content was identical. People consistently shifted their answers more toward the advice when they thought a machine had produced it.

Logg named the effect algorithm appreciation, and it inverted a long-standing assumption in psychology. Earlier work had described algorithm aversion — a stubborn human tendency to distrust formulas even when they outperformed us. But that aversion, it turned out, was fragile and context-dependent. When the task looked objective and quantitative, and when people lacked deep expertise of their own, appreciation dominated. We do not merely accept machine advice; we privilege it, granting the algorithm an aura of neutrality, comprehensiveness, and freedom from the messy motives we know human advisors to have.

The trouble is that this deference scales with the machine's fluency, not its correctness. A large language model does not "know" in the way an expert knows. It produces the statistically likely continuation of a prompt, delivered in the confident, well-organized prose we have learned to associate with competence. Algorithm appreciation means we bring to that output a lowered guard at exactly the moment a raised one is warranted — because the system that generated it has no model of the values, the context, or the stakes involved, only of what a good answer tends to look like.

Five mechanisms, and how they compound

Algorithm appreciation is one of five distinct routes by which AI erodes critical thinking. They are worth separating, because they operate differently and demand different responses — and because they reinforce one another in a way that makes the whole more corrosive than the sum.

The compounding is the danger. Cognitive offloading removes the practice that would let you catch an error. The illusion of understanding makes you believe you have checked when you have not. Algorithm appreciation predisposes you to defer in the first place. Bias amplification means the errors you fail to catch are systematic rather than random. And the collapse of intellectual humility strips away the one disposition — maybe I'm wrong, let me look again — that could interrupt the cycle. Each mechanism disables the defense against the next. A person can be caught in all five at once and experience it, from the inside, as simply getting work done efficiently.

The illusion of understanding

Of the five, the illusion of understanding deserves the closest look, because it is the one that hides its own operation.

When you reason through a problem yourself, you do a specific sequence of things. You notice what you do not know. You go looking for the missing piece. You weigh whether a source is reliable or slanted. You fit the new information against what you already believe, and when it clashes, you have to do something about the clash — revise the belief or reject the evidence, but not both silently. You make the trade-offs explicit because you are the one making them.

When you read AI-generated reasoning, almost none of this happens. The output arrives pre-assembled: the assumptions already chosen, the evidence already selected, the trade-offs already resolved and smoothed into confident prose. Your task shrinks to a binary — accept or reject — and, most of the time, accept. Yet the experience is deceptively similar. You have read words that look exactly like the trace of careful thought. Your brain, pattern-matching, registers "I have engaged with this problem analytically." This is why the MBA students in the case-analysis studies of 2025 could summarize an AI's business analysis fluently and then fall silent when asked to find the flaw in it or propose a different analytical frame. They had watched reasoning happen. They had not done it, and so they had no purchase on it.

The false confidence this produces is self-reinforcing in a particularly nasty way. Because you feel you understood, you see no reason to verify. Because you do not verify, you never discover that you did not understand. The loop closes, and each pass through it deepens a conviction that is not backed by competence. It is the cognitive equivalent of watching someone lift weights and feeling your own muscles grow.

Decision-making deskilling

There is a second, slower failure that unfolds not in a single essay but over months and years of professional practice. Psychologists studying workplace automation call it decision-making deskilling, and its arc is depressingly predictable.

graph LR
  A[AI takes over<br/>routine decisions] --> B[Human attention<br/>freed — feels beneficial]
  B --> C[Underlying judgment<br/>loses fluency from disuse]
  C --> D[Automation boundary<br/>expands upward]
  D --> E[Novel or failure<br/>situation arrives]
  E --> F[Human must decide<br/>alone — skill has atrophied]

It begins as an unambiguous good. The AI absorbs the repetitive, low-stakes decisions, freeing human attention for harder calls. But the routine decisions were where fluency was maintained. Deprived of that daily practice, judgment quietly rusts. Meanwhile the system improves, and the boundary of what it handles creeps upward, swallowing decisions that used to require a person. Then comes the moment the whole edifice was never designed for: the AI is unavailable, or the situation is genuinely novel, or the recommendation is subtly, catastrophically wrong — and the human who must now think independently discovers that the capacity to do so has withered at precisely the moment it is needed most.

This is not speculation about AI's future. It is documented history in every domain that automated ahead of the rest. Aviation learned it the hard way: pilots who fly mostly on autopilot show measurably degraded manual-flying skills, and accident investigators have traced disasters — Air France 447 being the grim textbook case — to crews who, when the automation abruptly handed control back, could no longer fly the aircraft they had been "flying" all night. Radiology is living through its own version now. A 2025 study in The Lancet Gastroenterology & Hepatology found that endoscopists who had grown used to AI assistance detecting polyps became worse at spotting them on their own once the AI was switched off — deskilling caught in real time, in working clinicians, not students. Algorithmic traders who cannot operate when their systems fail complete the picture. The pattern is invariant: automated help reduces the frequency of unassisted practice, and unassisted performance decays to match. What is new is only the breadth. AI is now doing this to analysis, planning, and judgment across all of knowledge work, in professions that never before had to worry about it.

The cruelty is in the incentives. Maintaining the skill requires deliberately slowing down — turning off the tool, working the problem by hand, courting inefficiency on purpose. Staying competitive requires keeping pace. And AI is what makes the pace possible. The rational short-term move for any individual is exactly the one that erodes the long-term capacity of everyone.

The bias amplification paradox

AI is marketed, more than anything else, as a corrective to human bias — a cool, consistent machine to replace hot, prejudiced people. In a narrow sense the pitch is fair: a formula applied uniformly does eliminate the idiosyncratic swings of individual human mood and prejudice. But it swaps a visible, variable bias for an invisible, systematic one, and that trade can leave things worse.

The mechanism is straightforward. AI systems learn from historical data, and historical data encodes historical discrimination. Amazon famously discovered, and scrapped, an internal résumé-screening tool that had taught itself to penalize the word "women's" and to downgrade graduates of two all-women's colleges — not because anyone programmed it to, but because it had faithfully learned the patterns of a decade of male-dominated hiring. The danger is not that such a system is biased; humans are too. The danger is that its bias wears the costume of objectivity. A biased human interviewer can be argued with, second-guessed, overruled. A biased algorithm, trusted as neutral by users primed with algorithm appreciation, applies the same skewed judgment to every candidate, consistently, at scale, and invisibly. The same dynamic runs through AI-generated news summaries that quietly inherit the slant of their source material, feeding readers a framing they might have questioned had they met it in the original.

Here is the paradox in full. Bias detection is itself a critical-thinking skill — built by years of interrogating sources, hunting for contradictory evidence, and catching one's own assumptions in the act. When AI becomes the primary lens through which information arrives, that skill atrophies along with the others. So AI does not reduce bias so much as relocate it: out of the visible realm of individual human judgment, where it can be spotted and contested, and into the invisible substrate of the system, where the very faculty that might have caught it has been allowed to fade.

A generation that never built the muscle

All of these effects grow more serious when we ask who is being deskilled, because there is a sharp distinction between losing a capacity and never having built it.

An experienced professional who leans on AI is like a fluent speaker who stops using a second language — the skill fades, but the scaffolding remains, and with effort it can be recovered. A young person who has spent their entire formative education with AI at hand is in a different situation entirely. They are not recovering a lost capacity. They may never have developed it. Critical thinking is built through iteration — hard problems, wrong answers, feedback, revision — and a student who routes that whole cycle through a chatbot completes the assignment, passes the class, and skips the only process that would have built the underlying ability. Surveys through 2025 found large majorities of recent graduates using AI routinely for exactly the assignments meant to teach analysis and argument.

Employers have started to notice. In a 2025 survey of hiring managers, 54 percent reported that recent graduates showed weaker analytical and critical-thinking skills than earlier cohorts, and many named AI dependency during education as the cause. Read carefully, that number is significant not because a single year's graduates are unusually weak, but because it suggests the baseline is shifting — that the decline researchers measure in the lab is already surfacing at the professional entry point, where the next generation of judgment is supposed to be forming. And the recovery problem is genuinely harder for this cohort. You cannot dust off a skill you never had. Rebuilding a capacity that was lost is remediation; building one that was never there, in adults past the developmental window that made it cheap to acquire, is a far steeper climb.

The societal stakes, and the honest limits of what we know

Scale these individual effects up and the stakes change in kind, not just degree.

Democracy runs on a distributed capacity that is easy to take for granted: enough citizens able to weigh competing claims, detect manipulation, and reason toward a judgment. Degrade that capacity across a population and the whole system grows more vulnerable — to misinformation, to propaganda, to the confident demagogue whose appeal is precisely that he saves people the labor of thinking. Complex institutions depend on the same distributed judgment. Governments, hospitals, corporations, and courts work because people at many levels are exercising independent scrutiny, catching each other's errors before they compound. When decision-makers up and down a hierarchy all defer to the same AI recommendations, an error introduced at the top can propagate untouched through every level that should have questioned it — a single point of failure wearing the mask of many independent checks.

Beneath the institutional risk sits something more personal: human agency itself. To outsource the reasoning that shapes your choices is, in a real sense, to become less the author of your own life and more the executor of an algorithm's. Technological lock-in sharpens the worry. A population that has grown dependent on AI to reason is poorly placed to cope when those systems fail, are compromised, or fall under the control of actors whose interests are not its own — and, having lost the skill, may not even be able to recognize that its reasoning is no longer its own.

Honesty requires holding all of this at arm's length. The evidence, weighed carefully, has real limits, and the book's obligation is to name them rather than paper over them with confident prose.

The first limit is causal. AI use is tangled up with a dozen other variables — time pressure, changing educational norms, shifting workplace culture, the general acceleration of everything. A correlation between heavy AI use and low critical-thinking scores, as in Gerlich's study, is compatible with AI causing the decline, with already-weaker thinkers gravitating to AI, or with some third factor driving both. The Microsoft/CMU work is self-reported, which measures what workers believe about their thinking, not the thinking itself. The Bastani math-tutor experiment is the strongest evidence precisely because it was randomized and measured performance directly — and stronger claims will require more of exactly that: controlled, longitudinal designs that follow the same people over years and can separate the AI's effect from everything else moving at once.

The second limit is conceptual, and it is the sharpest challenge to the whole thesis. Is what we are measuring degraded critical thinking, or merely changed critical thinking? Every previous cognitive offloading technology provoked the same panic — Socrates warned that writing would destroy memory, and in a narrow sense he was right, yet we do not mourn the loss. It is genuinely possible that some of what looks like decline is an adaptive reallocation: effort moving away from generating analysis and toward the higher-order task of supervising, prompting, and auditing an AI that generates it. That would be a skill, not a deficit — and a valuable one. The honest position is that we cannot yet cleanly distinguish the two. The deskilling evidence from aviation and radiology suggests the pessimistic reading is not paranoia; the history of offloading technologies suggests the optimistic reading is not naïve. Both may be true at once, for different skills.

The third limit is temporal. The gravest consequences — democratic fragility, institutional cascade failure, a hollowing of collective agency — are precisely the ones we cannot measure in advance. They are the kind of harm that becomes visible only after it has occurred, in the form of a specific failure traceable to a chain of humans who all trusted the same machine and none of whom checked. Whether these outcomes are predictable from current trends or will only announce themselves through some future event is, right now, unknown. That uncertainty is not a reason to dismiss the risk. It is the defining feature of it.

What to do about it

If the diagnosis is even partly right, the response cannot be to ban the tools — the productivity is real and the incentives are overwhelming. It has to be to redesign the moments where thinking matters most.

In high-stakes domains — hiring, lending, criminal justice, medicine — the design principle should be to make AI demand human evaluation rather than invite passive sign-off. A system that outputs a verdict — "reject this candidate," "deny this loan" — invites a rubber stamp. A system that surfaces the specific evidence, flags its own uncertainty, presents the strongest case against its recommendation, and requires the human to articulate a reason before proceeding forces the cognitive engagement that a clean verdict lets us skip. The goal is friction, deliberately placed: oversight that is meaningful rather than performative.

In education, the obligation is to make the practice of reasoning the measured output, not the finished product. When AI can produce a polished essay in seconds, grading the essay measures nothing. Assessment has to move toward the process — oral defenses where a student must reason live, drafts that show the thinking developing, in-class problem-solving without tools, assignments that require critiquing an AI's analysis rather than commissioning one. The point is not to catch cheaters. It is to relocate the assessment to the one place AI cannot stand in for the student: the act of thinking itself.

In organizations, the interventions should be proportionate to the stakes. For a firm whose competitive edge is the quality of its judgment, that might mean mandatory AI-free problem-solving sessions to keep the muscle alive, structured reasoning reviews where a decision-maker must show their work independent of any AI recommendation, and a deliberate policy of rotating people through unassisted practice — the professional equivalent of pilots being required to hand-fly, precisely so the skill is there on the day the automation isn't.

None of this reverses the trend on its own. But it points at the right target. The danger of AI is not that it thinks for us. It is that it lets us believe we are still thinking when we have quietly stopped — and the remedy is to build, deliberately and against the grain of every efficiency incentive, the moments that make us do the work.

Summary

1. The finding is real but bounded. Multiple 2025 studies — Gerlich's survey of 666 people, the Microsoft/Carnegie Mellon survey of 319 knowledge workers, classroom experiments showing roughly 17 percent lower conceptual understanding, and Bastani's randomized math-tutor trial — converge on a negative relationship between heavy AI use and critical thinking. The effect sizes are modest; the nature of the effect is the concern, because AI removes the very struggle through which judgment is built.

2. Five mechanisms compound. Cognitive offloading weakens skills through disuse; the illusion of understanding lets reading reasoning masquerade as doing it; algorithm appreciation drives uncritical deference to machines; bias amplification hides systemic prejudice behind a veneer of objectivity; and the collapse of intellectual humility strips away the disposition that could interrupt the cycle. Each disables the defense against the next.

3. Deskilling is documented, not hypothetical. Aviation and radiology — most recently a 2025 Lancet study of endoscopists who lost polyp-detection skill once AI was removed — show the same arc AI is now bringing to all of knowledge work: automation reduces unassisted practice, and unassisted skill decays to match.

4. The generational risk is qualitatively different. Young professionals who never built critical thinking without AI are not recovering a lost skill but attempting to build one past the point where it comes cheaply — and 54 percent of hiring managers already report the gap in recent graduates.

5. The stakes rise from the individual to the collective: democratic resilience, institutional error-correction, and human agency itself all rest on distributed, independent judgment that uniform deference to AI erodes invisibly.

6. Honesty about the evidence is non-negotiable. We cannot yet cleanly separate AI's causal effect from confounds like time pressure and educational change, nor distinguish degraded thinking from an adaptive shift toward supervising AI, nor measure the societal consequences before they occur. The response is not to ban the tools but to redesign the high-stakes moments — through friction in AI oversight, process-based assessment, and deliberate unassisted practice — so that the work of thinking still gets done.

Sources

• Algorithm appreciation: People prefer algorithmic to human judgment (Logg, Minson & Moore, 2019)
• AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking (Gerlich, 2025, Societies)
• The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort (Lee et al., Microsoft Research & CMU, 2025)
• Your Brain on ChatGPT: Accumulation of Cognitive Debt (Kosmyna et al., MIT Media Lab, 2025)
• Generative AI Can Harm Learning (Bastani et al., Wharton, 2024)
• Amazon scraps secret AI recruiting tool that showed bias against women (Reuters)
• Endoscopist deskilling after exposure to AI in colonoscopy (Lancet Gastroenterology & Hepatology, 2025)

Last updated: 2026-08-30