In this article
Efficiency has a price that never shows up on the bill: cognitive decline. Thinking less today doesn't save you from thinking tomorrow, it leaves you unable to. This is known in sport (don't use the leg, it atrophies) and ignored in cognition (don't think, nothing happens). The difference is that mental atrophy doesn't hurt at first. By the time it hurts, it has been quietly piling up for years, and reversing it costs far more than maintaining it would have.
The difference between the immobilized leg and the switched-off mind
Put someone in a leg cast for six weeks. When the cast comes off, the muscle has visibly shrunk. There's less mass. The patient feels the weakness on the first step. Climbing stairs hurts. The causal link between immobilization and loss is immediate, obvious, and above all flagged by the body through signals the patient's nervous system recognizes as such.
Now put someone to work delegating, for six months, nearly all their critical writing, argumentative comparison and complex structuring to an LLM. After six months, measure their ability to draft a three-page argument on a technical topic with no help. The skill has deteriorated. But the deterioration never warned them. No pain. No internal signal. If the person notices it at all, they pin it on tiredness, on lack of recent practice with the specific topic, on age, on a bad night's sleep, on anything before the chain of delegations stacked up over six months.
The asymmetry between the two kinds of atrophy is what's worth facing head-on, because it explains why the public debate about productive AI systematically leaves out half the ledger. The body has pain receptors that fire when an out-of-shape muscle is worked. The mind doesn't. Or it has them on another, much slower timescale, mediated by processes consciousness doesn't connect to their cause.
Why the muscle analogy isn't a metaphor
The reasonable objection to this line of argument is that the brain isn't a muscle and so the analogy is rhetorical. The objection collapses when you look at the literature.
Eleanor Maguire and her collaborators, in Navigation-related structural change in the hippocampi of taxi drivers (PNAS 97, 2000) and later work through 2006 and 2011, used structural MRI to measure the gray-matter volume of the posterior hippocampus in London cab drivers. London cabbies are a special case because they have to pass The Knowledge, an exam famous for its difficulty: memorizing some 25,000 streets and thousands of points of interest within roughly ten kilometers of Charing Cross. Only a fraction of candidates ever pass, after three or four years of study, with an extremely high dropout rate. What Maguire found is striking and worth holding onto. The cab drivers' posterior hippocampus was significantly larger than that of comparable control subjects. And the size correlated positively with the years worked as a cabbie.
The conclusion the work draws is that the posterior hippocampus houses part of the spatial representation of the environment and that gray matter in that region reorganizes with intensive use, showing structural neuroplasticity in the healthy adult. The ability to navigate space without external support, exercised over years, leaves a measurable anatomical trace.
The natural counterpoint is what happens when the use is transferred to an external support. There's specific literature, more recent and less consolidated, on the effect of intensive GPS use on spatial-orientation ability. The work of Dahmani and Bohbot at McGill (Scientific Reports, 2020) points in the direction physiological intuition would predict: habitual GPS use is associated with worse spatial memory when the subject has to navigate without assistance, and a later follow-up observed that the more GPS someone used, the more that memory deteriorated. This isn't clinical atrophy. It's measurable functional decline, with a pattern the authors describe as dose-dependent.
The muscle analogy, then, isn't free metaphor. It's an approximate description of a real neurobiological phenomenon: the adult brain exhibits bidirectional structural plasticity, and the networks underpinning specific cognitive functions strengthen with use and weaken with disuse. The difference from muscle is that the signaling of weakening is far subtler, and the subject's causal attribution is practically impossible without external measurement.
What goes first
The empirical question that follows is which cognitive functions are most sensitive to mass delegation. The MIT Media Lab study Your Brain on ChatGPT (2025) and Gerlich's work in Societies (2025) point to a consistent pattern, though the literature is young and the concrete figures are the first estimates of a field under construction.
The first thing to weaken seems to be the ability to hold a complex idea in mind for a long time while working it through. Working memory applied to argumentative tasks gets trained by writing, comparing, revising. When the LLM produces the first draft and the user "just reviews," working memory isn't exercised in building the argument. It's exercised in surface correction, which is a different and less demanding operation.
The second is the ability to write under pressure without help. Anyone who has spent a year delegating the first draft to the model feels, when sitting in front of a blank page, more resistance than they used to. It isn't creative block in the romantic sense. It's lack of recent practice with the specific cognitive muscle of going from thought to text, with all the micro-decisions that transition involves.
The third is critical thinking applied to the source. When the user receives the model's output as a starting point, their spontaneous disposition is to adjust and refine, not to contradict from scratch. The ability to spot a conceptual flaw in the first draft —this one, your own— requires having gone through the phase of building it and seeing where it doesn't fit. If that phase wasn't done, the ability to detect doesn't get exercised.
The fourth, and here the data are still preliminary, is the ability to tolerate ambiguity without premature closure. AI closes fast. It gives you an answer, a proposal, a structure, a decision. The human who loses the practice of staying in ambiguity until the problem is better understood starts seeking closure before comprehension has matured. It's deterioration of metacognition, not of instrumental cognition, which is why it's especially hard to measure.
Pedagogy as a silent victim
The uncomfortable institutional question arrives when you turn your gaze to the education system. Modern pedagogy, since the nineties, is built on the deliberate practice of processes: writing, calculating, comparing, justifying, defending. The acquisition of competence rests, in the terms of Sweller, van Merriënboer and Paas (Educational Psychology Review, 2019), on intentional cognitive load the student takes on while executing tasks the teacher designs to force it.
When the student can, outside the classroom, delegate most of that cognitive load to an LLM, pedagogy loses its founding assumption. The homework gets done, gets done fast, and the expected cognitive trace of the task doesn't form. The teacher reads the submitted work and, for the most part, can't tell the assisted writing from the unassisted. The grade rewards the output. The process, which was the thing meant to be taught, didn't happen.
The institutional response has been confused. Some universities banned LLMs for a term and then reversed course. Others have built the use into explicit pedagogy, with uneven results. The few that have tried to force the process —in-person exams without help, oral defenses, supervised writing— face operating costs that make it hard to scale.
Carr had anticipated this in The Shallows (2010) on digital reading. Postman, back in Technopoly (1992), had warned that no technology is neutral toward the human capacities it displaces. Newport, in Deep Work (2016), had codified in operational language what gets lost when deep cognition is fragmented. The three warnings, each at its own scale, converge here.
The problem isn't the teacher's. It's structural. Education is built on intentional cognitive friction. The available technology removes the friction by default, and the student, rationally, removes it when they can. The consequence, a decade out, is a cohort that has passed through the classrooms with the usual grades and without having exercised the cognitive functions the system assumed were being exercised. The metric doesn't detect the deficit. The labor market will detect it five years later, with the first promotions that don't come.
The deferred bill
There's a piece of accounting worth laying out without rhetoric. The balance sheet of productive AI in 2026 has three columns that get counted and one that doesn't.
The time saved gets counted. Cui and others (2025) measured a 26% increase in tasks completed by developers with Copilot assistance in field experiments with thousands of programmers. Consultancies have been publishing estimates of weekly hours freed up per office worker for years. The aggregate figure is real and the column is well kept.
The speed of output gets counted. Companies that have adopted generative AI produce more content, more emails, more reports, more code per unit of time. The charts go up.
User satisfaction gets counted. Surveys of employees who use AI report, by a majority, that they feel more productive, less stressed by routine tasks, freer for higher-order work.
The fourth column that doesn't appear
What doesn't get counted, or gets counted badly, is the accumulated cognitive cost. The fourth column appears on no balance sheet because its unit of measure isn't standardized, its evidence is recent, and its effect will show up on a timescale annual reports don't cover. The column exists all the same. The MIT EEGs, Gerlich's critical-thinking scores, Maguire's hippocampal volumes —the converging evidence points to that column being neither zero nor small.
The uncomfortable part of the situation is that the missing column won't appear until it's too late to correct it in aggregate. A cohort of young professionals who entered the market in 2023-2025 is learning to work with assistance from day one. The ability to work without assistance isn't being built. Ten years out, when these professionals are the seniors who have to make complex decisions under constraint, the missing skill will be visible, and individual recovery, possible but costly and uneven.
The question the ledger demands
There's a question worth posing calmly, to avoid alarmism. Is reversing this desirable? And if so, how?
The honest answer is that the first half of the question has no unanimous answer. There are reasonable positions in favor of accepting the displacement —the calculator displaced part of mental arithmetic and nobody wants to go back— and reasonable positions against —the calculator didn't displace thinking, only arithmetic; the LLM does displace thinking—. The answer will depend on what we consider an irreplaceable part of human mental exercise and what we consider delegable with no qualitative loss.
The second half of the question —how it's reversed— has an operational answer, and it's worth naming even if it's hard. It's reversed by introducing deliberate cognitive friction into the flow. On the individual level: stretches of voluntarily reduced use, tasks done by hand with awareness of the cost, reviewing your own texts without help. On the institutional level: pedagogy with AI-protected spaces during the process-building phases. On the legal and regulatory level, whatever can be done, though regulatory experience suggests that correction by rule always arrives late.
Thinking less today isn't free. The bill comes. Whoever pays it won't necessarily be the one who made the choice.
Definitions
Cognitive atrophy. Functional loss of mental capacities through disuse, with a neurobiological basis in the bidirectional plasticity of the adult brain. Distinct from muscular atrophy in that it lacks acute signaling and therefore isn't perceived by the subject as it accumulates.
Structural neuroplasticity. The healthy adult brain's capacity to reorganize the density and connectivity of specific regions in response to sustained use. Documented anatomically in London cab drivers by Maguire and others (2000-2011).
Germane (intentional) cognitive load. A concept from Sweller's Cognitive Load Theory. The processing effort the student invests in building mental schemas during the active execution of a task. Delegating it prevents consolidation.
The Knowledge. The British exam that London cab candidates must pass after several years of study, memorizing some 25,000 streets and thousands of points of interest, with a very high dropout rate. A paradigmatic case in the neuroplasticity literature for its demand of unassisted spatial navigation.
Critical thinking. The ability to evaluate claims, detect argumentative flaws, compare alternatives and suspend judgment until there's sufficient evidence. A cognitively costly operation whose delegation to an external system limits the consolidation of the skill.
References
Carr, N. (2010). The Shallows. What the Internet Is Doing to Our Brains. W. W. Norton. An early analysis of the cognitive impact of the digital medium on reading and deep attention.
Cui, Z. et al. (2025). The Effects of Generative AI on High-Skilled Work. Evidence from Three Field Experiments with Software Developers. Field experiments with thousands of programmers at Microsoft, Accenture and a Fortune 100 company; an average 26% increase in tasks completed with access to GitHub Copilot.
Dahmani, L. & Bohbot, V. D. (2020). Habitual use of GPS negatively impacts spatial memory during self-guided navigation. Scientific Reports 10, 6310. Association between intensive GPS use and worse hippocampus-dependent spatial memory in unassisted navigation.
Gerlich, M. (2025). AI Tools in Society. Impacts on Cognitive Offloading and the Future of Critical Thinking. Societies 15(1), art. 6. Negative correlation between intensive use of AI tools and performance on standardized critical-thinking tests.
Maguire, E. A. et al. (2000). Navigation-related structural change in the hippocampi of taxi drivers. Proceedings of the National Academy of Sciences 97(8), 4398–4403. Structural demonstration of neuroplasticity in the healthy adult as a function of intensive spatial-navigation use.
Maguire, E. A., Woollett, K. & Spiers, H. J. (2006). London taxi drivers and bus drivers. A structural MRI and neuropsychological analysis. Hippocampus 16(12), 1091–1101. Replication and refinement of the finding, comparing with bus drivers on fixed routes.
MIT Media Lab (Kosmyna, N. et al.) (2025). Your Brain on ChatGPT. Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. arXiv: 2506.08872. EEG data on reduced neural connectivity in the LLM group.
Newport, C. (2016). Deep Work. Rules for Focused Success in a Distracted World. Grand Central. An operational framework on cultivating deep cognitive functions against fragmentation.
Postman, N. (1992). Technopoly. The Surrender of Culture to Technology. Alfred A. Knopf. A structural critique of the cultural assumption that all technology is neutral toward the human capacities it displaces.
Sweller, J., van Merriënboer, J. J. G. & Paas, F. (2019). Cognitive Architecture and Instructional Design. 20 Years Later. Educational Psychology Review 31, 261–292. An updated synthesis of Cognitive Load Theory applied to pedagogy and cognitive delegation.
You might also like
- Cognitive delegation
- The future of human learning
- The extinction of human skills
- The cost of not thinking
Comments0
No comments yet.
Leave a comment