Does AI Make Us Stupider? It Depends on What We Outsource

Warnings that ChatGPT is making us less intelligent have become increasingly common. Students can generate essays without learning to write them. Workers can ask a model to construct arguments they have barely examined. One much-discussed MIT experiment found weaker patterns of measured brain connectivity among people writing essays with an LLM than among those working unaided.

Editorial image generated by the author.

There is a real problem here, but “using AI” is too broad a category to explain it. Someone who encounters a difficult question, asks ChatGPT for the answer and copies what appears on screen is doing something cognitively different from someone who starts with a hypothesis, uses an LLM to locate research, asks it to construct the strongest counterargument, checks the underlying evidence and revises the hypothesis accordingly. Both are using the same technology, but only one is necessarily outsourcing the thinking the exercise was supposed to develop.

The useful question is therefore which parts of thinking we choose to delegate. It becomes more interesting once AI is treated not merely as a faster way to complete existing work, but as something that changes the economics of intellectual exploration itself. If investigating an idea becomes dramatically cheaper, we may investigate far more ideas than we otherwise would. The potential gain is not simply faster thinking, but more thinking across a much larger territory. The risk is its mirror image: the same machine that makes an idea cheap to test also makes it cheap to rationalise.

We Have Always Outsourced Parts of Thought

Human cognition has never been confined to the human brain. Writing stores memories externally. Libraries preserve knowledge no individual could retain. Calculators perform arithmetic. Spreadsheets manipulate numbers. Search engines retrieve information. Calendars remember obligations and maps preserve geographical knowledge. Psychologists describe the broader phenomenon as cognitive offloading: altering the environment so that some part of a cognitive task no longer has to be performed internally.

There are obvious trade-offs. Someone who always navigates with GPS may become worse at navigating without it. Someone who constantly uses a calculator may lose speed at mental arithmetic. Most people now remember far fewer telephone numbers than they did thirty years ago. Yet it would be difficult to argue that external memory, calculators and searchable databases have reduced humanity’s overall intellectual capacity. They allow scarce attention to move elsewhere.

The mistake is to treat retention of every unaided skill as equivalent to total cognitive capability. An accountant who no longer performs long division by hand may have lost some arithmetic practice without becoming a worse accountant. If a spreadsheet removes hours of mechanical calculation while allowing more time to understand what the figures mean, the balance can easily be positive.

Large language models extend this old process into more sensitive territory. They do not merely store facts or execute calculations. They can summarise evidence, explain concepts, compare arguments, generate alternatives and produce coherent prose. They can therefore offload operations much closer to what we ordinarily call thinking.

That is also where the analogy with older tools begins to break down. A calculator cannot convincingly invent an argument for why its incorrect answer is actually right. An LLM can.

When Better Performance Hides Weaker Learning

Some of the clearest evidence for cognitive harm comes from education, where the difference between producing a correct answer and developing the capacity to produce one is unusually easy to observe. A 2025 study published in Proceedings of the National Academy of Sciences examined nearly a thousand high-school mathematics students. One group had access to an interface resembling ordinary ChatGPT, another used a specially designed GPT tutor that relied more heavily on teacher-designed hints, and a control group worked without AI.

While AI was available, both AI groups performed extremely well. Students using the relatively unrestricted system scored 48 percent better than the control group during practice; those using the structured tutor scored 127 percent better. When the assistance was removed, however, students from the unrestricted group performed 17 percent worse than those who had never received AI assistance. The tutor group did not suffer the same substantial penalty.

The interesting result is therefore not simply that AI harmed learning. The same underlying technology produced different outcomes depending on what the interface encouraged students to outsource. One system made it easy to obtain a solution. The other kept more of the reasoning process with the student.

An August 2026 NBER working paper points in a similar direction. In a randomized field experiment involving more than 6,000 middle-school students, structured AI assistance improved recovery after mistakes: students were more likely to answer the next attempt correctly and required fewer attempts to return to a correct answer. They also spent longer on supported questions, progressed more slowly and attempted fewer questions overall. The effect was not simply “AI tutoring works” or “AI tutoring fails.” The structure of the interaction changed what students did with their mistakes.

A review published in October 2026 gives this distinction a useful name: augmentation versus erosion. Its argument is that the educational effect of generative AI is not a property of the tool itself but of how cognitive labour is divided between learner and machine. Offloading a peripheral operation can free attention for the thing being learned; offloading the operation that is the learning task can bypass the practice required to acquire or retain the skill.

The mechanism is straightforward. Learning a skill normally requires performing the cognitive operation the skill consists of. Remove that operation systematically and immediate performance can improve while independent capability fails to develop. For a student learning algebra, having AI perform the algebra is quite different from having AI explain why the student’s own solution failed. The first replaces practice; the second can intensify it.

What the Brain-Scan Headlines Show

The most dramatic version of the cognitive-decline argument came from the 2025 MIT Media Lab preprint Your Brain on ChatGPT. Participants wrote essays either without assistance, with a search engine or with an LLM while researchers measured EEG connectivity. The unaided group showed the strongest and most distributed connectivity, search users occupied the middle, and LLM users showed the weakest. LLM users also had more difficulty accurately recalling what they had written and reported less ownership of their essays.

The results are suggestive, but they do not establish that ChatGPT makes people less intelligent. Fifty-four participants completed the main sessions and only 18 took part in the crossover stage, while a later methodological critique raised questions about sample size, reproducibility, reporting and interpretation of the EEG analysis. The work also remains a preprint rather than a settled longitudinal finding about cognitive decline.

More fundamentally, greater measured cognitive effort is not automatically evidence of a better outcome. A difficult calculation performed mentally may demand more effort than the same calculation performed with a spreadsheet. That difference matters enormously if the objective is to practise calculation and much less if the objective is to analyse a company’s finances. The MIT experiment is therefore more persuasive as evidence that LLM-assisted writers engaged less deeply with the particular writing task than as evidence that their general ability to think had deteriorated.

Substitution and Augmentation Are Not Properties of the Tool

The difference can be described as cognitive substitution versus cognitive augmentation:

Cognitive substitution	Cognitive augmentation
“Give me the answer.”	“Help me investigate this question.”
“Write my argument.”	“Find weaknesses in my argument.”
“Summarise this so I don’t have to understand it.”	“Explain the part of this paper I don’t understand.”
“Tell me what I should think.”	“Give me the strongest competing interpretations.”
“Solve the problem.”	“Show me where my solution may be wrong.”
AI supplies the judgement.	Human retains the judgement.

Reality is less tidy than the table. The same operation can be substitution in one context and augmentation in another. Asking AI to summarise a paper instead of reading it may be disastrous for a student who is supposed to learn close reading. Asking it to summarise thirty regulatory announcements that an experienced specialist already knows how to interpret may simply remove mechanical work. The relevant test is whether the delegated operation is itself a capability the user needs to develop or preserve.

A Microsoft study presented at CHI 2025 illustrates the ambiguity. Researchers surveyed 319 knowledge workers about 936 real-world instances of generative-AI use. Participants described critical thinking as shifting toward verification, response integration and oversight rather than simply disappearing. Higher confidence in the AI was nevertheless associated with less reported critical thinking, while greater confidence in one’s own expertise was associated with more.

Because the study concerns self-reported behaviour, it cannot establish that AI caused a measurable gain or decline in critical-thinking ability. The reported shift is nevertheless plausible: automation can move attention upward in a workflow, away from manually producing every component and toward evaluating what the system produced. That gain disappears if evaluation itself is gradually delegated. Once trust becomes high enough, assistance can turn into acceptance without any obvious moment at which the user decided to stop thinking.

AI Changes the Economics of Curiosity

Most research on AI productivity asks how much faster someone can complete an existing task. The gains can be substantial. In a study of 758 Boston Consulting Group consultants, now published in Organization Science, GPT-4 users completed suitable tasks more quickly, completed more of them and produced higher-rated work. The same study also demonstrated the “jagged technological frontier”: on a complex task deliberately chosen because the model handled it badly, AI-assisted participants were 19 percentage points less likely to reach the correct answer than the control group.

The more revealing detail is that AI still improved the coherence and persuasiveness of the responses, including incorrect ones. That finding captures a problem larger than hallucination. A system can simultaneously make an argument more convincing and make its user less likely to notice that the argument rests on a wrong conclusion.

Task-completion experiments may nevertheless understate the most consequential productivity effect. Before LLMs, an interesting but unfamiliar question carried a substantial activation cost. Suppose someone noticed that a particular tax policy appeared to coincide with changes in investment. Investigating the idea properly might require learning unfamiliar terminology, finding academic literature, locating historical data, identifying comparable countries, understanding competing theories and searching for evidence against the original hypothesis. Several evenings could disappear before the interesting part of the investigation even began.

Most casual hypotheses therefore died early, not necessarily because they were poor ideas but because discovering whether they were good cost too much. An LLM can compress much of that preliminary work. It can identify relevant concepts, suggest search terms, point toward unfamiliar literatures, generate possible mechanisms, compare jurisdictions and expose obvious counterarguments. None of those outputs deserves automatic trust, but they can move someone remarkably quickly from “I wonder whether...” to the point where serious investigation becomes possible. That is closely related to the broader potential of artificial intelligence as a discovery engine: its value may lie as much in lowering the cost of traversing unfamiliar intellectual territory as in generating finished answers.

If investigating an idea once required five hours and now requires one, the result need not be that the same person investigates the same number of ideas five times faster. They may investigate several times as many ideas. Questions that once fell below the threshold of being worth pursuing suddenly become cheap enough to test. AI’s largest intellectual productivity effect may therefore be extensive rather than intensive: expanding the number of subjects a person can afford to think seriously about.

The same economics makes confirmation bias cheaper. Begin with the belief that a particular tax causes entrepreneurs to leave a country and ask an LLM to find supporting evidence. Within seconds it can produce examples, plausible mechanisms, references and a polished structure. The existence of that structure says very little about whether the hypothesis is true; the system was asked to construct a case and did so.

A better investigation would also ask what evidence would falsify the hypothesis, which jurisdictions contradict it, what alternative mechanisms could explain the observation, and what a sceptical economist would attack first. The technology capable of making hypothesis testing extraordinarily efficient is equally capable of making confirmation bias extraordinarily efficient.

This may be a more serious risk for experienced users than the obvious temptation to have AI do one’s homework. A sophisticated user can retain ownership of the initial idea while quietly outsourcing too much of the adversarial process needed to decide whether the idea deserves to survive. LLMs are unusually dangerous in that role because they are extremely good at coherence. A mediocre premise can acquire an impressive intellectual superstructure long before anyone has established that it is the best explanation for the facts.

Fluency was never the same thing as truth. Generative AI has simply made the gap easier to industrialise.

Creativity Gains, Intellectual Convergence

Research on creativity reveals another version of the same trade-off. In a Science Advances experiment by Anil Doshi and Oliver Hauser, writers given access to generative-AI ideas produced short stories rated as more creative, better written and more enjoyable, with the largest gains among participants who initially scored lower on creativity.

Yet the AI-assisted stories also became more similar to one another. The individual outputs improved while collective diversity narrowed. Large language models have learned patterns from enormous quantities of existing human expression, making them exceptionally good at producing plausible associations and competent variations. The same ability gives them a gravitational pull toward recognisable patterns.

The stage at which AI enters the process may therefore matter as much as whether it enters at all. Beginning with “give me ten interesting ideas to write about” allows the model to define much of the search space before the human has contributed anything. Beginning with an observation, irritation, contradiction or intuition developed independently and then using AI to explore it preserves more of the original divergence. The broader cultural consequences of cheap generative production are explored in the wider economics of AI-generated culture; for individual thinking, the narrower lesson is that AI may be most useful after the first spark and before the final judgement.

What Should We Actually Outsource?

Resisting cognitive offloading altogether is not a serious option. There is too much information in the world and too little human attention. Requiring people to perform every calculation manually, search every document themselves, repeatedly convert information between formats or reconstruct standard summaries from scratch would preserve plenty of cognitive labour without preserving much that is valuable.

Routine retrieval, transcription, formatting, repetitive comparison, straightforward calculation, initial literature discovery and the mechanical parts of synthesis are obvious candidates for aggressive automation. More caution is warranted around deciding which questions are interesting, forming hypotheses, identifying hidden assumptions, judging evidence, distinguishing correlation from causation, deciding what would change one’s mind, weighing competing explanations and taking responsibility for the final judgement.

The boundary will move as models improve. Some activities that currently require expert judgement will eventually be performed better by machines. Capability, however, is not the only issue. A society that delegates judgement still needs people capable of recognising when the delegated system has failed. Otherwise increasing machine competence creates a peculiar dependency: humans become progressively more reliant on systems they are progressively less able to supervise.

The problem is one of cognitive comparative advantage. We do not need to preserve a task merely because humans once performed it manually. We do need to preserve enough of the surrounding cognitive architecture to understand what the machine is doing, challenge it when necessary and decide whether its output deserves to matter.

The Cognitive Exoskeleton

A physical exoskeleton offers a useful analogy. Give a worker a machine that allows her to lift five times as much weight and her productive capacity immediately rises. That fact alone says little about what happens to her muscles. If the machine allows heavier work while she continues exercising normally, she gains capability without much physical loss. If it performs every movement for her, her muscles may weaken even while her ability to move objects increases. Augmentation and atrophy can occur at the same time.

LLMs may produce the cognitive equivalent. Someone might become worse at writing first drafts from scratch while becoming better at evaluating arguments. They may retain fewer isolated facts while becoming capable of synthesising evidence across a much larger field. They may become less skilled at manually searching for information while becoming better at deciding which information is relevant. None of those changes can be classified simply as “smarter” or “stupider” without specifying which capability we care about.

We do not yet have decades of longitudinal evidence showing what happens to people who conduct much of their intellectual lives with AI assistance. Caution is particularly justified in education, where the point of an exercise is often not its output but the struggle required to develop the underlying skill. A child who has never acquired a capability is poorly placed to decide whether that capability is safe to delegate.

For adults who already possess the relevant skills, the calculation can be different. The emerging evidence does not support treating every use of an LLM as equivalent. An AI that supplies an answer can reduce learning; one that makes a student work through mistakes can support the process. An AI that writes an argument can substitute for reasoning; one instructed to attack that argument can provoke reasoning that might otherwise never have occurred. An AI that generates the initial idea may narrow the search space, while one used to investigate an independently generated idea can widen it dramatically.

The divide is less between people who use AI and people who do not than between people who retain cognitive ownership of what they are doing and people who gradually surrender it. Maintaining that ownership may become harder as the systems improve. Bad systems invite scepticism because their mistakes are obvious. Reliable systems invite dependence. The better a machine becomes at producing polished, plausible and usually correct answers, the easier it becomes to stop asking whether one particular answer deserves to be trusted.

The challenge is therefore not to preserve every inefficient way humans used to work. Much of that inefficiency really was wasted effort, and removing it can free enormous amounts of time for more interesting questions. The harder task is distinguishing friction that merely consumed attention from friction that provided the practice required to retain judgement.

If that balance can be maintained, the largest intellectual benefit of artificial intelligence may not be that it thinks so we do not have to. By removing large quantities of cognitive grunt work, it can make curiosity cheaper: more hypotheses tested, more evidence compared, more unfamiliar subjects explored and more weak ideas discarded before anyone invests heavily in them.

If the balance is lost, we may still produce more than ever while becoming progressively less capable of determining whether any of it is worth producing.

The question for the AI age is what we can safely stop thinking about without losing our ability to think about the things we cannot safely delegate.