The Cognitive Obesity Hypothesis: AI and the Risk of Mental Atrophy

1. Introduction

In 1953, Morris and colleagues found that physically active bus conductors had substantially lower mortality from coronary heart disease than the more sedentary drivers working in the same transport system. Later post-mortem (necropsy) evidence likewise showed more ischaemic myocardial fibrosis (heart-muscle scarring from a restricted blood supply) among men in light occupations than among those in physically demanding ones [1,2]. The observation, simple and now seventy years old, established a principle whose eventual scale was barely imaginable at the time: that the systematic removal of effort from daily life is itself a pathology. Seven decades of epidemiology have since confirmed that the modern human body, freed by machines from sustained physical demand, has become measurably weaker, less metabolically resilient, and more prone to chronic disease than that of any previous generation [3,4,5,6].

The same logic now applies, at a velocity its physical predecessor never had, to the human mind. In roughly 30 months, generative artificial intelligence has moved from research curiosity to default cognitive prosthesis. Falling inference costs have likely accelerated this: Epoch AI estimates that the cost of large-language-model inference at a fixed level of performance has been falling by roughly 10- to 900-fold per year, depending on the task [7]. Today, AI systems are coding software, synthesizing information, conducting research, reviewing documents, performing medical, legal, and financial analyses, and automating decisions, taking on work that people previously did through deliberate cognitive effort [8]. The transition is now extending beyond passive chat assistants toward AI agents. An AI agent is a system that does not merely answer a single question but plans a multi-step task, chooses and operates tools, takes actions, and checks its own outputs; such systems substitute not for one answer but for an entire workflow of decision and execution [9,10]. The question we pose is empirical, not rhetorical: if removing physical effort produced quantifiable physical decline, what should we expect from removing cognitive effort?

We treat the answer as a hypothesis, not a verdict. We argue that the three-step mechanism behind the body’s two-century decline (substitution, redesign, and rebound) is now plausibly operating on cognition, but at a different scale: broader in scope, faster in diffusion, and potentially self-undermining in ways that make reversal harder. Using the historical record of physical decline as a guide, we specify what signals to look for, identify five domains where early evidence may already be detectable, confront the strongest opposing case (that AI augments rather than depletes the mind), examine why the cognitive trajectory is unlikely to self-correct as physical fitness partially did, and outline the design and policy responses whose absence, in the case of the body, helped produce a global epidemic of non-communicable disease. Throughout, we are explicit about both the strengths and the limits of the evidence behind each step, not to present cognitive decline as inevitable, but to issue an early warning that, if the same forces that weakened the body are now being applied to the mind, the trajectory must be recognized, debated, and deliberately shaped before it becomes harder to reverse.

Two arguments in this paper are, to our knowledge, relatively underdeveloped in the existing debate, and we state them plainly to distinguish the argument from the long-standing worry that technology makes us dependent. The first is a mechanistic homology: the substitution → redesign → rebound mechanism is not asserted by analogy but transferred from a fully worked, century-scale physical case whose course and consequences are documented. The second is a cellular argument: the biology of effort (effort-dependent survival of new neurons, and the neurotrophic signaling that supports it) yields a falsifiable prediction that the substituted faculties will not merely lag but atrophy. These two claims are the distinctive contribution of this paper, and they are the claims we ask to be tested.

Because “mental fitness” appears in our title, we define the term before using it as an object of measurement. By cognitive (mental) fitness, we mean two related but distinguishable capacities. The first is fluid capacity: the ability to reason through novel problems, detect relations, adapt to unfamiliar situations, and sustain attention without relying primarily on previously learned knowledge [11]. This is the component most closely analogous to muscle because it reflects the active, effort-dependent capacity to solve new problems. The second is crystallized competence: the accumulated knowledge, procedures, vocabulary, and domain-specific skills acquired through education, practice, and experience, including writing, coding, mathematical technique, and clinical judgment [11]. This distinction matters because a tool may erode practiced competences by replacing their use while leaving fluid capacity largely intact; alternatively, as the hypothesis considered here suggests, prolonged substitution may eventually affect both. We therefore map each signal discussed below to the component of mental fitness it most directly indexes. The cross-cutting terms used throughout the paper are collected in Table 1.

2. The Body as a Model System: Two Centuries of Physical Decline

The physical case is not used here as a loose analogy, but as a worked historical example. Its trajectory, mechanisms, and consequences are sufficiently documented to help identify what should now be examined in the cognitive domain. As summarised in Table 2, the decline of human physical capacity unfolded through three overlapping waves: occupational, domestic, and environmental. Together, these waves displaced large amounts of routine muscular work from everyday life.

The first wave, occupational, displaced muscular labor from production. In 1790, about 90% of the US labour force were farmers; agriculture employed 41% in 1900, 1.9% by 2000, and under 1% by 2024 [14,15,16]. Mechanization compressed the labor needed to produce 100 bushels of wheat from roughly 370 hours in 1800 to about 7 hours in the early 1980s [14,15]. Analyzing five decades of US Bureau of Labor Statistics data, Church and colleagues documented a fall of 142 kcal/day in occupational energy expenditure for men (and 124 kcal/day for women) between the early 1960s and 2008 [3].

The second wave, domestic, mechanized the household. Washing machines, refrigerators, vacuum cleaners, and dishwashers compressed a weekly housework load of about 58 hours in the early 1900s to roughly 18 hours by 1975 [12]. Using direct calorimetry, Levine and colleagues estimated that replacing hand-washing, hand-laundering, and walking commutes with machines eliminated roughly 110 kcal/day from a typical adult’s expenditure [17], a reduction only partially offset by leisure exercise [18].

The third wave, environmental, redesigned space itself. Between 1960 and 2009, the share of US workers walking to work fell from about 10% to 3%, with active travel declining especially among children, women, and older adults [13]. The International Physical Activity and Environment Network study of 14,000 adults across 12 countries found that built-environment features (walkability, land-use mix, transit access) are reproducibly associated with physical activity and obesity; a 2.1-million-participant smartphone study using residential relocations as natural experiments was later able to probe this link causally [19,20]. The car-dependent environment did not merely permit inactivity; it built movement out of the set of available choices.

Taken together, these reductions amount to several hundred kilocalories per day, enough to account for a large share of the population-level weight gain since the 1960s [3,18]. The physiological consequences are by now uncontroversial. The NCD Risk Factor Collaboration’s 2024 Lancet synthesis (3,663 studies, 222 million participants) found that adult obesity more than doubled between 1990 and 2022, and that over one billion people across all ages are now living with obesity [5]. Ekelund and colleagues’ meta-analysis of more than one million people quantified a 59% mortality excess in the most sedentary quartile [4]. Lee and colleagues attributed about 9% of premature mortality worldwide (5.3 million deaths a year) to physical inactivity, a toll comparable to tobacco [21].

The change is not merely behavioural but structural, and recent enough to measure directly: grip strength, an integrative biomarker of musculoskeletal health, fell roughly 20% in young adult men between 1985 and 2016 [22,23], and youth aerobic fitness in mostly developed countries declined steadily through the 1980s and 1990s before stabilising in many high-income countries in the early 2000s [24,25,26,27]. Over far longer timescales the skeleton tells the same story: trabecular bone density (the spongy interior bone that remodels under load) and limb-bone strength fell by roughly a fifth to a third as mobility declined with the adoption of farming [28,29,30]. But that Neolithic transition unfolded over millennia and for a different reason (the spread of agriculture), and we lean on it only to show that disuse reshapes bone, not to time the modern decline.

Communities that still live physically demanding lives make the contrast vivid. Old Order Amish, whose routines preserve nineteenth-century patterns of manual labor, average roughly 18,000 steps a day (men) and had a community obesity rate of about 4%, against roughly 31% among contemporaneous US adults [31]. The Hadza of Tanzania accumulate around 135 minutes of moderate-to-vigorous activity daily (more than ten times the industrialized average) and show no detectable rise in cardiovascular risk factors across the lifespan [32]. These are not genetic outliers but the same human body meeting a higher demand: the pathology emerged not from the body itself but from the disappearance of the demands for which it was built.

The two-century decline in physical capacity can be understood as a three-step mechanism: effort is first substituted by machines, the surrounding environment is then redesigned around that substitution, and the resulting efficiency gains produce rebound effects that further reduce everyday exertion. It is this mechanism, rather than the specific loss of muscular work, that provides the relevant model for the cognitive domain.

• Substitution. A machine substitutes for an effort. The substitution is welcomed: it relieves drudgery, expands output, and is in nearly every individual instance unambiguously good. The reaper, the washing machine, and the automobile each removed a specific physical burden no rational person would choose to keep.
• Environmental redesign. The substitutions accumulate, and the environment is rebuilt around them. Cities are designed for cars rather than pedestrians, homes for sitting rather than working, and workplaces for keyboards rather than tools. The default option, the path of least resistance, no longer requires the effort it once did. The choice architecture (the way the available options are arranged and defaulted) inverts: physical activity, once the unavoidable substrate of life, becomes a discretionary leisure pursuit, increasingly the preserve of those with time, income, and self-discipline.
• Rebound (Jevons paradox). Efficiency gains do not reduce aggregate consumption of the now-cheaper resource; they expand it. William Stanley Jevons observed in 1865 that James Watt’s improvements to the steam engine had not conserved coal but had made it so useful that consumption exploded [33]. The pattern is one of the most durable regularities in the economics of general-purpose technology, though its size is contested and not every efficiency gain rebounds fully [34]. UK lighting consumption rose by several orders of magnitude over three centuries as the cost of light collapsed [35]; aviation fuel use and emissions quadrupled between 1960 and 2006 despite a roughly 70% gain in fuel efficiency per passenger-kilometre [36,37]; and US vehicle-miles travelled nearly tripled between 1971 and 2019 even as new-vehicle fuel economy improved [38,39]. In the case of the body, the rebound took the form of suburban sprawl, longer commutes, larger homes, and expanding sedentary leisure, each absorbing the time and money that mechanization had freed. The combined effect of substitution, redesign, and rebound was a population-level transition that no individual chose and no regulator authorized. The body did not fail. The species had built an environment in which the body’s former function was no longer demanded.

3. The Mechanism Repeats: From Body to Mind

We argue that the same three-step mechanism is now beginning to operate on cognition. Its structure is analogous to the physical transition, but its tempo is radically compressed: what unfolded over centuries in the body is now occurring over years in the mind. Table 3 compares the two transitions directly and introduces the three differences that recur throughout the argument: cognitive substitution is broader, faster, and potentially self-undermining.

• Substitution. Generative AI substitutes for cognitive effort across a wider range of tasks, and more quickly, than any prior tool. Reading, summarising, drafting, coding, planning, and analyzing are increasingly delegated to large language models, whose performance has advanced rapidly in reasoning-intensive domains while inference costs at a fixed performance level have fallen by roughly 10- to 900-fold per year, depending on the task [7]. The adoption figures are striking but come almost entirely from commercial or self-report surveys and should be read as such: by early 2026, an industry impact report drawing on 400-plus companies estimated that more than nine in ten developers used AI tools at least monthly (more than two in five daily), with AI authoring roughly a quarter of merged code [40]; a separate developer survey put daily use among AI adopters at about 72% and AI-generated or assisted code at about 42% [41]. The OECD reports that, on average across member countries, occupations at highest risk of automation account for about a quarter of jobs [42]. The substitution now reaches past single answers: agentic systems (autonomous, multi-step pipelines that plan, select tools, act, and verify their own outputs) substitute for the entire workflow, removing the human not only from the final cognitive step but from the intermediate ones as well [9,10].
• Environmental redesign. The second wave is already visible. Search engines now answer rather than retrieve, so the act of synthesizing from sources is increasingly delegated. Schools are reorganizing assessment around the assumption that drafted text will be machine-generated; the OECD’s 2026 Digital Education Outlook documents pervasive integration of generative AI into formal instruction across member states [43]. Email, documents, and code are being rebuilt with embedded copilots whose default presence makes unaided cognition the deliberately effortful path. As with the suburb, the choice architecture is inverting: the default option will no longer require the cognitive effort it once did.
• Rebound. The Jevons paradox appears to be operating in AI with unusual speed: efficiency gains are not reducing aggregate demand for computation but are lowering barriers to use, thereby encouraging wider adoption, heavier deployment, and more frequent delegation of cognitive work [44,45]. As AI-mediated cognition becomes cheaper, it is consumed in greater quantity rather than reliably freeing human capacity for harder thinking. The January 2025 release of DeepSeek’s low-cost frontier models, widely interpreted as a challenge to the economics of large-scale training, strengthens rather than weakens this argument: by showing that frontier-level capability can be produced and deployed at lower cost, it makes broader and more intensive AI use more likely [46]. The cognitive analog of the suburb is thus emerging at software speed: an AI-mediated workflow in which questions are increasingly posed to a model rather than worked through by a person.

The three differences are these. First, the substitution is broader: where mechanization displaced muscular labor across production, transport, the household, and leisure over centuries, AI is being introduced across many cognitive domains in parallel. Second, the diffusion is faster: ChatGPT, released in late 2022, reached more than 700 million weekly active users (close to a tenth of the world’s adults) within about two and a half years [47]. Third, and most consequential, the substitution is self-undermining: the faculty being replaced is the very faculty needed to judge the replacement. A weakened bicep does not impair the judgment that decides whether to use it. A weakened capacity for sustained reasoning does.

4. Five Empirical Signals of Incipient Mental Decline

If AI use is producing cognitive decline, the claim should leave measurable traces. We catalog five early signals, each preliminary on its own but together forming a pattern consistent with the hypothesis (summarised in Table 4). Before presenting them, three distinctions and one caution are essential, because the strength of the whole argument turns on them.

A central claim of this paper depends on distinguishing three progressively stronger forms of cognitive loss. The first, offloading, is a reversible behavioral choice: a person delegates a mental operation to a tool when that tool is available. The second, deskilling, is the weakening of a practiced skill through disuse, usually recoverable with renewed practice; in the terminology of Section 1, it concerns crystallized competence. The third, atrophy, is the strongest claim: a durable and harder-to-reverse loss of underlying cognitive capacity, or the fluid substrate. Strong evidence of atrophy would be premature to expect, since generative AI has been in widespread use for only two years, and durable loss would likely require a longer period of sustained disuse. Yet some early findings already point in the predicted direction. Most of the evidence reviewed below, therefore, demonstrates offloading and short-term deskilling.

The question itself is not new. A 30-year body of research has already established that people offload cognitive tasks to machines. Sparrow et al. showed that when people expect information to remain available on a computer, they remember the information itself less well but remember where to find it, the original “Google effect” [48]. Risko and Gilbert’s review defined cognitive offloading and showed it has a cost-benefit structure that can be adaptive or maladaptive [49]. What is new is not offloading but its reach: search engines offload storage and retrieval; generative AI offloads a layer deeper (synthesis and reasoning), and AI agents offload the whole task. The signals below ask whether that deeper offloading is tilting the balance toward the maladaptive.

4.1. Signal 1: Neural Engagement During AI-Assisted Production

Kosmyna and colleagues at the MIT Media Lab recorded brain activity (EEG) in 54 adults asked to write essays under three conditions: unaided, search-engine assisted, and ChatGPT-assisted [50]. The ChatGPT group showed the lowest neural engagement, the weakest coordinated signaling between brain regions tied to executive control and meaning (cross-frequency directed connectivity), and the poorest recall of their own essays minutes after writing. When groups were swapped, and ChatGPT users wrote unaided, their connectivity stayed suppressed relative to the original unaided group, an after-effect the authors call “cognitive debt.” The result is preliminary in ways that matter: the sample is small (54 across three sessions, only 18 in the cross-over arm), the task is short and laboratory-bound, and a published methodological commentary argues the study is underpowered for several specific claims and that the “cognitive debt” interpretation outruns the data [51]. We therefore read it as the first direct neurophysiological measurement of the substitution effect, no more: a hypothesis-generating result whose direction matches offloading theory and whose value depends on replication at scale and over longer horizons. This signal speaks primarily to offloading and short-term after-effects. Nevertheless, because it shows reduced neural engagement that partly persists after the tool is removed, it may be read as an early warning signal in the direction predicted by the atrophy hypothesis.

4.2. Signal 2: Cognitive Offloading and Critical Thinking in Adults

Three 2025 surveys point the same way, though all are cross-sectional and none can establish direction. Gerlich, studying 666 UK adults, found that frequent AI-tool use was negatively associated with Halpern critical-thinking scores; the association tracked self-reported cognitive offloading and was strongest among younger participants who reported higher trust in AI [52]. Lee and colleagues, surveying 319 knowledge workers reporting on 936 AI-use episodes, found that workers report less cognitive effort on AI-assisted tasks and that their critical thinking shifts from forming ideas to verifying outputs; higher confidence in the AI was associated with less critical thinking enacted, higher self-confidence with more [53]. Tian and Zhang, surveying 580 Chinese university students, found AI dependence associated with lower critical thinking, an association that tracked self-reported cognitive fatigue and was buffered where information literacy was high [54].

A complementary line studies the failure mode directly. Le and Kunz, in six experiments (N = 1,370) on the adoption of AI outputs in service work, document AI complacency: workers accept AI-generated content without scrutiny even when it contains errors, the more so when no one is accountable for checking [10]. Romeo and Conti’s review of 35 studies on automation bias finds systematic miscalibration of trust, with users under-verifying AI recommendations, especially when verification is effortful [60]. Ibrahim and colleagues, synthesizing the overreliance literature, document that even skilled professionals, including developers approving AI-written code containing security flaws, accept fluent, confident outputs uncritically [9]. Fernandes and colleagues add experimental weight from two large studies (with a randomized no-AI control in the second): adults solving logical-reasoning problems with ChatGPT improved their scores yet most accepted the model’s answer after a single prompt, monitored their own correctness only weakly, and substantially overestimated how well they had done [61].

The convergence of self-report (Gerlich, Lee, Tian), behavioral and experimental (Le, Fernandes), and review (Romeo, Ibrahim) evidence is suggestive, but we resist treating it as confirmation: the survey studies share a framing and a plausible common confound (see Section 4.6), so their agreement is weaker evidence than the count of studies implies. The experimental studies, by contrast, are less vulnerable to that confound. These signals index offloading and, in the workplace studies, the beginnings of deskilling; they do not reach the fluid substrate.

4.3. Signal 3: Adolescent and Student Reasoning Under AI Dependence

Younger users are adopting these tools fastest, and the early data are mixed in instructive ways. The Common Sense Media 2024 survey of 1,045 US teenagers found that 70% had used at least one generative AI tool and that two in five had used AI for school assignments, nearly half of them without their teacher’s permission [55]. Lim argues, reviewing classroom evidence, that overreliance on generative AI can weaken students’ critical thinking and metacognition by reinforcing automation, confirmation, and anchoring biases unless use is supported by deliberate scaffolding [62]. The clearest demonstration to date is a field experiment with about 1,000 high-school students, reported in the OECD’s Digital Education Outlook: access to GPT-4 raised short-term task performance substantially while simultaneously leaving performance markedly worse once the model was withdrawn: output rose, the underlying skill fell [43]. The effects are recent, and confounding cannot be excluded; we therefore do not treat this as strong evidence of atrophy. We note only that the pattern is exactly what the offloading-to-deskilling hypothesis predicts: assisted output rises while unaided skill falls. Still, the result may represent an early phase in the direction predicted by the atrophy hypothesis, in which repeated offloading first improves performance with the tool, then weakens performance without it, and only longer exposure can show whether that weakness becomes durable.

4.4. Signal 4: Deskilling in Knowledge Work, with Agents as Accelerator

Deskilling is the signal where the historical literature is strongest, and it long predates AI. Bainbridge’s “ironies of automation” established four decades ago that automating a task erodes the operator’s skill and vigilance, so that when the automation fails the human is least able to step in [56]. That is precisely our self-undermining difference, observed in cockpits and control rooms long before chatbots; AI agents simply extend it from psychomotor tasks to reasoning itself. In medicine, El Tarhouny and Farghaly argue that AI overreliance diminishes diagnostic reasoning and pattern recognition, the clinical analog of Bainbridge’s deskilled operator [63]. In software, AI was authoring roughly 27–42% of committed code by early 2026, depending on the (commercial) survey [40,41]. Ferdman argues that what is at stake differs qualitatively from the historical deskilling of trades: AI threatens capacity-hostile environments in which the general capacities of personhood (epistemic, social, moral, creative) are displaced, not merely craft-specific skills [64,65,66].

The accelerant is the shift from passive assistants to agents. Where a chat assistant substitutes for the final output, an agent substitutes for the entire workflow (planning, tool selection, intermediate checks, and delivery), so the human is removed from the steps in which a skill is exercised and maintained. Le and Kunz find that when accountability for monitoring is weak, willingness to evaluate outputs falls and commission errors rise [10]. Direct evidence on agentic versus passive use remains sparse: published studies focus on chatbots, whereas these days the fastest-growing deployments are agents that can automate entire cognitive tasks. Closing that measurement gap is a precondition for diagnosing cognitive decline before, rather than after, the substituted skills have eroded.

4.5. Signal 5: Displacement of Desirable Difficulty

This signal differs in kind from the first four: it is a theory-based prediction, not yet a direct measurement. Four decades of cognitive psychology, developed by Bjork and others, show that the conditions making learning feel easy in the moment often impair long-term retention and transfer; the desirable difficulties (spacing, interleaving, retrieval practice, generation, productive struggle) are effortful precisely because the effort is what consolidates learning [57]. Generative AI, by design, removes generation effort and minimizes productive error, so the theory predicts impaired consolidation even where short-run performance improves. Bellwether’s 2025 synthesis applies this body of work to AI in K–12 settings [67].

But the prediction must be stated carefully, because not all difficulties are desirable. Sweller’s cognitive load theory distinguishes extraneous load (busywork that should be removed) from germane load (the effortful processing that builds understanding) [58,68]; the Bjorks’ own term is desirable difficulty, defined against undesirable difficulty. The same literature the alarm rests on also tells us that most friction is waste: Wilson and colleagues’ “eighty-five percent rule” implies that optimal learning occurs when learners err on about 15% of attempts, so roughly 85% of struggle can be removed without loss [59]. Mechanization removed undesirable physical effort, and we do not mourn the scythe. The real claim, then, is narrower and sharper than “preserve the difficulty”: AI tends to strip load indiscriminately, removing germane and extraneous difficulty alike, and the danger is the removal of the germane. Which cognitive effort is worth preserving, and who decides, is the question Section 6 must answer rather than assume.

4.6. Alternative Explanations

Before reading these signals as a decline, three non-causal accounts must be taken seriously, because each could generate the same pattern. First, reverse causation: people with weaker discernment or critical-thinking habits may adopt AI more readily, so the association runs from low ability to high use rather than the reverse. Second, selection and composition: heavy AI users differ from light users in age, occupation, and motivation, and the surveys above cannot fully adjust for this. Third, and most corrosive to the adult-decline claim specifically, construct validity: that claim rests on critical-thinking instruments (Halpern and similar) whose scores may not capture the capacity the title cares about, and which were not designed for an AI-saturated environment. None of these is fatal to the hypothesis, but each is a live competing explanation that only longitudinal, baseline-anchored measurement can adjudicate. We flag them here, where the signals are presented, rather than quarantining them in a limitations box.

4.7. The Augmentation Counter-Claim

The strongest case against this paper is not that the signals are weak but that the framing is wrong: a respectable tradition holds that cognitive tools augment rather than diminish the mind. Engelbart’s Augmenting Human Intellect and Licklider’s Man–Computer Symbiosis envisioned exactly the human–machine partnership now arriving, and the “bicycle for the mind” framing treats such tools as multipliers of human capacity [69,70]. Cognitive science supplies a mechanism: offloading low-value cognition can free working memory for higher-order work [49]. The most principled version is the extended-mind thesis of Clark and Chalmers, on which a tool tightly coupled to the brain becomes part of the cognitive system, so that measuring “unaided” performance is the wrong yardstick: the right unit is the human-plus-tool [71]. Distributed cognition is, moreover, not new: Wegner’s transactive-memory theory treats memory as a property of a group rather than a single mind, with knowledge distributed and retrieved across its members [72].

The most direct empirical challenge comes from inside the literature we surveyed. Hong and colleagues’ quasi-experiment (N = 240) found that scaffolded AI use (offloading routine subtasks while the learner focuses on critique and reflection) produced large gains in critical thinking and essay quality rather than losses (effect sizes reported in the source) [73]. This is counter-evidence to Signals 2–4, and we place it here rather than in a policy footnote: difficulty is not eliminated by AI but reallocated, and the reallocation can be designed for or against the learner. The design is quasi-experimental with a self-report mediator and, crucially, includes no AI-free transfer test, so the gains index student-plus-AI performance rather than internalized capacity.

Yet augmentation and depletion need not be alternatives; the same study can show both. Fernandes and colleagues, across two large studies of adults solving logical-reasoning problems with ChatGPT, found that AI use raised task performance substantially, confirming augmentation, while participants overestimated their scores by about four points, their confidence tracked correctness only weakly, and the Dunning-Kruger effect collapsed into near-uniform overconfidence as AI leveled performance across skill [61]. Higher self-rated AI literacy predicted worse calibration rather than better, and a monetary incentive for accurate self-assessment failed to repair it. The augmentation is thus real but partial: the tool makes users measurably “smarter” on the task while leaving them “none the wiser” about how they performed, degrading exactly the self-monitoring needed to catch a confident model’s errors, the augmentation thesis confirmed on the performance axis and undercut on the metacognitive one.

Why, then, expect decline rather than augmentation? Three reasons, and the historical record sharpens them. Earlier cognitive prostheses (writing, print, the pocket calculator) coincided with the twentieth-century rise in measured intelligence (the Flynn effect), not its fall, which is exactly what an augmentation theorist will cite. We answer that generative AI differs along the three axes named throughout: it is broader (a calculator offloaded arithmetic; a chatbot offloads synthesis, judgement, and, via agents, whole tasks), faster (prior tools diffused over generations, leaving time to adapt institutions and assessment), and self-undermining (a calculator does not erode the judgment that checks its answer; a reasoning prosthesis may). To the extended-mind objection specifically, we offer a narrower reply: even if the right unit is sometimes human-plus-tool, unaided capacity still matters, because the tool is not always present, it can be confidently wrong, and catching its errors requires exactly the undelegated judgment the loop erodes: Bainbridge’s irony, restated for cognition. What is genuinely new, against this 30-year baseline, is not that we offload but the character of the partner (always available, never pushing back) and the universal scale at which it now operates.

5. Why the Cognitive Transition Is More Dangerous and Will Not Self-Correct

The three differences do more than accelerate the timetable; they remove the margin of error that, in the physical case, lets the damage be recognized before it becomes irreversible. Put plainly: the faculty being replaced is the same faculty we need in order to judge the replacement. The body’s transition unfolded across generations, and its worst consequences did not become measurable until roughly a century after mechanization began. The cognitive transition denies us that interval. The automobile took about five decades to reach mass household penetration in the United States [74,75], while ChatGPT reached comparable population penetration in about two years [47]. And because the rebound now operates on the substituted faculty itself, the early signals point to a feedback loop in which substitution erodes the very capacity needed to assess it. The self-correcting forces that partly rescued the body (examined first) are largely absent, and the biology of effort (examined second) implies that the endpoint of the loop may not be stagnation but atrophy. Table 5 sets the two transitions side by side.

5.1. The Self-Correcting Forces That Slowed the Body Are Absent

The physical transition was eventually slowed in part by a spontaneous leisure-fitness culture, and two of its engines have no cognitive counterpart. The first was economic: a long-run wage premium for cognitive over manual labor created the time and means for discretionary exercise [76,77]. But generative AI is now compressing the returns to many of the cognitive skills it can perform, so that incentive may not push the same way over the next century. The second was visibility: muscularity and athletic competence are legible, socially rewarded signals, and the body’s changes can be seen and corrected through individual effort, whereas mental fitness has no equivalent visible marker: there is no cognitive analog of muscularity acquired in a gym and displayed at the beach, and reasoning ability stays invisible until it is tested. The motivational machinery that recruited a fitness culture for the body has little purchase on the mind. A third engine was cultural, and it has the longest history of the three. Deliberate physical training has been organised as a mass pursuit for millennia: the ancient Olympic Games and the Greek gymnasion made athletic effort a civic institution, and that tradition was revived and globalised in the modern Olympic movement, in compulsory school physical education, in youth leagues, and in the recreational running, cycling, and gym culture that now enrols a large share of the population, children included, in effort undertaken for its own sake. When mechanization finally removed muscular work from daily life, this tradition was already in place; the fitness response did not have to be invented, only scaled, because the infrastructure, the role models, and the social reward for physical effort already existed. The mind has no comparable inheritance. There are, to be sure, organized contests of intellect, chess and Go, competitive debate, mathematical and quiz olympiads, the World Memory Championships, and arguably the rise of esports, and we do not dismiss them. But they are marginal in participation beside physical sport, they reward narrow domain expertise rather than the general, transferable fitness our hypothesis concerns, and none is embedded in universal schooling as physical education is. Formal education does train the mind, but instrumentally, for credentials and employment, rather than as a lifelong, celebrated culture of cognitive exercise pursued voluntarily for its own sake. The reservoir of practice and motivation that the body could draw on would, for the mind, have to be built almost from nothing.

5.2. The Brain-Biology of Effort: Atrophy or Lag

The body literature provides more than a behavioral analogy; it also suggests a plausible biological mechanism, and this is where our second original claim lies. Effortful learning is associated with brain changes to a greater extent: animal studies show that it can determine whether newly generated neurons survive. Shors and colleagues showed that newborn neurons in the adult hippocampus are preserved only when animals subsequently engage in effortful learning: easy tasks do not rescue these cells, whereas more demanding learning does [78]. The signaling pathway that supports this plasticity is also partly characterized. Aerobic exercise increases brain-derived neurotrophic factor (BDNF), a protein involved in neuronal growth and survival, and BDNF in turn supports synaptic plasticity and adult neurogenesis [79]. The cellular logic is therefore a biological version of “use it or lose it”: cognitive effort helps maintain the neural substrate on which future cognition depends. At the behavioral level, this is the same principle captured by the desirable-difficulty literature, where effortful retrieval, generation, and productive struggle strengthen learning precisely because they require the learner to do cognitive work.

A related but distinct concept must be handled carefully. Cognitive reserve, the brain’s resilience to age-related damage, is supported by decades of cohort evidence showing that lifelong intellectual engagement can delay the clinical onset of dementia [80,81,82,83,84]. We use this literature only as suggestive of a mechanism, not as evidence that AI depletes cognitive reserve. The reserve literature concerns resilience to neuropathology in aging populations, whereas our argument concerns possible deskilling and reduced cognitive effort in otherwise healthy users, including young adults. Conflating the two would make the argument vulnerable to a simple objection: that we had mistaken aging-related decline for an effect of chatbot use. Similarly, fluid reasoning, the capacity to solve novel problems independently of acquired knowledge, declines across adulthood from a peak before age thirty [85]. The neurogenesis and reserve literatures therefore support only a limited inference: the cognitive baseline is not fixed, and effortful engagement may help maintain it. They do not show that AI use has already reduced it.

The implication for AI is structural, but we stress, a candidate mechanism rather than an observed effect. If new neurons survive only with effortful learning, and if frictionless cognitive prostheses displace the generation, retrieval, and productive struggle that constitute that effort, then the prediction is not merely behavioral lag but a measurable change in the neural substrate: atrophy rather than disuse. No direct evidence of this exists in the AI context; obtaining it is the purpose of the longitudinal infrastructure we recommend below, and the absence of any current test for it is a gap we treat explicitly in Section 7.

5.3. A Related Behavioural Effect: Sycophancy

One striking recent result sits adjacent to the five signals above, rather than within them, and we treat it accordingly. Cheng and colleagues’ 2026 Science study measured AI sycophancy across eleven leading models and tested its behavioural consequences in preregistered experiments involving 2,405 participants [86]. Across models, AI systems affirmed users’ actions far more often than humans did, including in scenarios involving deception or harm. A single interaction with a sycophantic model reduced participants’ willingness to take responsibility for interpersonal conflict and to repair it, while increasing both their trust in the model and their desire to continue using it [86]. The authors’ earlier ELEPHANT work had characterized this form of “social sycophancy” across the same models [87]. We therefore flag sycophancy as a related behavioral effect, not as a sixth signal of cognitive decline. Its dependent variable is moral and social conduct, rather than unaided reasoning, working memory, attention, and writing competence, which are the focus of our hypothesis and falsifiability tests. The underlying mechanism, however, is familiar from social media: systems optimized for engagement tend to select outputs that users prefer in the moment over outputs that serve them in the long run [88]. In that narrower behavioral sense, the cognitive analog of ultra-processed food, immediately palatable, rewarding, and potentially depleting, is already empirically visible.

6. What Worked, What Failed, and What to Do

The policy and design history of the physical transition is largely a record of what was not done in time. Several interventions did work: walkability standards, school physical education mandates, Amsterdam-style cycling infrastructure [89], and Finland’s North Karelia project, which, over thirty-five years, reduced cardiovascular risk factors at the population scale [90]. These show that the trajectory is reversible by environmental redesign, but only at substantial cost and with multi-generational lag. Mass media and individual-level encouragement, by contrast, had only limited population effects [91]. The lesson is unambiguous: when an environment is engineered against a behavior, no amount of individual exhortation will restore it. Four responses follow, summarised in Table 6.

Preserve productive difficulty, and specify which difficulty matters. The cognitive analog of the gym is not a blanket rejection of AI, but the deliberate preservation of effortful learning where effort is educationally productive. This includes assessments that test unaided competence, learning environments in which AI is sometimes not available, and tool defaults that introduce productive friction, such as delayed answers, Socratic prompts, or a required attempt before assistance is provided. But “preserving difficulty” is not self-evidently desirable, and we do not present it as such. The relevant criterion, developed in Section 4.5, is to preserve productive difficulty, the effort that builds understanding, while removing extraneous difficulty, which the 85% rule suggests may account for much of the learner’s struggle [58,59]. We also acknowledge the legitimacy cost of coercive versions of this prescription. Age-graded AI restrictions and mandated unaided assessment raise real questions of paternalism, and reasonable people will weigh autonomy against the externalities of mass deskilling differently. AI-literacy frameworks proposed for schools and universities offer a less coercive starting point: rather than banning the tool, they can teach when to use it, when to withhold it, and which forms of cognitive effort must remain undelegated [67,92].

Measure the trajectory. The body’s decline became visible only because longitudinal data on grip strength, aerobic capacity, bone density, and obesity had been collected, often serendipitously, for decades. There is no equivalent for cognition. A precondition for evidence-based intervention is population-scale, longitudinal measurement of unaided reasoning, working memory, sustained attention, and writing competence, in cohorts of children and adults, powered to detect cohort effects on the order Tomkinson detected for aerobic fitness, and frequent enough to separate acute from cohort effects of AI use [24,25]. Without it, the cognitive analog of the obesity epidemic will be diagnosed only in retrospect.

Constrain the rebound. The rebound literature finds that what reliably contains rebound is binding regulation (efficiency standards, hard caps, or pricing), with the refrigerator and the Montreal Protocol as canonical successes and unmanaged fossil fuels as the canonical failure [93]. For AI, the analogous instruments would target the substitution itself: assessment-integrity standards, age-graded restrictions in formative education (analogous to those on driving, alcohol, and screen time), and design standards for high-impact deployments that keep a human a meaningful cognitive participant. The closest existing analog to age-graded restriction is the screen-time debate, whose evidence base is genuinely contested [96], a cautionary precedent for any restriction justified by “cognitive harm.” The European Union’s AI Act, together with the newly established AI Office and AI Board, is the first attempt to build such a regulatory architecture at scale [94,97]. Its relevance here, however, is institutional rather than substantive. The Act is primarily organized around risk management, product safety, and fundamental-rights protection; it does not directly target cognitive deskilling, assessment integrity, or the preservation of unaided competence. Critics also question how effectively its governance machinery will constrain deployment in practice [95]. We therefore treat the AI Act as a precedent for regulatory capacity, not as a solution to the problem identified in this paper.

Build a training culture for the mind. The body’s fitness response did not start from zero; it drew on a pre-existing tradition of organized sport (Section 5.1). The mind has no such inheritance, so the corresponding response is not only to preserve difficulty but to build the culture and institutions that make cognitive effort a popular, visible, and lifelong pursuit, the mental counterpart of sport. Concretely, this means a “cognitive physical education” of regular, timetabled, AI-free deliberate practice in reasoning, memory, and writing, framed as training rather than examination; mind-sport competitions and leagues designed to reward general, transferable cognitive fitness rather than the narrow expertise that chess or memory contests now reward; community clubs that make cognitive training social, voluntary, and lifelong, the analog of recreational running or the local gym; visible champions and spectatorship that supply the status signal cognition currently lacks; and gamified personal metrics, linked to the measurement infrastructure above, that let people see their unaided capacities the way a runner sees a finishing time. The central tension is that the analogy can mislead. A game optimized as a target invites teaching to the game rather than transfer (Goodhart’s law); narrow mind-sports may not generalize to the fluid capacity at issue; and, unlike infrastructure, a culture cannot simply be mandated into being but must be made desirable. Yet the physical precedent shows that a mass culture of voluntary effort is achievable, and that where it exists, it does part of the work that no regulation can.

The deeper logic behind the four responses proposed here is the precautionary principle: acting under uncertainty to reduce the likelihood of serious and potentially irreversible harm. We name that logic explicitly, together with its standard objection. Precautions can become a brake on beneficial innovation, and the burden is therefore on us to show that the risk is serious enough to justify pre-emptive design and governance. That is the case this paper has tried to make.

7. Conclusion and Limitations

This paper does not claim that artificial intelligence has already produced a population-level decline in human cognition. Its claim is narrower, but more urgent: the conditions under which such a decline could occur are now visible, and they resemble a historical transition whose consequences are no longer in doubt. The modern body was not weakened because people chose weakness. It was weakened because machines removed effort, environments were rebuilt around that removal, and the freed capacity was absorbed by systems that made exertion optional, inconvenient, or unnecessary. We argue that the same mechanism, substitution, environmental redesign, and rebound, is now plausibly beginning to operate on the mind.

The central contribution is therefore not the familiar warning that AI may make people think less. It is a framework for turning that warning into a testable scientific problem. From the physical case, we derive the mechanism to be examined. From the biology of effort, we derive the stronger and more consequential prediction: if frictionless cognitive offloading systematically removes the generation, retrieval, revision, and productive struggle on which neural plasticity depends, then prolonged exposure may produce not only temporary performance loss or recoverable deskilling, but measurable atrophy in the cognitive substrate itself. That is the claim that matters most, and it is precisely the claim for which the necessary evidence does not yet exist.

The evidence reviewed here is suggestive, not conclusive. Generative AI has been present at mass scale for only two to three years, whereas durable cognitive atrophy, if it occurs, would likely require longer exposure to detect. Most of the five signals, therefore, remain limited by short time horizons, cross-sectional designs, possible reverse causation, selection effects, and uncertainty about whether existing critical-thinking instruments measure the capacities now at risk. The strongest current evidence supports offloading and early deskilling, not established atrophy. Even so, the pattern is difficult to ignore: reduced and partly persistent neural engagement after AI-assisted writing; cases in which assisted performance improves while unaided performance falls once the model is removed; growing evidence of complacency and overreliance; and the rapid shift from tools that answer questions to agents that plan, execute, and verify whole workflows. Agents intensify the problem because they remove not only the final answer but also the intermediate acts of questioning, checking, revising, and deciding through which cognition is normally exercised.

The hypothesis is refutable. If future cohorts show stable or improving unaided reasoning, working memory, sustained attention, writing competence, and domain-specific judgment as AI exposure rises, the central behavioral claim would be weakened or disconfirmed. If the early neural and educational signals fail to replicate in larger, better-controlled longitudinal studies, the warning should be revised. The stronger biological version of the hypothesis is even more demanding: if high-exposure cohorts show no exposure-related change in substrate markers predicted by effort-biology, including hippocampal structure, neurotrophic signaling, and functional connectivity, then the atrophy claim fails even if behavioral dependence is observed. Conversely, sustained changes in these markers that scale with AI exposure, after adjustment for prior ability, education, age, socioeconomic status, motivation, and baseline cognition, would make the analogy to physical disuse far harder to dismiss.

The practical implication is immediate. The absence of conclusive evidence is not reassurance; it is the central danger. The body’s decline became undeniable only after the built environment had already changed, after movement had been designed out of daily life, and after remediation required gyms, public-health campaigns, urban redesign, and medical treatment at enormous cost. We should not wait for the cognitive equivalent of that retrospective diagnosis. The necessary infrastructure is clear: longitudinal measurement of unaided cognition; educational designs that preserve productive difficulty while removing needless friction; assessment systems that distinguish human competence from human-plus-tool output; and deployment standards that keep people meaningfully involved in high-consequence cognitive work.

The purpose of such measures is not to reject AI. The benefits of artificial intelligence are real, and many forms of offloading are not only harmless but liberating. The question is which forms of effort can be safely removed and which must be preserved because they maintain the very faculties that make human judgment possible. A civilization can mechanize labor without abolishing the need for strength; it can automate cognition without abolishing the need for thought. But only if it decides, deliberately, where effort still matters.

The body taught this lesson late: an environment engineered against effort can produce disease at the population scale. The mind may now be entering its own version of that experiment. The scientific and moral task is to measure it before the baseline disappears, to design against the most dangerous forms of substitution, and to preserve the forms of difficulty without which intelligence may remain productive in appearance while becoming weaker at its source.