⚡ Published May 30, 2026 — researched and written by Aditya Kumar Jha. Every figure is sourced from peer-reviewed or named-institution research. Key facts: MIT Media Lab's EEG study on 54 students found the ChatGPT-only group showed up to 55% lower cognitive engagement than students who wrote unaided (arXiv:2506.08872, June 2025). Neuroscientist Dr. Jared Cooney Horvath testified before the U.S. Senate Committee on Commerce, Science, and Transportation in January 2026 that Gen Z is the first generation in recorded history to test cognitively below their parents. A Brookings Institution report published January 2026 — covered by Fortune in January and March 2026 — analyzed 400+ academic studies and 500+ educator and parent interviews across 50 countries and concluded the risks of generative AI in children's education currently outweigh its benefits. 92% of college students in 2026 now use AI when studying — but only 18% submit unedited AI-generated work as their own. Pearson analyzed nearly 80 million student interactions (February 25, 2026) and found a single embedded AI study-tool interaction triples the likelihood of active reading behavior — and a single AI interaction inside instructor-led courseware increases it by 23 times. The difference between AI that strengthens your mind and AI that atrophies it is not which tool you use. It's how.
In mid-2025, MIT Media Lab researchers wired 54 university students to EEG headsets, split them into three groups, and asked each to write a series of essays — one group using only their own brains, one using Google Search, and one using ChatGPT. Then they measured what was happening inside each participant's skull in real time. Congressional witnesses cited the findings. Boardrooms from Silicon Valley to Beijing debated them. Follow-up research replicated the results across three continents. The answer to 'is AI making students dumber?' turns on a single variable that most coverage of this topic misses entirely.
The MIT Brain Study That Changed the Conversation
MIT Media Lab researcher Nataliya Kosmyna and her colleagues ran a four-month longitudinal study that has become the most-cited piece of neuroscience in the AI-and-education debate. The setup: 54 students from Boston-area universities including MIT and Harvard, each wearing an electroencephalography headset measuring brain electrical activity in real time. Three conditions. Three sessions of essay writing. Then, in a fourth session, the groups switched — the ChatGPT users wrote without AI for the first time, and the no-tools group tried ChatGPT.
The brain data left no ambiguity. Students who wrote without any tools showed the strongest neural connectivity across the frontal-parietal networks — the regions responsible for executive function, deep synthesis, and complex reasoning. The ChatGPT group showed the weakest. The search-engine group sat in the middle. Cognitive debt had accumulated silently in the AI group: with each session, their brains did less and less.
The 55% Number and What It Actually Means
Across several cognitive engagement metrics, the ChatGPT group showed up to 55% lower brain activity than the no-tools group. This is a measure of real-time cognitive engagement during the task — not IQ, and not permanent neurological damage. If a student's brain is measurably less engaged every time they use AI to write, and they do this hundreds of times over a degree, what is the cumulative effect on the learning that writing is supposed to produce?
The fourth session gave part of the answer. When the ChatGPT group tried to write without AI, their brain connectivity did not recover to the no-tools group's level. The reduced engagement pattern persisted. The researchers used the term 'cognitive debt' deliberately: outsourcing thinking to AI incurs a cost that compounds with repeated outsourcing. 83% of ChatGPT-group students could not accurately recall key arguments from their own essays moments after submitting them. They had produced the words without forming the memories.
Congress Heard the Evidence in January 2026
Neuroscientist and education researcher Dr. Jared Cooney Horvath testified before the U.S. Senate Committee on Commerce, Science, and Transportation in January 2026 on the relationship between technology and cognitive development. His central finding, backed by standardised assessment data compiled across 80 countries: Gen Z is the first generation in modern recorded history to score below their parents on cognitive measures. He attributed this not to a single cause but to a compounding pattern — digital technology in classrooms creating cognitive offloading habits at precisely the developmental stage when effortful independent practice matters most.
- Approximately half of American 12th-grade students now score below basic level in mathematics, according to National Assessment of Educational Progress data cited in Horvath's Senate testimony — a benchmark that would have triggered national alarm in any prior generation.
- Roughly one in three American 12th-grade students lacks basic reading skills. The decline correlates with the expansion of screen time and digital tool use in classrooms, not with changes in student aptitude.
- Horvath's testimony framed this explicitly as a technology-design problem: 'This is not a debate about rejecting technology. It is a question of aligning educational tools with how human learning actually works. Evidence indicates that indiscriminate digital expansion has weakened learning environments rather than strengthened them.'
- A February 2025 study by researchers at Microsoft and Carnegie Mellon University, reported by Fortune, found AI use was associated with measurably worse judgment and critical thinking skills — among adult professionals, not just students.
- Dr. Michael Gerlich's 2025 peer-reviewed study in the journal Societies documented cognitive offloading effects across 666 participants, finding a statistically significant association between heavy AI tool use and reduced performance on critical-thinking tasks once the AI was removed.
92% of Students Use AI. Most of Them Aren't Cheating.
In 2023, 43% of college students admitted using AI tools in their coursework. By 2026, 92% use AI in their studying, based on the Higher Education Policy Institute's survey data and analysis published by plagiarismcheck.org drawing on Anson Alexander's 2026 research. That jump — from 43% to 92% in three years — is the fastest adoption curve of any academic tool in modern educational history.
The 92% figure conceals an important distribution. Only 18% of students use AI to complete and submit unedited work as their own — a figure remarkably close to the 17% who texted answers via phone in 2012. Academic dishonesty rates appear stubbornly stable across generations regardless of the available technology. The remaining 74% who use AI for legitimate purposes — research assistance, concept explanation, essay feedback, study planning — are not cheating. The question for them is whether they are using AI in a way that builds or erodes their thinking.
400 Studies, 50 Countries — The Brookings Verdict
The most comprehensive research synthesis on this topic is a Brookings Institution report published January 2026, covered extensively by Fortune. Researchers at the Brookings Center for Universal Education analyzed over 400 academic studies and conducted interviews and focus groups with more than 500 educators, parents, and students across 50 countries. Their conclusion, stated plainly: the risks of deploying generative AI in children's education currently outweigh its benefits.
Mary Burns, an education consultant and co-author of the Brookings report, told Fortune that the cognitive offloading AI enables — and the cognitive decline associated with it — is real, measurable, and affecting the development of critical thinking, reading proficiency, and knowledge of basic facts. The report's framing is important: its authors are not arguing AI is inherently harmful to learning. They are arguing that the current default deployment — unsupervised, unrestricted, without intentional pedagogical design — is producing negative outcomes at scale across the full diversity of educational contexts they studied.
What Cognitive Offloading Actually Does to a Developing Mind
Cognitive offloading — outsourcing a mental task to an external tool — is not new or inherently harmful. You use it when you write a grocery list instead of memorising it. The concern in the current research is not offloading generally, but offloading in domains where the practice of the task is itself the learning. Writing is not just a way to record thoughts that already exist. Constructing an argument in writing is the mechanism through which the argument becomes understood at depth. Outsource the construction before the understanding forms, and the understanding never forms.
- Retrieval practice — trying to recall information from memory — is among the most robustly documented learning mechanisms in cognitive science. When AI retrieves and presents information for you, the retrieval attempt is eliminated. That missing attempt is not neutral: retrieval effort, even when unsuccessful, is what encodes information into long-term memory.
- Error correction learning — recognizing a mistake, understanding why it is wrong, and building the corrected understanding — is how conceptual knowledge develops beyond surface familiarity. When AI produces error-free output, the entire error-correction cycle disappears from the learning process.
- Productive struggle — the frustration and persistence required to work through a genuinely difficult problem — builds cognitive resilience and tolerance for complexity. Students who consistently bypass this by asking AI to solve hard problems show measurably lower frustration tolerance in tests, job interviews, and any professional context where AI is unavailable.
- Anthropic researchers Judy Hanwen Shen and Alex Tamkin studied 52 engineers learning an unfamiliar Python library for the first time. Engineers who fully delegated to AI produced working code — but scored 17% lower on conceptual quizzes afterward, the equivalent of nearly two letter grades. They could not debug what the AI had written. These were people with existing programming expertise, and they still underperformed their no-AI counterparts significantly. For a student encountering a concept for the first time, the substitution is more complete: there is no existing understanding to compare the AI output against.
The Surprising Side: AI Can Also Make You Sharper
The MIT study, Senate testimony, and Brookings report represent the concerning half of the research picture. They are real, should be taken seriously, and point to a genuine pattern. They are not the whole story. In February 2026, Pearson published an analysis of nearly 80 million student interactions with AI study tools embedded in digital textbooks. A single intentional AI study-tool interaction increased the probability of a student exhibiting active reading behavior by three times. Repeated use pushed that to 3.5 times. When that same AI tool was embedded inside instructor-led courseware rather than a standalone eTextbook, a single interaction raised the likelihood of active reading by 23 times — a figure that has not received the attention it deserves. A Swansea University study with over 800 participants found AI augments rather than replaces human creativity — but only when the human directs the process rather than passively accepting AI output.
The pattern across both the negative and positive studies is unmistakable: AI's effect on cognition is not determined by the tool. It is determined by the role the human plays in the interaction. The same interface, the same model, the same capability — produces opposite cognitive effects depending on one variable that most AI-in-education debates skip entirely.
The One Variable That Determines Everything
Active use means: the human attempts the task first, encounters difficulty, and then uses AI to diagnose a specific error, understand a specific gap, or challenge a weakness in their reasoning. The AI responds to the human's thinking. Passive use means: the human presents the task to AI before attempting it, accepts the output, and submits or applies it. The AI replaces the human's thinking.
This distinction is structural, not about effort or integrity. Active use preserves the cognitive struggle that learning requires. Passive use eliminates it. Only active use produces the brain connectivity that the MIT EEG data measured in the high-engagement no-tools group.
| Usage Pattern | Effect on Cognitive Development | Research Basis |
|---|---|---|
| Submit task to AI before attempting it yourself | Reduces cognitive engagement; weakens long-term retention and independent problem-solving capacity | MIT EEG study (arXiv 2025); 83% recall failure in ChatGPT group |
| Attempt task first, then use AI to evaluate your work | Preserves the learning cycle — improves error recognition, retention, and cognitive resilience | Pearson 80M interactions (Feb 25, 2026); Swansea creativity study (Nov 2025) |
| Use AI to explain a concept you're stuck on | Accelerates genuine understanding when used after a real attempt — not as a substitute for it | Consistent with retrieval-practice literature across cognitive science |
| Use AI to generate practice questions for self-testing | Significantly improves learning outcomes — activates deliberate practice at scale on demand | Active learning research; consistent across education science literature |
| Use AI to draft essays or assignments before writing yourself | Bypasses writing-as-thinking; weakens argument development, ownership, and long-term recall | MIT cognitive debt study; Shen & Tamkin 2026 Anthropic skill-formation study |
| Use AI to surface research sources | Beneficial for discovery when primary sources are then read; harmful when AI summary replaces reading | Neutral to positive as a discovery tool; negative when synthesis replaces source engagement |
The Cognitive Debt Loop — Why This Gets Worse Without Design
AI context windows expanded from 512 tokens in 2017 to 2 million tokens by 2026 — a nearly 4,000-fold increase. Over the same period, measurable human sustained-attention capacity has declined, based on longitudinal behavioral data through 2020. As AI becomes more capable, the cognitive threshold at which people choose to delegate drops. As delegation increases, the cognitive practice that maintains independent thinking capacity decreases. The loop reinforces itself: AI gets better, humans offload more, humans become less capable of the tasks AI now handles, which deepens reliance on AI further.
The Brookings researchers, the MIT neuroscientists, and the Pearson learning scientists are all making the same underlying point from different angles: the tool is not the problem. The absence of pedagogical design around the tool is the problem. A student with a deliberate protocol for how they use AI is building a cognitive superpower. A student without one may be incurring a debt they will not notice until the AI is unavailable — in an exam room, a job interview, or any professional situation that requires thinking they have not practiced.
Here Is What Most Reviews of This Research Won't Tell You
The Brookings study, Horvath's Senate testimony, and the MIT findings are all describing a school system problem, not fundamentally an AI problem. Evidence that cognitive decline correlates with indiscriminate digital tools in classrooms long predates ChatGPT. Horvath's data shows the inflection point around 2010 — the year iPads entered classrooms at scale, seven years before ChatGPT existed. AI made the pattern more visible and accelerated it. It did not originate it. If educators had banned calculators in 1985 instead of redesigning math curricula to build understanding alongside computation, we would be having the same conversation today about arithmetic atrophy. The research does not say AI is the enemy of learning. It says undesigned deployment of any powerful cognitive tool produces the same outcome: offloading without anchoring. The fix is not less AI. It is better design.
What to Actually Do About It
- Enforce the 'attempt first' rule — before asking AI anything, spend at least five minutes working on the problem independently. Even an unsuccessful attempt activates the frontal-parietal networks that the MIT EEG data identifies as protective against cognitive debt. The attempt, not the correct answer, is what matters neurologically.
- Test your own understanding after AI assistance — after AI explains something or reviews your work, close the interface and explain the concept in your own words. If you cannot, the AI created the illusion of understanding, not understanding itself.
- Use AI to generate tests, not answers — prompt AI to create practice questions at your level, then attempt them without AI open. Self-testing under difficulty is the single most evidence-backed learning practice in cognitive science. AI makes unlimited personalised practice available on demand — a capability that had never existed before, and one that actively strengthens cognition when used this way.
- For parents: ask 'explain to me how you solved this' instead of 'did you use AI?' A child who explains their reasoning clearly is using AI as a learning amplifier. A child who cannot explain it used AI as a bypass. The explanation — not the presence or absence of AI — is the evidence of learning.
- For teachers: the Pearson data shows that AI tools designed to require active engagement — asking questions before giving answers, locking responses to source material students must read first — triple the likelihood of active learning behavior in standalone tools, and multiply it by 23 times when embedded in instructor-led courseware. Banned AI versus unrestricted AI is not the only choice. Designed AI is the third option, and the data shows it works.
- For students in high-stakes environments — US college applications, gaokao, A-levels, JEE, professional licensing exams — the AI-versus-no-AI gap surfaces most clearly under timed conditions, in oral assessments, and in any professional role requiring independent judgment. Building genuine capability alongside AI fluency is not optional. It is what separates candidates who can demonstrate their skills from those who relied on AI to simulate having them.
LumiChats Study Mode was built specifically in response to the research in this article. Rather than generating answers to exam questions, Study Mode locks every AI response to the specific pages of your uploaded course materials — NCERT chapters, research papers, textbook PDFs — requiring you to engage with the source content before receiving AI assistance. The Quiz Hub generates practice questions from your actual study materials and shows explanations only after you answer, deliberately activating the retrieval practice that the MIT EEG data identifies as protective against cognitive debt. The design principle is direct: AI that makes you engage with difficulty, not AI that removes it.
The single most protective AI study habit you can build today: after every AI-assisted session, spend five minutes writing without AI — summarising what you just learned, explaining in your own words what you now understand that you did not before, and identifying one thing the AI helped you with that you could now explain without it. This metacognitive close converts passive AI use into active learning — and it is exactly what the MIT brain data predicts will maintain the neural connectivity the ChatGPT group lost.
Frequently Asked Questions
- Is AI actually making students dumber? The MIT EEG study found the ChatGPT group showed up to 55% lower cognitive engagement during essay writing compared to unaided students — and that reduced engagement persisted when they tried to write without AI in a later session. The effect is real and measurable. Students who use AI actively (attempt first, then consult AI) do not show the same pattern.
- Does using ChatGPT reduce brain activity? Yes, according to the MIT Media Lab's 2025 EEG study (arXiv:2506.08872). The LLM user group showed the weakest neural connectivity of all three groups across all four sessions. The specific metric was dDTF signal magnitude — a measure of information flow across brain regions. Brain-only writers showed the strongest connectivity; search engine users were in between.
- What did the MIT study on ChatGPT and brains actually find? The study tracked 54 participants across four sessions. The ChatGPT group showed the weakest brain connectivity, the lowest essay quality on linguistic analysis, the lowest ownership scores, and the poorest recall — with 83% unable to accurately quote arguments from essays they had just written. When switched to writing without AI, their brain engagement did not recover to the level of students who had practiced unaided throughout.
- Can AI make you smarter? Yes — when it requires active engagement rather than replacing it. Pearson's February 2026 analysis of nearly 80 million student interactions found that a single intentional AI study-tool interaction tripled the likelihood of active reading behavior. When embedded in instructor-led courseware, a single interaction raised that likelihood by 23 times. The Swansea University creativity study (800+ participants) found AI augments human creativity when the human directs the process.
- What is cognitive offloading and why does it matter for students? Cognitive offloading means delegating a mental task to an external tool. The problem is offloading in domains where the practice of the task is the learning mechanism itself. When a student asks AI to write their essay, they bypass the writing-as-thinking process that is supposed to form the understanding. The output exists; the learning does not.
- How should students use AI without losing their ability to think? Three rules from the research: (1) Attempt the task yourself before consulting AI. (2) After AI explains something, close the interface and explain it back in your own words — if you cannot, the understanding did not transfer. (3) Use AI to generate practice questions and tests, not to generate answers. These three patterns appear consistently in the positive-outcome research and avoid the cognitive debt patterns in the negative-outcome studies.
