A college junior sat across from me last fall, fidgeting with his phone. He had been referred for a cognitive evaluation after his grades collapsed despite, by his own account, spending more time than ever on coursework. “I don’t understand,” he said. “I use ChatGPT for everything now. My essays are better. My research is faster. But when I sit down for an exam without it, my mind goes completely blank. It’s like I forgot how to think.”
His cognitive testing revealed something I am now seeing with increasing frequency: strong verbal comprehension, solid perceptual reasoning, but a working memory score that had dropped 11 points from a baseline assessment taken three years earlier. His processing speed was unchanged. His fund of knowledge was intact. But the cognitive system responsible for holding information in mind, manipulating it, and generating original responses had measurably weakened.
He was not developing a neurological condition. He was experiencing what researchers at MIT have begun calling “cognitive debt,” the accumulated deficit that builds when you consistently outsource your thinking to an artificial intelligence system.
With over 900 million people now using ChatGPT weekly and 92 percent of UK university students reporting AI use in their studies, this is no longer a theoretical concern. It is a cognitive experiment being conducted at planetary scale, and the early results should give everyone pause.
The Google Effect Was Just the Beginning
To understand what AI is doing to your brain, you need to understand what Google already did to it.
In 2011, psychologists Sparrow, Liu, and Wegner published a landmark study in Science that documented what they called the “Google Effect on Memory.” Across four experiments, they demonstrated that when people believe information will remain digitally accessible, they encode significantly less of it into memory. Participants remembered where to find information far better than they remembered the information itself.
This was not necessarily alarming. Humans have always used “transactive memory,” distributing knowledge across social networks. You remember that your spouse knows the neighbors’ names and your colleague knows the software system, and you rely on them rather than memorizing everything yourself. Google simply became a new transactive memory partner.
But GPS navigation research revealed a darker pattern. Dahmani and Bohbot’s 2020 study in Scientific Reports tracked 50 regular drivers and found that greater lifetime GPS use predicted worse spatial memory during self-guided navigation. In a three-year longitudinal follow-up, increased GPS use predicted steeper decline in hippocampal-dependent spatial memory, suggesting GPS caused the decline rather than merely compensating for it. Brain imaging confirmed the mechanism: when participants navigated using turn-by-turn GPS instructions, hippocampal activity was significantly reduced compared to self-guided navigation. The brain’s spatial processing system simply disengaged when the work was outsourced.
Google changed how we remember. GPS changed how we navigate. AI is changing how we think. And the early evidence suggests the cognitive costs may be substantially larger.
The Study That Should Worry Every Student
The most rigorous evidence for AI’s cognitive effects comes from a randomized controlled trial published in Proceedings of the National Academy of Sciences in 2025. Bastani and colleagues assigned approximately 1,000 Turkish high school math students to three conditions: standard ChatGPT access, ChatGPT with pedagogical guardrails that provided hints rather than answers, or no AI access.
The results were striking. With AI assistance, students using standard ChatGPT solved 48 percent more practice problems, and those using the tutoring version solved 127 percent more. By every measure of practice performance, AI made students look dramatically more capable.
Then the AI was taken away, and students sat for unassisted exams.
Students who had used standard ChatGPT scored 17 percent worse than the control group that never had AI access. They had practiced more, felt more confident, and learned less. The AI had not supplemented their learning. It had substituted for it.
The students’ self-assessments were dangerously optimistic. They believed they had learned more than they actually had, a metacognitive failure that I find particularly concerning from a clinical perspective. If you do not know what you do not know, you cannot correct the deficit.
The study’s most important finding, however, was that the pedagogically designed AI tutor, which gave hints and guidance rather than direct answers, preserved learning outcomes entirely. The same underlying technology produced opposite cognitive effects depending on how it was designed. This distinction between AI as answer machine and AI as thinking partner turns out to be everything.
Your Brain on ChatGPT: The MIT Evidence
A 2025 study from MIT’s Media Lab provided the first neuroimaging evidence of AI’s effects on cognitive processing. Researchers fitted 54 participants with EEG headsets across four months of essay-writing sessions and compared brain activity across three conditions: writing entirely independently, writing with ChatGPT assistance, and writing with only a human editor.
The findings were sobering. Participants using ChatGPT showed the weakest neural connectivity across alpha, theta, and delta frequency bands, the brain waves associated with creativity, memory encoding, and deep semantic processing. Their brains were doing less cognitive work, not because the writing was easier but because they were delegating the hard parts of thinking to the machine.
The most troubling finding emerged from a crossover condition, where ChatGPT users were asked to write independently after months of AI-assisted work. Their brain activity did not return to baseline. The reduced neural engagement persisted even after the AI was removed, suggesting that months of cognitive offloading had produced lasting changes in how their brains approached the writing task. The researchers described this accumulated deficit as “cognitive debt,” an apt metaphor for something I now look for in every young adult I assess.
I want to be transparent about limitations: this study had a small sample size and remains a preprint rather than a peer-reviewed publication. But its biological specificity, measuring actual neural changes rather than self-reported habits, makes it an important early signal.
The “Use It or Lose It” Principle Is Not a Metaphor
The biological plausibility of AI-induced cognitive decline rests on one of neuroscience’s best-established principles: use-dependent plasticity is bidirectional. The brain grows with use and shrinks with disuse.
Eleanor Maguire’s London taxi driver studies provide the canonical demonstration. Drivers who spent years navigating London’s 25,000 streets developed significantly larger posterior hippocampi than controls, with volume correlating positively with years of experience. Her 2011 longitudinal study scanned 79 trainee drivers before and after three to four years of training and found that every single trainee who qualified showed increased hippocampal gray matter. Those who failed showed no changes. Cognitive training causes structural brain growth.
But here is the finding that matters for the AI question: retired taxi drivers who had stopped navigating showed smaller posterior hippocampi than those still active. The structural gains partially reversed with disuse. Draganski’s juggling studies confirmed this reversibility more directly: gray matter increases from three months of juggling practice reversed nearly to baseline after three months without practice.
At the population level, the Whitehall II Study tracking 3,433 civil servants found that verbal memory decline was 38 percent faster after retirement than before. Rohwedder and Willis, using cross-national pension policies as natural experiments, confirmed that early retirement causally accelerates cognitive decline. Arthur and colleagues’ meta-analysis of 189 data points quantified skill decay: effect sizes reached d = -1.4 after one year of nonuse, with cognitive tasks decaying faster than physical ones.
The implication is straightforward: if you consistently outsource reasoning, analysis, and composition to AI, you are reducing the cognitive demands on the neural systems that maintain those abilities. The “use it or lose it” principle predicts exactly the kind of decline that the PNAS math study and the MIT EEG study documented.
Automation Bias: Why Your Brain Trusts the Machine Too Much
There is a second, subtler mechanism through which AI degrades cognitive performance, and it has been studied for decades under the label “automation bias.”
Parasuraman and Manzey’s influential 2010 review in Human Factors documented two signature failure modes when humans work with automated systems. Omission errors occur when people fail to notice problems because the automation did not flag them. Commission errors occur when people follow automated recommendations even when they contradict available evidence. In aviation research, cockpit crews showed a 100 percent commission error rate on a false engine-fire alert: every single crew followed the automation’s incorrect recommendation.
These dynamics are now appearing with AI systems. Harvard researchers demonstrated that people frequently accept AI recommendations even when wrong, and that adding AI explanations paradoxically increased overreliance rather than reducing it. Only “cognitive forcing” interventions, designs that compel users to form independent judgments before seeing AI output, reduced the bias.
Perhaps the most alarming evidence of AI-induced deskilling comes from medicine. A 2025 study published in The Lancet Gastroenterology and Hepatology examined 1,443 patients across four endoscopy centers. After clinicians gained experience with AI-assisted colonoscopy, their detection rate for precancerous growths in non-AI-assisted procedures dropped from 28.4 percent to 22.4 percent, a clinically significant six-percentage-point decline. Physicians had become dependent on the AI’s pattern recognition, and when the AI was removed, their own perceptual skills had atrophied.
This is precisely the cognitive dynamic I see playing out with students, professionals, and knowledge workers who use AI for writing, analysis, and decision-making. The more you delegate, the less you practice. The less you practice, the weaker the skill becomes. And because AI makes the output look polished regardless of how much thinking you actually did, you may not notice the decline until the machine is taken away.
But Can AI Actually Make You Smarter?
The evidence is not entirely negative, and intellectual honesty requires presenting the other side with equal rigor.
The strongest empirical case for AI-enhanced learning comes from Kestin and colleagues’ 2025 study published in Scientific Reports. This Harvard randomized controlled trial compared an AI-powered physics tutor against in-class active learning. Students using the AI tutor learned significantly more in less time, with effect sizes between 0.73 and 1.3 standard deviations, approaching what educational researchers call the “2 sigma” threshold, the finding that one-on-one tutored students outperform 98 percent of conventionally taught peers.
The Bastani PNAS study itself contained the most instructive finding: while unguarded ChatGPT harmed learning by 17 percent, the pedagogically designed AI tutor preserved learning outcomes entirely. The same technology produced opposite effects depending on its design.
Vaccaro, Almaatouq, and Malone’s 2024 meta-analysis in Nature Human Behaviour, covering 370 effect sizes from 106 experiments, found that human-AI teams consistently outperformed humans alone. However, the pattern was task-dependent: creative and generative tasks showed genuine synergy, while decision tasks showed performance losses from collaboration.
The philosopher Andy Clark’s “Extended Mind” thesis provides a framework for understanding this: if an external tool functions equivalently to an internal cognitive process, reliably accessible, trusted, and action-guiding, it can be considered part of the cognitive system itself. By this logic, AI is not replacing cognition but extending it, much as writing, maps, and calculators extended cognition for previous generations.
Socrates warned in Plato’s Phaedrus that writing would “create forgetfulness in the learners’ souls.” He was partly right: literacy did reduce oral memorization capacity. But he was catastrophically wrong about the net effect. Writing enabled entirely new forms of analytical thinking impossible in oral cultures.
The question is whether AI follows the trajectory of writing, expanding what human minds can do, or the trajectory of GPS, atrophying a specific cognitive capacity that we needed more than we realized.
The Distinction That Matters: Delegation vs. Amplification
After reviewing every major study in this field, I believe the critical distinction is not between “using AI” and “not using AI.” It is between cognitive delegation and cognitive amplification.
Cognitive delegation means outsourcing your thinking entirely. You paste a question into ChatGPT and accept the answer without engaging your own reasoning. You let the AI write your essay, compose your email, or make your decision. You consume the output without processing it deeply. This is the pattern that consistently produces cognitive costs in the research.
Cognitive amplification means using AI to extend and enhance your own thinking. You draft your own argument first, then use AI to identify weaknesses. You solve the problem yourself, then check your reasoning against the AI’s approach. You use AI to access information you could not otherwise find, but you integrate and evaluate that information through your own judgment. This is the pattern associated with maintained or enhanced cognitive performance.
A Microsoft Research study of 319 knowledge workers captured this distinction perfectly: higher confidence in AI tools was associated with less critical thinking engagement. The workers who trusted AI the most thought the least. And the researchers noted the central irony: by automating routine cognitive tasks, AI deprives users of the routine practice opportunities that maintain their cognitive skills, leaving them “atrophied and unprepared” when exceptions arise.
What This Means for Your IQ
Everything I have described in this article converges on a single practical question: what is AI doing to the cognitive abilities that IQ tests measure?
IQ tests assess working memory, processing speed, verbal reasoning, and abstract pattern recognition. These are precisely the abilities that cognitive offloading research suggests are vulnerable to disuse-dependent decline. If you consistently let AI handle your verbal reasoning (writing emails, composing arguments, summarizing information), your working memory engagement (holding complex problems in mind while you work through them), and your abstract reasoning (analyzing data, identifying patterns, making inferences), you are reducing the cognitive demands on exactly the systems that IQ tests measure.
This does not mean using ChatGPT will lower your IQ. It means that the pattern of using AI as a replacement for thinking, rather than as a complement to thinking, may gradually reduce the cognitive fitness that IQ tests capture. The analogy is physical: using a wheelchair when you can walk does not break your legs, but it does weaken them over time.
Understanding where your cognitive abilities currently stand gives you a baseline against which to measure whether your AI habits are helping or hurting. A comprehensive cognitive assessment tells you not just a number but a profile: your relative strengths in verbal reasoning, spatial processing, working memory, and processing speed. That profile tells you which cognitive systems you most need to keep exercising and which you can more safely augment with AI.
How to Use AI Without Losing Your Mind
Based on the converging evidence from the studies discussed in this article, here is the framework I now share with patients, students, and professionals who ask me how to use AI wisely.
Think first, then consult. Form your own position, draft your own argument, or attempt your own solution before asking AI for input. The PNAS math study showed that AI tutoring that preserved productive struggle maintained learning, while AI that eliminated struggle harmed it. The cognitive work of generating your own response, even an imperfect one, is what builds and maintains the neural pathways that support reasoning.
Use AI to challenge, not to confirm. Ask ChatGPT to argue against your position, find flaws in your reasoning, or present perspectives you have not considered. This engages your critical thinking rather than replacing it. The automation bias research shows that we tend to accept AI output uncritically. Deliberately using AI as a sparring partner rather than an oracle counteracts this tendency.
Maintain AI-free cognitive practices. Read books on paper. Write by hand when learning new material. Solve problems without digital assistance regularly. Navigate without GPS occasionally. These practices are not nostalgic indulgences. They are cognitive exercise that maintains the neural systems AI tends to atrophy.
Monitor your own cognitive patterns. If you notice that you can no longer compose an email without AI assistance, that you struggle to remember information you recently read, or that your mind goes blank when you try to reason through a problem independently, these are warning signs that cognitive delegation has gone too far.
And know your baseline. You cannot track cognitive change without knowing where you started. Understanding your current cognitive profile, your working memory capacity, your verbal reasoning strength, your processing speed, gives you both a benchmark and a map for where to focus your cognitive maintenance efforts.
Final Thoughts
In three decades of assessing human cognition, I have watched technology reshape the cognitive landscape in ways that are neither entirely good nor entirely bad, but always consequential.
AI represents the most powerful cognitive tool our species has ever created. It can extend human thinking in ways that were genuinely impossible a decade ago. But the same power that makes it transformative also makes it dangerous, because the easier it becomes to outsource your thinking, the less your brain practices the skills that make you capable of thinking well.
The research is early, the sample sizes are small, and the long-term effects remain genuinely unknown. We are conducting an uncontrolled cognitive experiment on a billion people, and the longitudinal data that would reveal whether AI-induced changes are temporary adaptations or durable deficits simply does not yet exist.
What does exist is a clear principle, supported by decades of neuroscience: cognitive abilities that are not exercised will decline. The brain is not a museum where skills are preserved in glass cases. It is a living system that maintains only what it actively uses.
Your intelligence is not fixed. It is maintained. And in a world where machines can think for you, the choice to keep thinking for yourself is no longer automatic. It is a decision you make every time you reach for the chatbot instead of reaching for the answer yourself.
