Cognitive delegation. The muscle you don't use atrophies

The MIT Media Lab study Your Brain on ChatGPT (2025) showed with EEG that LLM users had the weakest brain connectivity of three groups (brain-only, search engine, LLM) during a writing task. You're externalizing parts of thinking every time you ask for something to be "summarized for me" or "told whether it's right." It isn't harmless. Each delegation is a muscle you don't use and that atrophies. This isn't alarmism, it's basic cognitive physiology. And nobody is measuring the wear — we'll see it in ten years, when there's no turning back.

The figure that has been measured

It's worth starting with the data because the data exists and because the public conversation about cognitive delegation is still conducted in speculative terms when it's already past that point. Researchers at the MIT Media Lab published in 2025 Your Brain on ChatGPT. Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task (arXiv 2506.08872), with 54 participants in three experimental conditions: a "brain-only" group (no tools), a "search engine" group (with a search engine) and an "LLM" group (with ChatGPT). Brain activity was measured with EEG during the task of writing a short essay.

The results, in a single sentence: the LLM group showed the weakest neural connectivity of the three groups, and the brain-only group the strongest. The search-engine group fell in between. The metric wasn't user satisfaction or the surface quality of the text produced. It was measurable electrical activity of the neural networks during the execution of the task. When the model does part of the work, the brain stops doing that part of the work. It's not a metaphor. It's what the electrode records.

There was an added behavioral finding worth holding on to. After the first session, 83% of the LLM-group participants reported difficulty quoting a sentence from the essay they had just presented as their own — a proportion that dropped to 33% in the third session. The LLM group also reported the lowest sense of authorship of the three. The essays produced by the LLM group were statistically more homogeneous among themselves, with less thematic variation, than those of the other two groups. The text was theirs in a legal and administrative sense. Cognitively it hadn't been.

The precedent rarely talked about

The MIT researchers didn't discover a new class of phenomenon. What they documented has a direct precedent in the literature. Sparrow, Liu and Wegner, in Google Effects on Memory. Cognitive Consequences of Having Information at Our Fingertips (Science 333(6043), 2011), measured what they called the Google effect: when participants expected to be able to look up a piece of information again, their ability to recall it dropped significantly compared to when they expected not to be able to. Memory adjusted to the availability of the external source. The figure had been there since 2011. The only thing that changed in 2025 was the sophistication of the externalizer.

Risko and Gilbert, in Cognitive Offloading (Trends in Cognitive Sciences 20(9), 2016), generalized the framework. Cognitive offloading, the partial transfer of mental processes to external tools, is a universal and adaptive phenomenon in itself. Using a notebook so as not to memorize the shopping list is offloading. Noting an appointment in the calendar is offloading. The question isn't whether we delegate cognition — we always have, with whatever supports we had at hand — but what specific function is delegated, how often, and what cognitive price the delegating agent pays.

What a notebook doesn't do, and an LLM does, is execute the higher-order cognitive functions that make up thinking: argumentative structuring, comparison between alternatives, critical evaluation, writing that organizes thought as it's written. The notebook stores; the LLM composes. The difference matters because composition is the place where thought is formed.

What you delegate without noticing

It's worth listing, without moralizing, what an active AI user delegates in an ordinary week of intellectual work.

They asked the model to summarize two long reports so as not to read them in full. They asked it to rephrase a paragraph of their own to sound more professional. They asked it to compare two technical options and tell them the pros and cons. They asked it to structure an argument they had half-cooked. They asked it to review a text and tell them whether it was good. They asked it to decide between two phrasings for a difficult email. They asked it to search and filter information about a topic rather than open a search engine and read five sources. They asked it to explain a technical concept so as not to read the paper. They asked it to draft a first version that they would later "only touch up."

Each of these operations, on its own, looks like a reasonable saving of time. The cognitive bill isn't paid in the individual operation. It's paid in aggregate, when the muscle of doing each of those operations by hand has gone unpracticed for months. The practice was the operation. The operation was the thinking.

The muscle and bidirectional plasticity

And here's where the muscle analogy stops being a comfortable metaphor. Muscle physiology is well known. An arm immobilized for six weeks loses measurable mass. Critical thinking, argumentative capacity, the skill of writing under pressure, working memory applied to complex topics — all these are functions cognitive neuroscience treats as use-dependent, with bidirectional plasticity. What isn't exercised weakens. What is exercised holds or strengthens. The literature on cognitive load theory by Sweller and others documents this principle in educational contexts since the nineties.

Common intuition says the brain isn't a muscle and so the analogy doesn't apply. Common intuition is half wrong. The brain isn't a muscle, yes, but the rule use it or lose it has an empirical basis in specific cognitive skills. What the literature debates is the magnitude of the effect and the recovery time after periods of disuse, not whether the effect exists. It exists.

The temporal asymmetry nobody offsets

There's a property of saving-through-delegation worth naming because it explains why willpower isn't enough to correct it. The benefit is immediate and visible. The cost is deferred and invisible.

When you delegate the drafting of the paragraph to the LLM, you save fifteen minutes. Those fifteen minutes are yours today, verifiable, spendable on something else. When, six months later, you sit down to write something important without the LLM available and notice it costs you more than before, you don't associate the difficulty with the hundreds of small prior delegations. You associate it with a bad day, with tiredness, with the topic being hard. The causal chain is too distributed in time for the cognitive system to recognize it.

This pattern of temporal asymmetry is the same as the unhealthy diet, the sedentary lifestyle, consumer debt, retirement saving. Behavioral economics has spent decades documenting that humans discount the future non-linearly: we give more weight to the small immediate benefit than to the much larger future cost. Kahneman formalized it. Daniel Ariely wrote books about it. The cognitive physiology of delegation is subject to the same asymmetry, and willpower alone doesn't correct it because willpower operates in the present and the cost isn't seen.

What could correct it is information distributed over time about the cost. But that information doesn't exist. There's no marker telling you "your ability to write under pressure has dropped 11% in six months of intensive use." There's no personal critical-thinking benchmark that updates when you delegate. There's no alarm. And since there's no alarm, there's no correction.

Gerlich and the other end of the data

Michael Gerlich, in AI Tools in Society. Impacts on Cognitive Offloading and the Future of Critical Thinking (Societies, 2025), contributes cross-sectional evidence complementary to MIT's. He found a significant negative correlation between frequency of AI-tool use and scores on standardized critical-thinking tests, with a larger effect in younger participants. Causality, as in any cross-sectional study, isn't established. What is established is the coincidence between intensive use and low performance on the specific measures.

There's also recent work by Jiang, Wu and Leung published in Frontiers in Computational Neuroscience (2025) that adds EEG data on LLM interactions in problem-solving and decision-making tasks. Its authors read the data in an almost opposite key: the model's assistance reduces cognitive load by externalizing the reasoning, which is reflected in lower frontal theta activity during the task. The same measurable drop in effort that MIT interprets as debt, this team interprets as efficiency. That two labs measure the same thing and disagree on whether it's good or bad is, in itself, the most honest data point: what the brain stops doing isn't in dispute, what's in dispute is whether it matters that it stops doing it.

What's notable is that the raw data is already on the table, whatever its sign. The public conversation about AI productivity nonetheless keeps telling only half the balance sheet: the 26% increase in tasks completed by developers with Copilot (Cui and others, 2025), the hours saved by office workers, the multiplied writing speed. That half is real. The other half, the accumulated cognitive cost, doesn't appear in the McKinsey reports or in the investor presentations. Not because it doesn't exist. Because it's not yet time to tell it.

The conversation that isn't being had

There's a question that exceeds this article and is worth posing anyway. What happens with a generation of young professionals who enter the labor market having delegated, since university, the part of writing, comparing, structuring, deciding, reviewing? The optimistic intuition says that, freed of those tasks, they'll devote their time to higher-order tasks. The pessimistic intuition says that the higher-order tasks require precisely the skills they haven't practiced.

Both intuitions have the same flaw: they're speculation. What is known, in light of the available literature, is that the ability to write critically under pressure doesn't appear by osmosis. It's built with hours of deliberate practice, with friction, with error and correction, with time alone facing the text. Whoever hasn't practiced it between 18 and 25 won't have it at 30 without an explicit recovery program. And explicit programs to recover cognitive skills are long and costly, and don't scale to the general population.

Carr, in The Shallows (Norton, 2010), framed the problem in its earlier form — the internet and scattered attention — and framed it well. The part of his book that aged worst is the prescriptive part, the recipe for mitigating the effect. The part that aged best is the descriptive: when a new and very powerful technological support mediates a large part of everyday cognitive operations, the user's cognitive architecture reorganizes around that support, at a speed greater than the speed at which society can debate whether the reorganization is desirable.

Is it wrong to delegate to AI? The question put that way is badly framed. It's wrong to delegate what makes up the thinking you wanted to develop. It's fine to delegate what you didn't want to develop anyway. The line between the two isn't obvious and, above all, it isn't drawn by the interface, it's drawn by the user. The interface, for its part, is designed so that crossing the line in one direction is easy and coming back the other way is costly.

Definitions

Cognitive offloading. The partial transfer of mental processes to external tools — paper, calendars, search engines, now LLMs. A universal and adaptive phenomenon in general; problematic when the delegated function is structural to thinking.

Cognitive debt. A term coined by the MIT Media Lab researchers (2025) to describe the deferred cost of intensive LLM use on cognitive tasks: immediate saving of effort and accumulated deterioration of the capacities involved.

Neural connectivity. A pattern of coordinated activity between brain regions, measurable with EEG. It's associated with the intensity and depth of cognitive processing during a task. It decreases consistently when the subject delegates part of the task to an LLM.

Cognitive Load Theory. A framework proposed by John Sweller and collaborators. It posits that learning and the acquisition of cognitive skills depend on the active effort of processing in working memory. Delegating that effort reduces consolidation.

Google effect. A pattern documented by Sparrow, Liu and Wegner (2011) by which the memory of information adjusts to its external availability. A direct precedent of the phenomena observed with LLMs in 2025.

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 support on attention and reading depth.

Cui, K. Z. et al. (2025). The Effects of Generative AI on High-Skilled Work. Evidence from Three Field Experiments with Software Developers. Management Science. A 26% increase in tasks completed with Copilot assistance; cited as a counterpart to the absent balance of cognitive cost.

Gerlich, M. (2025). AI Tools in Society. Impacts on Cognitive Offloading and the Future of Critical Thinking. Societies 15(1), 6. A negative correlation between intensive AI-tool use and performance on critical-thinking tests, with a larger effect in young participants.

Jiang, T., Wu, J. & Leung, S. C. H. (2025). The cognitive impacts of large language model interactions on problem solving and decision making using EEG analysis. Frontiers in Computational Neuroscience 19. EEG data on LLM interaction; the authors read the lower frontal theta activity as a reduction of cognitive load.

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. Weaker neural connectivity in the LLM group versus the search-engine and brain-only groups; difficulty quoting one's own essay in 83% of LLM-group participants after the first session, dropping to 33% in the third.

Risko, E. F. & Gilbert, S. J. (2016). Cognitive Offloading. Trends in Cognitive Sciences 20(9), 676–688. A general metacognitive framework on the transfer of mental processes to external tools.

Sparrow, B., Liu, J. & Wegner, D. M. (2011). Google Effects on Memory. Cognitive Consequences of Having Information at Our Fingertips. Science 333(6043), 776–778. The foundational study on the adjustment of memory to the availability of external sources.

Sweller, J., van Merriënboer, J. J. G. & Paas, F. (2019). Cognitive Architecture and Instructional Design. 20 Years Later. Educational Psychology Review 31. An updated synthesis of Cognitive Load Theory applied to educational and cognitive-delegation contexts.

Going deeper

Newport, C. (2016). Deep Work. Rules for Focused Success in a Distracted World. Grand Central. A practical application of the cost of interruptions and constant delegation to contemporary intellectual work.

You may also like

• The cost of thinking less
• Augmented cognitive laziness
• What happens to your brain when you use ChatGPT, according to MIT
• The externalization of knowledge

Elsewhere

• Your Brain on ChatGPT — Project Overview (MIT Media Lab)
• The 2025 AI Index Report (Stanford HAI)