Will We Still Be Smarter Than Chatbots by 2030?

This did not start with ChatGPT

The most important thing to understand about generative-AI-driven cognitive decline is that generative AI did not invent the mechanism. It is the latest and deepest move in a long sequence: layers of tools that hand different parts of the mind to machines, each reaching a little further inward than the one before. They also arrived in a clear order. You can date each layer by the device that made it ordinary. So let’s walk through them in the order they entered daily life.

Layer one: the hand, displaced by the keyboard (the home computer of the 1990s, then a laptop in every lecture hall through the 2000s). It starts with the fingers. For most of human history, getting a thought onto a page meant forming each letter by hand, and that physical act was quietly doing cognitive work. Pam Mueller and Daniel Oppenheimer’s longhand-versus-laptop study found that typists recorded far more words while doing worse on conceptual questions, apparently because handwriting forced students to compress and reframe ideas in their own words. That specific behavioral result has a notable failed replication, so the cleaner claim is narrower: handwriting and typing do not recruit the same cognitive machinery. High-density EEG work from Audrey van der Meer’s group finds richer connectivity during handwriting than during typewriting, especially in networks associated with memory formation and encoding. fMRI work similarly suggests that forming letters by hand recruits visual-word and spatial-processing areas that key-pressing can skip. The hand was never just an output device; it helped the mind etch language into itself. The keyboard, the touchscreen, autocomplete, and now a generative-AI assistant that writes the whole sentence have been removing that physical encoding one stroke at a time. 54 55 56 57

Layer two: navigation, displaced by GPS (the dashboard satnav of the mid-2000s, then Google Maps in every pocket after 2008). The cleanest natural experiment we have is the satnav. In a 2020 Scientific Reports study, researchers tracked fifty drivers and found that people with greater lifetime GPS use had measurably worse spatial memory when forced to navigate on their own. The follow-up was especially important: heavier GPS use predicted subsequent decline, and the people who used GPS most did not appear to be doing so simply because they already had a poor sense of direction. The evidence does not prove every act of GPS use causes atrophy, but it makes the compensation story much harder to maintain. 1

The neuroscience is specific. Finding your own way engages the hippocampus, which builds a true cognitive map by comparing routes, predicting turns, and holding the city in your head. Following turn-by-turn directions leans more heavily on stimulus-response execution: just do the next thing the voice says. fMRI work has found lower hippocampal engagement when people follow externally supplied directions than when they actively navigate. Because these systems can compete, habitual reliance on execution can crowd out mapping. 2

The reverse is also possible, which is what makes this a choice rather than a fate. Eleanor Maguire’s famous London taxi-driver studies showed that drivers who spent years memorizing 25,000 streets for “The Knowledge” developed measurably larger posterior hippocampi, with size tracking years on the job. The brain is plastic. Exercise the function and it can grow; stop loading it and the gains can recede. The hippocampus is not literally a muscle, but the metaphor is close enough to be useful. GPS is the elevator we take instead of the stairs. 3 4

Layer three: attention, captured by social media (the smartphone arrived in 2007; the feed was inescapable by the early 2010s). While the satnav was quietly retiring our sense of direction, social media was retraining something underneath all thinking: what our brains find rewarding.

Social platforms exploit reward-learning dynamics, including unpredictable reinforcement. Neuroscientist Kent Berridge’s work established the crucial distinction: dopamine is closely tied to wanting, not simply liking. The spike comes in anticipation, before the reward, not only during it. That helps explain why scrolling can feel compulsive without feeling satisfying, and why checking the phone becomes a reflex rather than a decision. 5 6

The cognitive cost is not mainly the time. It is the appetite shift. A brain conditioned to expect rapid, unpredictable reward signals finds long, effortful, low-stimulation tasks increasingly intolerable: reading a dense argument, sitting with a hard problem, waiting for understanding to form. Social media does not need to make anyone directly worse at reasoning. It can do something more strategic by lowering tolerance for the slow state in which reasoning happens. It prepares the ground.

Layer four: depth, flattened by the screen (by the mid-2010s, the phone and tablet were where most reading happened). Here is a good place to run a small experiment on yourself, because the numbers are unusually concrete. If you have read this far, actually read it line by line rather than skimming for the bold bits or pasting the essay into a chatbot for the gist, you belong to a genuinely small group. Analytics firms that study scrolling behavior have long found brutal drop-off: a large share of visitors leave within seconds, the median reader does not reach the end, and only a minority finish long articles. The new layer is AI-generated search summaries. Pew found in 2025 that 65% of U.S. adults at least sometimes encounter AI summaries in search results, and a separate Pew browsing analysis found that when an AI summary appears, users click traditional results in 8% of visits versus 15% when no summary appears. The exact “top 8%” figure is therefore a back-of-the-envelope estimate, not a measured statistic. Still, the direction is clear: the slice of people who will carry a dense, multi-thousand-word argument all the way to the end, without skimming and without outsourcing the reading, is small and probably shrinking. If you are still here, that is the faculty this section is about. 66 67 68

Cognitive scientist Maryanne Wolf has spent her career on what this does to the brain. Reading, she stresses, is not natural: learning to read physically builds a new circuit, and what and how we read reshapes it. "Deep reading", slow, immersive, effortful, is the mode in which we connect new information to what we already know, draw inferences, weigh an argument's truth, and pass over into another person's perspective. When we skim, Wolf writes, we literally don't have time to think, or to feel. Read only on skim-friendly screens for long enough and the circuit grows less elaborate: comprehension drops even as the sense of speed rises, and in children the critical-thinking and empathy processes that deep reading is supposed to build never fully form. The screen did not just change the page. It changed how far below the surface of the words we are willing to go, and a large language model that reads the article for you, and hands back three bullet points, is the frictionless bottom of that slope. 58 59 60 61

Layer five: reasoning itself, outsourced to the large language model (ChatGPT, late 2022, one of the fastest-adopted consumer technologies ever measured). Then, into a population whose hands had left the page, whose sense of direction had thinned, whose attention had been retrained, and whose reading had flattened into skimming, arrived a tool that offered to do the one thing none of the others could: the thinking.

Each layer reaches deeper than the last. The keyboard took the hand out of encoding. GPS outsourced where. Social platforms rewired what we crave. Screens dissolved how deeply we read. And now the large language model reaches for the last thing left, the how of thought itself: the derivation, the synthesis, the working-through that was the irreducible core of intellectual labor. This is the leap. And the evidence that it carries the same atrophy signature as every layer before it is arriving fast.

Why this time is different in kind

It is tempting to file generative AI under “just the next tool,” another entry in the long lineage of writing, calculators, and search engines that humans have always used to lighten the mind’s load. That lineage is real, and much of it was benign. The important distinction is that generative AI changes the kind of work being offloaded.

Earlier cognitive tools usually offloaded the periphery of thinking while leaving the core intact. A calculator does the arithmetic, while you still decide what to compute and why. A search engine, even Google Scholar, retrieves the literature, while you still have to read it, weigh it, reconcile contradictions, and build the argument yourself. The scholar who once walked the stacks library by library was offloading the legwork of finding sources. The scholarship itself, meaning the synthesis, judgment, and chain of reasoning, stayed theirs. Every act of doing it kept the muscle under load. Search changed where the information lived. It did not change who did the thinking.

Large language models break precisely that boundary. As cognitive scientists have begun to point out, generative AI is qualitatively different from everything before it: it offloads not retrieval but ideation, structuring, reasoning, and synthesis, the generative acts that were the thinking. When a tool turns a vague human prompt directly into polished, structured, reasoned prose, it absorbs the exact sequence that builds understanding: converting confusion into a representation, a representation into a procedure, a procedure into transferable skill. That friction was never a bug in the old way of working. It was the mechanism of learning itself, and LLM interfaces are engineered to remove it. 16 39 40

So the comparison to search is a category error. Offloading memory or retrieval touches a single, peripheral faculty. Offloading reasoning touches the integrative core, the general-purpose machinery of logic, derivation, decision, and mental computation on which every other intellectual capacity is built. We are not automating one more narrow sub-task. We are, for the first time, automating the part of cognition that was supposed to stay ours. That is exactly what makes the damage so hard to undo: you can regrow a specialized skill, but the thing being hollowed out here is the foundation the specialized skills stand on.

The "just look it up" fallacy

There is a comforting story we tell about all of this, and demolishing it is the single most important thing the science can do, because almost everyone believes it, and it is wrong. The story goes: facts are cheap now. Memorizing them is obsolete busywork. Offload the raw information to Google or to an LLM, and you free the mind for the good stuff: analysis, creativity, judgment. Why clutter your head with what you can look up in two seconds?

Cognitive science has a one-word answer: because that is not how thinking works. Knowledge held in long-term memory is not the rival of reasoning. It is the substrate reasoning runs on, and you cannot offload it without starving the very faculty you were trying to liberate.

Daniel Willingham, distilling decades of the field, puts it flatly: critical thinking is bound to background knowledge, and memory is the mind's tool of first resort, faced with a problem, the mind searches memory for a solution before it reasons from scratch. The deep mechanism is chunking. Knowledge already in your head lets you compress new information into meaningful units, which frees up the tiny, precious space of working memory to actually manipulate ideas rather than just hold them. This is why experts crush novices: studies of chess masters, even under blitz controls that leave almost no time to reason, show their edge comes from fast pattern recognition, not faster logic. They hold thousands of chunked positions in memory; the novice, grinding through first principles in real time, loses anyway. Expertise is stored knowledge, organized. 62 64

Strip the knowledge out and the higher faculties do not get liberated; they buckle. Reading comprehension itself, it turns out, depends far more on how much you already know about a topic than on any general "comprehension skill." E.D. Hirsch's warning lands precisely on the generative-AI moment: without knowledge already in your head, you cannot evaluate a source, judge whether a claim is plausible, or even know which question to ask, and a novice, by definition, doesn't know what they don't know. The classic demonstration is almost comic: children handed a dictionary and told to use new words produced nonsense sentences, because looking a word up is not the same as knowing it. Access is not possession. 63 65

Now see why this is the hidden bottom of the entire argument. Every earlier layer offloaded a process of thought, the navigating, the reading, the reasoning. The faith that you need not know anything because you can always look it up is more corrosive still, because it removes the raw material those processes run on. And note the escalation: a search engine at least made you read the result, so some knowledge still trickled in; asking an LLM to use the knowledge for you removes even that trickle. The popular defense, "let the machine hold the facts so humans do the higher-order thinking", rests on a premise cognitive scientists have spent forty years dismantling: that knowledge and thinking are separable. They are not. You think with what you know. A mind that has shipped its knowledge off to a model has not been freed to reason at a higher level; it has been quietly emptied of the very thing reasoning is made from. Offloading does not just empty the warehouse; it convinces you the warehouse is still full. Here is the deeper cut: the warehouse was never separate from the factory. It was the factory's raw material, and we are shipping it offsite. 14 15 62 63

The evidence is no longer speculative

For a while, the worry about generative AI and thinking was mostly a generalized unease. It is now becoming data.

Start with the closest thing we have to watching it happen in the brain. In 2025, a team at the MIT Media Lab (Kosmyna and colleagues) put 54 people in three groups to write essays: one using ChatGPT, one using a search engine, and one using only their own heads, while wearing EEG caps. The brain-only writers showed the strongest, most distributed neural connectivity. Search users showed less. The LLM group showed the weakest connectivity of all; cognitive engagement scaled down as the tool took on more of the work. Two findings cut deepest. First, the disengagement appeared to persist: when some LLM users were later asked to write unaided, they remained under-engaged. Second, ownership weakened. LLM users struggled to quote the essays they had just “written,” because in a real sense they had not encoded them. The authors called the result cognitive debt. A published methodological comment raises real cautions about sample size, EEG interpretation, and transparency, so this study should be treated as a strong signal rather than a verdict. 7 8

The pattern also appears where skill can be measured directly. In a 2026 randomized controlled trial, Anthropic’s own researchers had 52 mostly junior engineers learn an unfamiliar Python library, half with an AI coding assistant and half without. The assistant group finished about two minutes faster, a difference that was not statistically significant. On a quiz testing whether they had actually learned the material, the assistant group averaged 50% against the manual group’s 67%, with the largest gap in debugging. They had shipped the task and skipped part of the learning. You trade competence for a speed boost that often is not even there. 9

The education literature is converging on the same shape. A global survey found 86% of university students now use AI in their studies, mostly meaning generative tools in this context. Controlled experiments show that unguided generative-AI use produces cognitive offloading without any improvement in reasoning quality, while only carefully structured use avoids it. Longitudinal analysis of real student prompts shows the natural drift: left alone, students converge toward low-effort, minimally engaged requests, asking for the answer, not the method. Researchers describe the risk as a loss of "epistemic agency," the capacity to evaluate, justify, and take ownership of one's own knowledge. 10 11 12 13

And underneath it sits the oldest finding of the digital age, which now reads like a prophecy. In 2011, Betsy Sparrow's "Google effect" study in Science showed that when people expect to be able to look something up later, they remember less of the information itself and more about where to find it. We shifted from storing knowledge to storing pointers. The genuinely dangerous part is the sequel: people consistently mistake the internet's knowledge for their own. Offloading does not just empty the warehouse. It convinces you the warehouse is still full. Extend that from facts to reasoning, from "I feel like I know this" to "I feel like I could work this out", and you have a population losing capability while feeling more capable than ever. 14 15

This is the white-collar story now unfolding in real time. The intellectual middle of knowledge work, drafting, summarizing, structuring a first-pass argument, debugging within a bounded code context, synthesizing a report, is exactly the layer where LLMs can produce competent-looking work often enough to make offloading irresistible. Every one of those acts was, until recently, a rep at the cognitive gym. We are skipping the reps en masse.

The decline you cannot see

Now comes the deepest problem of all, and the reason the others can run unchecked. A society losing its physical fitness can watch it happen. A society losing its cognitive fitness may not be able to, even in principle, because the instrument you would use to detect the loss is the very faculty being lost.

Return to the gym one last time. Physical decline broadcasts itself: the mirror, the number on the bar, the breathlessness at the top of the stairs, where you finish in the race. The feedback is external, objective, and indifferent to your opinion of yourself. Your body is not the thing judging your body. Cognitive decline has no mirror. The only instrument capable of evaluating the quality of your thinking is your thinking, and as it degrades, so does its ability to register that anything has degraded. The judge and the defendant are the same faculty, declining together.

This is not armchair speculation; it is a robust metacognitive finding. In 1999, Kruger and Dunning identified what they called a “double curse”: the skills you need to perform well in a domain are often the same skills you need to recognize whether you are performing well. Incompetence does not merely produce errors. It also weakens the metacognitive capacity to detect those errors. Later researchers sharpened the point: metacognition is not a free-floating talent. It requires a mental model of what good performance looks like before you can measure your own against it. As competence erodes, the yardstick erodes with it. You do not just get worse; you lose part of the instrument that would have told you so. 41 42

Now drop a large language model into that loop, and the trap tightens. Aalto-affiliated researchers found in two large-scale studies that ChatGPT use improved task performance while weakening calibration: participants overestimated how well they had done, and higher self-rated AI literacy correlated with lower metacognitive accuracy. The familiar Dunning-Kruger pattern did not simply persist under LLM use; in their model, it ceased to exist. The tool sold as making us smarter can also mute the alarm that tells us when we are not. 43 45

The mechanism has been isolated, and it is mundane and damning. Researchers studying how explanation style shapes judgment found that longer, more fluent LLM explanations reliably raise a user's confidence without improving their ability to tell correct answers from wrong ones. Verbosity masquerades as authority. Worse, people cannot easily distinguish between having understood an explanation and having merely read a fluent one, so they quietly revise their sense of their own understanding upward on the strength of prose the machine wrote. It is the illusion of explanatory depth with a turbocharger: the felt click of insight, fully detached from the having of it. 44 45

Cognitive scientists now have a term for the result: metacognitive decoupling, a widening gap between four things that are supposed to track together, the output you produce, the understanding you actually hold, how calibrated your self-assessment is, and how skilled you believe you are. Fluent answers strip out the "difficulty signals" that normally trigger self-monitoring; with no felt struggle, nothing flags that anything needs checking. And because the machine's own confidence is itself badly calibrated, stating wrong answers in the same assured tone as right ones, a miscalibrated human meets a miscalibrated model and the two compound instead of correcting each other. 45 46

Hold this against the spine of the whole argument. Physical fitness is legible; there is always a mirror. Cognitive fitness is not, and LLMs fog the one mirror we had: the friction of trying to reason something through and feeling ourselves come up short. Remove the friction and you remove the feedback. A population can slide a very long way down this slope while every individual on it feels, with sincere and growing confidence, that they have never been sharper. That is what makes this decline uniquely dangerous. It is not merely unmonitored; it is anti-monitored. The loss quietly deletes the evidence of itself.

Why your brain is doing this on purpose

Here is the part that should genuinely unsettle you: none of this is a failure of the brain. It is the brain working exactly as designed. Offloading isn't a bug in human cognition. It is close to its governing principle.

Cognitive scientists Evan Risko and Sam Gilbert formalized “cognitive offloading” as a metacognitive cost-benefit decision: the mind constantly weighs doing a task internally against pushing it onto an external tool, and picks whichever is cheaper for the expected payoff. Their conclusion about the consequence is blunt: when a tool reliably performs a function, the brain, ever economical, reduces its investment in that function. It is disuse atrophy, and it is automatic. 16 17

The organ that runs this accounting is the dorsal anterior cingulate cortex. The dominant model, Shenhav, Botvinick, and Cohen's Expected Value of Control, holds that this region computes a running cost-benefit analysis: the payoff of thinking hard, the amount of mental control required, and the cost of that control in effort, then decides whether the effort is worth it. The companion "opportunity cost" account explains why hard thinking feels actively unpleasant: the aversiveness is the price tag. Effort hurts because spending a cognitive resource means not spending it elsewhere, and the brain wants you to know it. 18 19 20 21

How much effort you will tolerate is set, chemically, by dopamine. John Salamone’s animal work showed that depleting dopamine in the nucleus accumbens makes a creature abandon high-effort, high-reward options for low-effort, low-reward ones. Westbrook and Braver extended this to mental effort: striatal dopamine helps set your willingness to think hard by tuning how you weigh the benefits against the costs. Dopamine is the throttle on cognitive labor. 22 23

At the network level, this economy has a concrete geometry. Stroud and colleagues asked what recurrent neural networks do when they must solve cognitive tasks under two biological constraints: neural noise and metabolic cost. The answer was not simply "represent everything and think harder." Under cost, the networks changed the geometry of their internal representations. They kept task-relevant variables available while suppressing dynamically irrelevant stimuli through activity-silent, subthreshold dynamics. Recordings from primate prefrontal cortex during learning moved in the same direction, toward a minimal representational strategy. That is the mechanism missing from the loose "energy minimization" story: deliberate control is expensive because the circuit has to preserve the right information while preventing irrelevant dimensions from consuming activity and causing interference. 69

Ali and colleagues make the same point from another direction. When they trained recurrent networks to minimize energy consumption in predictive environments, predictive coding emerged without being hardwired: the networks self-organized into prediction units and error units and learned to inhibit predictable input. Zénon, Solopchuk, and Pezzulo give the information-theoretic bridge. A cognitive task is costly to the extent that it forces the system to update its internal model; unfamiliar and dual tasks are expensive because they require larger informational changes. Musslick and Cohen add the architectural bottleneck: shared representations are efficient for learning and generalization, but they create interference when several control-dependent processes run at once, so the system must gate, separate, or serialize them. 70 71 72

This makes the "path of least resistance" mechanical rather than moral. Thinking hard is not merely choosing to burn more willpower. It is asking a noisy, metabolically constrained, interference-prone network to hold a precise task geometry in place, update its internal model, suppress irrelevant channels, and route control through a limited architecture. Cognitive scientists call the resulting behavior resource-rational: given that internal computation is genuinely costly, offloading is often the correct local answer, not a weak one. 32 33 69 70 71 72

That is why LLM offloading can behave less like a preference than like a gradient. The resource-rational optimum depends entirely on the cost of each option. Generative AI collapses the cost, latency, and effort of the external route to nearly zero, for almost every task, for almost everyone. The optimization landscape has been redrawn so that offloading dominates. The brain is not being seduced into a mistake. It is solving the wrong local problem: minimizing effort now, blind to the skill it is liquidating for later.

The gym objection

Here is the strongest objection, and it deserves a real answer. Physical labor was automated a century ago; we have cars, elevators, dishwashers, forklifts. And yet people voluntarily go to the gym. They lift heavy things for no reason but to stay strong. So why won't the same thing happen for the mind? Won't a culture of "cognitive fitness" simply emerge, with people doing their mental reps the way they do their squats? Isn't this just like chess, a hard thing people choose to grind because the difficulty is the point?

It is a comforting argument, and it is wrong, for four reasons that map directly onto the neuroscience above.

First, the reward signal is structurally different, and LLMs sever the one that matters. The brain does have an intrinsic reward for hard thinking. It's called curiosity, and it runs on the same dopamine circuitry as food and money. Here is the critical finding: that reward does not fire for information itself. It fires for the rate of learning, the felt sense of figuring something out. Midbrain dopamine tracks learning progress, and learning progress is what links cognitive effort to engagement. When you struggle to derive an answer, prediction errors resolve in a stream and you get paid in dopamine; the effort is rewarded. When an LLM hands you the finished answer, there is no error to resolve, no gradient to climb, and the reward never fires. The effortful path now costs effort and pays nothing. Physical exercise, by contrast, still delivers its felt payoff: the pump, the fatigue, the visible result. Generative AI uniquely removes the intrinsic reward for the activity it replaces. The gym still pays you. The cognitive gym, once an LLM is in the room, does not. 29 30 31

Second, physical fitness is visible and social; cognitive decline is invisible and self-concealing. People go to the gym partly because the results show: in the mirror, in a deadlift number, in how others see them. Strength is legible and high-status. Cognitive capability has no such signal, and worse, the loss is masked by the very illusion of knowledge that offloading creates. The person whose reasoning has atrophied feels, if anything, more capable, because the LLM keeps producing competent-looking output under their name. You cannot be motivated to repair a deficit you have been neurologically convinced you don't have. There is no cognitive mirror.

Third, the gym and chess are opt-in walled gardens. Offloading is the default flow of everything else. This is the chess point, and it cuts the other way from how the objection intends. Chess survives precisely because it is a sealed arena where the difficulty is deliberately preserved, a hobby, a sport, a protected ritual of desirable difficulty. So is the gym. They work as exceptions. Yet you cannot run your whole life as a chess problem. For the eight hours of actual knowledge work where cognition is load-bearing, the emails, the analysis, the code, the decisions, the LLM is right there, free, and faster, and the resource-rational machine will choose it every time. A few people maintaining cognitive hobbies no more preserves general human reasoning than a population of weekend powerlifters reverses a sedentary society's metabolic decline. The walled garden does not generalize.

And we already ran this experiment, with mathematics. Math is the cleanest case imaginable of a hard cognitive skill that the modern world made radically more accessible. Calculators, then spreadsheets, then Wolfram Alpha, then Khan Academy, then a free chatbot that will solve and explain any equation you can type. If accessibility produced practice, the last two decades should have created a renaissance of adults voluntarily doing math for the joy of a now-frictionless skill. Ask yourself the honest question: since ChatGPT, or since the internet, or since the graphing calculator, do you personally know a single adult who, finding math suddenly easier to reach, started sitting down to solve equations they no longer had to? In most people's experience the answer is nobody. Not one person. The accessibility went up and the voluntary practice went to zero.

The data is at least consistent with this, and it is sobering. The OECD’s PIAAC survey, the largest study of adult skills in the world, reports that while internet use among adults rose from 76% in 2012 to 93% in 2023, literacy and numeracy proficiency were stable or declining in many participating countries, especially among less-educated adults. In the United States, the share of adults at the lowest numeracy level rose from 29% in 2017 to 34% in 2023. The German longitudinal arm of PIAAC sharpens the interpretation: cognitive skills declined with age mainly among adults with below-average skill usage, while high-usage white-collar workers continued gaining into midlife. Maintenance is not a function of access alone. It depends on whether daily life still forces the reps. 36 37 38

That is the whole argument in miniature. We made math easier to access, and broad adult practice did not automatically follow once school stopped compelling it. There is little reason to expect reasoning, writing, analysis, and judgment to behave differently now that an LLM will do those on demand too. The gym is the rare opt-in exception. Math is the rule. And the rule is abandonment.

Fourth, and this is the heart of it, hard thinking is not wired into most people's daily survival, and never really was. Going to the gym is a recent, effortful cultural workaround for a problem evolution didn't design us to have. It requires intrinsic drive, social scaffolding, identity, and leisure, and even with all of that, most people don't sustain it. The grim base rate of gym memberships is the real analogy: most are bought and abandoned. Now ask people to do the harder, less visible, less rewarded, more easily skipped version, for their minds, against a tool engineered to make skipping frictionless. The gym example is not the optimistic counterargument. It is the warning. Most people don't go.

The plasticity objection

The last refuge of optimism is plasticity. The brain bounces back; skills return with practice; look at the taxi drivers whose hippocampi regrew. Decline, the argument goes, is a thermostat that resets the moment we start thinking again. There is real truth in this, but three findings turn the consolation into something far more fragile.

First, broad capacity is hard to rebuild with narrow exercise. This is one of the most robust and least appreciated results in the science of cognitive training. Across enormous samples, including Owen and colleagues’ eleven-thousand-person Nature study, and a long line of meta-analyses since, the pattern is stubborn: training a mental task makes you better at that task and ones closely resembling it (near transfer), while improvements usually do not generalize far beyond the trained domain (far transfer). A second-order meta-analysis pooling working-memory training, brain-training games, music, chess, and video games found that once placebo effects and publication bias are addressed, transfer to general cognitive ability is statistically indistinguishable from zero. The practical lesson is not that training never matters. It is that trained skills tend to stay welded to the domain in which they were trained. 47 48 49 50

See what this does to both the gym analogy and the hopeful intuition that “some games make you smarter.” Specific regimens can sharpen specific faculties. Certain action games, for example, can improve narrow visual-attention processes. That is exactly the point: the enhancement is narrow, and broad transfer is limited. The corollary matters. If narrow practice has weak evidence for building broad reasoning, it is also a weak candidate for rebuilding broad reasoning after disuse. And what LLMs threaten to erode is not a single narrow faculty. It is the integrative core: derivation, synthesis, judgment, and mental ownership. A Sudoku streak, crossword habit, or brain-training app may help a local skill. The concern here is general capacity, and local repairs do not automatically reach it.

Second, the reversibility evidence is mostly short-term, and mostly about the wrong kind of cognition. The encouraging studies typically pull a tool away for days or weeks and watch a single isolated function, recall of a word list, a memory-task score, recover. None of that tells us what happens to foundational reasoning after years of continuous outsourcing across an education or a career. We are extrapolating from a sprint to a marathon nobody has yet run. 53

Third, and this should end the complacency, for the most exposed people, the problem may not be atrophy at all. Researchers have begun drawing a hard line between atrophy and foreclosure. Atrophy is the weakening of a skill you once built; the foundation exists, so recovery is at least possible. Foreclosure is never building the skill in the first place. The distinction breaks sharply along age, because prefrontal systems involved in judgment, planning, and higher-order reasoning remain unusually plastic and still under construction into the early twenties, making adolescence and early adulthood the sensitive window in which these capacities get wired. A forty-five-year-old leaning on generative AI is letting a built muscle soften. A fifteen-year-old who has never reasoned without an LLM is not softening a muscle; they are skipping the developmental step in which the muscle was supposed to form. You cannot atrophy what was never built, and you cannot "recover" a capacity you never had. One early survey of 666 people illustrates the fault line: younger participants showed higher dependence on AI tools, mostly generative tools in this context, and lower critical-thinking scores than older participants. That is correlational evidence, not proof of developmental foreclosure, but it is exactly the pattern the foreclosure hypothesis would predict. The generation being told that fluency with generative AI is the essential modern skill may be the one quietly foreclosing the deeper skill underneath it. 39 51 52

Put the three together and the resilience argument narrows. Yes, the brain is plastic, but broad reasoning is built through years of effortful, general use, often inside sensitive developmental windows. The transfer literature gives little reason to think it can be restored cheaply with targeted drills once it has thinned or was never laid down. Reversibility is real for some people, for some functions, on some timescales. It is not a general guarantee, and it is weakest exactly where the exposure is heaviest.

The crossing is a choice, not a friction-free one

So return to the graph and its two changing lines. Model capability is improving quickly. The threshold we should watch is not a fixed altitude; it is a relationship between what machines can do and what humans still practice unaided. The familiar debate watches the model curve. This essay has followed the human one.

That second curve is not a metaphor. It is made of habits that become easy to miss because each one is individually rational. GPS takes the map out of the hippocampus. Feeds train the reward system away from slow effort. Screens and summaries thin the deep-reading circuit. Search and AI summaries turn knowledge into pointers. LLMs move from retrieval into derivation, synthesis, and judgment. Then the metacognitive problem locks the loop: fluent output makes the user feel more capable even as the work that would have built capability has been skipped.

This is why generative AI changes the practical comparison. It improves the model while also changing the human baseline. The comparison point is not the best human mind, fully exercised and richly stocked with knowledge. In most markets, classrooms, workplaces, and institutions, the comparison point is the human who has spent years letting a system draft the argument, solve the bug, compress the reading, choose the route, supply the fact, and polish the prose. Practical parity arrives when the model can replace that human in the actual loop, not when it has matched the strongest version of human cognition in principle.

The hopeful fact is still plasticity, but hope has to be specified. Taxi drivers show that sustained, high-stakes use can build and rebuild a concrete cognitive map. Transfer research shows that narrow drills rarely restore broad reasoning. Learning science shows that durable skill comes from desirable difficulty, retrieval, error correction, and ownership. The dopamine literature explains why that regime will feel uphill once generative AI has made the shortcut almost free. So the escape hatch is real, but it is not ambient. It is a deliberately engineered practice: using LLMs as tutors, critics, examiners, and adversaries, not as ghostwriters; asking them to preserve the derivation instead of replacing it; doing the first pass from memory; checking sources yourself; letting the struggle remain long enough for the learning-progress reward to fire.

This is also where the architectural point matters. Avah Banerjee has framed the distinction as one between extraction and embodied dynamics: current LLMs are extraordinary extractors of patterns from already-captured traces, not physical systems living through the world they model. They can spoof thinking because language contains the residue of many human thoughts. But the physical world is not a text corpus, and human reasoning is not merely next-token continuation. A brain is a recurrent, living control system coupled to eyes, hands, metabolism, social feedback, fear, motion, and time. It has to simulate possible futures while acting inside the world, revising itself as the world pushes back. Mathematically, that is a different kind of computation from a model that compresses past trajectories and samples plausible continuations from them. The model can imitate the trace of thought, but matching the trace is not matching the engine. It is not paying the same embodied, serial cost that produced the trace. 73

This distinction matters because polished output is easy to misread as superior thinking. LLMs are strongest where the task is interpolation over many examples: pattern matching, specific semantic search, translation between familiar formats, boilerplate, and code in tight feedback loops where tests, compilers, logs, and human review supply a gradient. But coding is not software; search is not research; retrieval is not scientific discovery; a plausible plan is not engineering; and pattern completion is not business judgment, system architecture, original creativity, or the abstract layer of design. In daily use, the more a task resembles a dense cloud of prior data points, the more impressive the model looks. The more it requires deciding what should exist, what matters, what trade-off to own, or what future to build, the less the surface fluency tells us.

That is why the loss of friction matters so much. Before the shortcut, a person had to hold complexity in working memory, build a rough model of the world, ask what follows if this step is true, notice what breaks two steps later, and keep revising the imagined future. With an LLM in the loop, the surrogate fills the imagination gap. In interpolation-heavy tasks, the surface artifact may be cleaner, the bug may be fixed faster, or the plan may contain more branches than one tired person would have sketched alone. But the human can end up moving only one step at a time: prompt, accept, nudge, accept. The sprint engine never fires. The brain gets a constant slow walk through terrain the model has already paved, and the frightening part is that the output can look more polished while the internal machinery that would have generated it stays cold.

As a co-founder building Attryx, this is the product belief underneath the essay: AI should make the process more inspectable, not make the process disappear. A serious research tool should not train people to aim only at outcomes. It should help them climb the reasoning step by step, with assumptions, failed attempts, diagnostics, derivations, code, and intermediate results visible enough to examine. AI should not replace the friction that used to build understanding. It should move the frontier of discomfort: help someone attempt a harder math problem than they would have tried alone, run a more complex simulation than they could have staged by hand, compare more hypotheses, and still know how the answer was earned.

The human prescription cannot be only better AI interface design. Keep some thinking in the body and the physical world: play sports, navigate sometimes without turn-by-turn instructions, solve problems on paper, argue with people face to face, build things with your hands, and spend time in rooms where another person's reaction changes your next thought. The next interface will keep moving closer to real-time assistance; the pressure to outsource thinking on the spot will only grow. The best integration is not abstinence. It is deliberate alternation: use AI to extend the horizon, then step back into unaided reasoning, physical effort, and social reality so the brain still pays the costs that make it a brain.

Practical parity arrives at the crossing. The model will keep improving. The question is not whether AI can produce useful artifacts in the domains where interpolation is enough. It can. The question is whether we will use that power to climb toward harder worlds, or let it make the world small enough that we never have to climb.