Is AI Making Us Dumber? What 30+ Studies Actually Say

Last Updated: July 30, 2026

Thirty-plus studies have now examined what AI is doing to human cognition. The honest answer is more specific than most coverage suggests - and more alarming for one particular type of AI use.

MIT Media Lab researchers measured brain activity in 54 participants across three groups: people writing SAT essays alone, people using Google search, and people using ChatGPT. They used EEG headsets to track neural activity over four months. The least cognitively active brains were in the ChatGPT group. The most active were writing alone. When the ChatGPT group was asked to write without AI assistance, they consistently underperformed at neural, linguistic, and behavioral levels compared to where they started.

That is not anecdote. That is a controlled EEG measurement showing a specific cognitive cost from a specific type of AI use. The paper, published on arXiv in June 2025, calls the effect "accumulation of cognitive debt."

At the same time, a Harvard University physics study found that students using AI tutors learned more than twice as much in less time compared to traditional classroom instruction. A Harvard Business School randomized controlled trial of 758 consultants found AI users completed 12.2% more tasks, 25.1% faster, with 40% higher quality.

Both sets of findings are real. They are measuring different things. And the distinction between them is the entire answer to whether AI is making us dumber.

🎯 Before you read on - we put together a free 2026 AI Tools Cheat Sheet covering the tools business leaders are actually using right now. Get it instantly when you subscribe to AI Business Weekly.

Get the Free Cheat Sheet

The evidence for cognitive costs from AI use has accumulated quickly in 2025-2026. The research comes from multiple independent teams using different methodologies and arriving at consistent findings.

The MIT Media Lab brain activity study (2025):

Kosmyna et al. at MIT Media Lab measured neural engagement across three essay-writing conditions over four months using EEG. Participants who used ChatGPT showed the lowest brain activity, the weakest memory retention, and the lowest sense of ownership over their work. When reassigned to write without AI, they struggled to recall or quote their own previous essays. The brain-only group showed the strongest cognitive engagement throughout.

As Nextgov/FCW reported on the findings: participants who exclusively used AI to write essays showed weaker brain connectivity, lower memory retention, and a fading sense of ownership over their work. And even after they stopped using AI, the effects lingered. The study is a preprint and not yet peer-reviewed - an important caveat - but its findings are consistent with every other study in this category.

The Microsoft knowledge worker survey (319 participants):

Microsoft's own research found that the more workers used AI to handle cognitive tasks, the less critical thinking they reported doing - and the less confidence they had in their own independent judgment over time. The researchers described a feedback loop: AI handles the thinking, you practice it less, you become less confident doing it, you defer to AI more. Source: Scott et al., Microsoft Research, via Prof G Media

The Gerlich study (666 participants, 2025):

A mixed-methods study across diverse age groups found heavy AI tool use was strongly associated with decreased critical thinking ability. The mechanism: cognitive offloading - delegating reasoning to AI - reduces the practice of reasoning, which reduces the skill. The study identified what the author calls "cognitive laziness" - a decline in inclination to engage in deep, reflective thinking. Younger participants showed stronger effects than older ones. As cited directly in the International AI Safety Report 2026: a study with 666 participants found that heavier AI tool use was strongly associated with lower scores on a self-assessment scale related to critical-thinking behaviours, mediated by cognitive offloading.

The Brazilian retention study (FGV/UFRJ, 2025):

A preregistered randomized controlled trial measured knowledge retention at 45 days between AI-assisted and non-AI-assisted learners. The gap: 11 percentage points. Students who used AI to work through problems learned faster but retained significantly less six weeks later. The Cognitive Divergence arXiv paper cites this RCT as one of the strongest methodological demonstrations of the practice-loss mechanism in AI-assisted learning.

The clinical accuracy finding:

The International AI Safety Report 2026 cites a study finding that three months after the introduction of AI support, clinicians' ability to detect tumors without AI assistance had dropped by 6%. AI-assisted diagnostic accuracy improved. Unassisted accuracy declined. The skill atrophied when not practiced - the same mechanism as every other finding in this section, applied to a high-stakes professional context.

The HBS/BCG performance paradox:

The Harvard Business School and Boston Consulting Group research found AI boosts short-term performance by 14-40% while simultaneously eroding critical thinking and independent judgment. The researchers pointed out that by automating routine tasks and leaving exception handling to human users, AI deprives users of the routine opportunities to practice their judgment and strengthen their cognitive musculature. You get faster and worse at thinking simultaneously.

The national security concern:

The Council on Strategic Risks launched a formal debate series in 2026 specifically about whether AI is degrading critical thinking in the national security workforce. As Prof G Media covered in May 2026, the concern is direct: if cognitive offloading degrades critical thinking, and the national security workforce depends on clear thinking, rapidly deploying AI tools may create a security risk rather than a capability advantage. When an institution responsible for national security runs formal debates about whether its productivity tools are making its people worse at their jobs, that is worth noting.

For our complete data on AI accuracy limitations and where AI systems fail, our AI hallucination statistics guide covers what AI gets wrong and why human verification remains essential.

The same body of research also contains evidence that AI improves cognitive outcomes - and the conditions under which this happens are equally specific.

The Harvard physics study (2025):

Students using AI tutors learned more than twice as much in less time compared to those in traditional active-learning classrooms. This is not a marginal finding - it matches Bloom's 1984 result showing one-on-one human tutoring produces the largest learning gains of any educational intervention. AI tutoring achieved equivalent results at essentially zero marginal cost per student. The mechanism: AI provides immediate, personalized, patient feedback on exactly the mistake a student just made. That feedback loop is what drives learning - and AI enables it at scale. Source: Kestin et al., Scientific Reports, June 2025. Covered in full in our AI education statistics guide.

The HBS consultant RCT (758 participants):

A randomized controlled trial with 758 consultants found AI users completed 12.2% more tasks, 25.1% faster, with 40% higher quality scores. One of the most rigorous AI productivity studies available because of its scale and randomized design. Source: Dell'Acqua et al., Harvard Business School. Full productivity data in our AI productivity statistics guide.

The UTS structured AI use finding (March 2026):

The University of Technology Sydney's March 2026 report found that structured AI use - where AI handles grammar and formatting while the human handles the reasoning - does not show the same cognitive costs as unstructured AI use. This is the most actionable single finding in the entire literature. The problem is not the tool. It is which part of the thinking the tool is doing.

Every study showing cognitive harm and every study showing cognitive benefit is measuring a different type of AI use. The research is not contradicting itself. It is describing two different relationships with the technology.

As the AI Tool Discovery synthesis of 30+ studies puts it: the problem is not AI - it is offloading the specific mental work that builds the skill. If AI writes for you, you produce better output and weaker skills. If AI corrects your grammar, you produce better output and your skills are unchanged.

The mechanism is straightforward: cognitive skills are built through practice. Skills you do not practice atrophy. If AI handles the reasoning component of a task, you practice reasoning less. If AI handles the formatting component, your reasoning practice is unchanged.

This maps cleanly onto every finding in the literature:

The GPS analogy holds. University of Melbourne researchers note that decades of research confirm over-reliance on GPS reduces spatial memory and navigation skills in passive users while active users - those who think about the route while using GPS - maintain and sometimes improve their skills. Same technology. Different cognitive outcome depending entirely on whether the user is exercising or outsourcing their judgment.

The research consistently shows that younger users are more affected than older users. The explanation is developmental: people over 46 who use AI heavily already have established cognitive frameworks. They are applying existing skills with AI assistance. People still developing those frameworks - students, early-career professionals, anyone in a steep learning phase - risk building a dependency rather than developing independent capability.

The MIT Media Lab study specifically flagged this concern for younger users: as NPHIC reported, the findings highlight the need for particular caution in public health and education sectors where younger developing minds are most affected.

This is the most significant implication of the research for education and professional development. A 22-year-old using AI to write every first draft may never develop the writing skills that require working through the difficulty of structuring an argument independently. A 45-year-old using AI to write first drafts is applying a skill they already have.

The risk is not uniformly distributed. It concentrates in learning phases and in high-stakes professional domains where independent judgment cannot be replaced. For the full picture of what AI means for education and the learning outcomes research, our AI education statistics guide covers every finding.

For what AI means for teachers and the education profession specifically, our will AI replace teachers guide covers the employment and practice data.

Is AI making us dumber?

For people using AI in an unstructured way - where AI reasons and synthesizes on their behalf - the evidence shows specific and measurable cognitive costs: declining brain activity in repeated sessions per MIT Media Lab, an 11-point knowledge retention gap at 45 days per the FGV/UFRJ RCT, and a 6% drop in unassisted clinical skill after three months of AI-assisted practice per the International AI Safety Report 2026.

For people using AI in a structured way - where AI handles logistics, formatting, and grammar while the human handles reasoning, judgment, and argument - the same research shows no cognitive costs and measurable productivity gains.

When Harvard Gazette asked ChatGPT whether AI can make us dumber or smarter, it answered: "It depends on how we engage with it - as a crutch or a tool for growth."

That is correct. And it is the most important thing anyone using AI tools professionally needs to understand about how to use them without losing the skills that make their judgment worth having.

Three principles that follow directly from the research:

Use AI downstream of your thinking, not instead of it. Write the outline yourself. Form the argument yourself. Use AI to improve what you have written, not to replace the act of writing it. The HBS finding - 40% higher quality from AI-assisted consultants - came from consultants who were still doing the reasoning. The MIT finding - declining brain activity - came from participants who outsourced the reasoning entirely. The structured prompting approach that keeps your reasoning in the loop is covered in our how to write better AI prompts guide.

Protect deliberate practice in learning phases. If you are building a skill, the cognitive cost of AI assistance is highest. A medical student who uses AI to reason through every differential diagnosis is building AI dependency, not diagnostic competence. Use AI to check your reasoning after you have done it independently, not before.

Maintain unassisted practice in high-stakes domains. The 6% drop in unassisted tumor detection after three months of AI-assisted reading is the finding that should shape policy in medicine, law, finance, and any professional domain where independent judgment under pressure matters. AI improves assisted performance. It may degrade unassisted performance. In most situations that tradeoff is acceptable. In emergency situations where AI is unavailable, it is not.

The ROI data at the enterprise level confirms the same pattern - organizations that use AI to augment human judgment outperform those that use it to replace it. Our AI ROI statistics guide covers the financial returns data showing why the high performers are those who kept humans in the reasoning loop.

AI Education Statistics 2026
The full data on AI in schools - the Harvard physics study, student adoption, and the academic integrity picture.

AI Productivity Statistics 2026
The productivity gains data - what AI does to output quality and work speed when used correctly.

AI Hallucination Statistics 2026
Why AI gets things wrong - the accuracy limitations that make human judgment irreplaceable.

Will AI Replace Teachers? The 2026 Data
The employment data - what AI in education means for teaching as a profession.

How to Write Better AI Prompts
The structured AI use approach - prompting that keeps your reasoning in the loop.

AI ROI Statistics 2026
The productivity and return data - when AI helps and when it stalls.

Is AI making us dumber?
The honest answer depends on how you use it. For people using AI in an unstructured way - where AI does the reasoning, writing, or problem-solving on their behalf - controlled studies show measurable cognitive costs: declining brain activity in repeated EEG sessions per MIT Media Lab's 2025 study of 54 participants, an 11-point knowledge retention gap at 45 days per the FGV/UFRJ preregistered RCT, and a 6% drop in unassisted clinical skill after three months of AI-assisted practice per the International AI Safety Report 2026. For people using AI in a structured way - where AI handles grammar, formatting, and logistics while the human handles reasoning and judgment - the same research shows no cognitive costs and measurable productivity gains per the UTS March 2026 report.

What does the MIT study say about AI and brain activity?
MIT Media Lab researchers (Kosmyna et al., June 2025) measured brain activity via EEG in 54 participants divided into three groups: ChatGPT users, Google search users, and a brain-only group using no tools. All groups wrote SAT essays over four months. The ChatGPT group showed the lowest neural activation, the weakest memory retention, and the lowest sense of ownership over their work. When reassigned to write without AI, they consistently underperformed compared to their own baseline. The brain-only group showed the strongest cognitive function throughout. The study is a preprint and not yet peer-reviewed, but its findings are consistent with multiple independent studies across different methodologies and populations.

What does the research say about AI and critical thinking?
Multiple independent studies show a negative relationship between unstructured AI use and critical thinking. Gerlich 2025 surveyed 666 participants and found heavy AI tool use strongly associated with lower critical thinking scores, mediated by cognitive offloading. Microsoft's survey of 319 knowledge workers found more AI use correlated with less critical thinking and lower confidence in independent judgment. The HBS/BCG study found AI boosts short-term performance 14-40% while simultaneously eroding critical thinking and independent judgment. The International AI Safety Report 2026 cited these findings as a documented systemic concern. All studies use the same explanatory mechanism: cognitive skills are built through practice, and AI use that replaces practice reduces the skill.

Does AI help or hurt learning?
Both, depending on how it is used. The Harvard University physics study (Kestin et al., Scientific Reports, June 2025) found students using AI tutors learned more than twice as much in less time - a 2-standard-deviation improvement matching the best educational interventions ever measured. The FGV/UFRJ preregistered RCT found AI-assisted students showed 11-point lower retention at 45 days compared to non-AI learners. The key distinction: the Harvard study used AI as a tutor that prompted students to reason through problems themselves. The FGV study used AI as a tool that provided answers. AI that teaches you to reason improves learning. AI that reasons for you reduces retention.

What is cognitive offloading and why does it matter?
Cognitive offloading is the delegation of mental tasks to an external system - historically to notebooks, calculators, and GPS. AI enables cognitive offloading at unprecedented scale and breadth. The concern, cited directly in the International AI Safety Report 2026, is that cognitive offloading can free up cognitive resources in the short term while producing long-term effects on the development and maintenance of cognitive skills. The GPS analogy is the best-documented prior example: passive GPS use reduces spatial memory over time while active GPS use - where the user thinks about the route while navigating - does not show the same effect. The technology is the same. The cognitive outcome depends on whether the user is exercising or outsourcing their judgment.

Are younger people more affected by AI's cognitive risks?
Research consistently shows younger users are more affected by AI cognitive offloading than older users. The MIT Media Lab study specifically flagged younger developing minds as a particular concern. The Gerlich 2025 study of 666 participants found stronger effects in younger age groups. The explanation is developmental: people who already have established cognitive frameworks are applying existing skills with AI assistance and face lower risk. People still developing those frameworks risk building competency that depends on AI rather than developing independent capability. The practical implication is direct: students and early-career professionals face the highest cognitive cost from unstructured AI use.

How should I use AI without losing my cognitive skills?
Three evidence-based principles from the research. First, use AI downstream of your thinking: form your own outline, argument, or diagnosis first, then use AI to improve, check, or expand it. The HBS productivity gains came from consultants who were still doing the reasoning. Second, protect deliberate practice in learning phases: if you are building a skill, the cognitive cost of AI assistance is highest during that phase. Use AI to verify your reasoning after completing it independently, not before starting. Third, maintain unassisted practice in high-stakes domains: the 6% drop in unassisted clinical skill after three months of AI-assisted practice suggests professionals in domains requiring independent judgment under pressure need to regularly practice without AI to maintain that skill.

Thirty-plus studies and the honest answer is two sentences.

AI use that replaces cognitive practice produces cognitive atrophy in the specific skills being replaced. AI use that augments cognitive practice without replacing it produces better outputs without cognitive cost.

The MIT Media Lab data, the FGV retention gap, the clinical accuracy decline, and the Microsoft critical thinking survey are all measuring the same phenomenon in different contexts. Cognitive skills built through practice atrophy when AI handles the practice for you.

The Harvard physics result, the HBS consultant productivity data, and the UTS structured AI use finding are measuring the other phenomenon. AI that gives better feedback, removes logistical friction, and accelerates output without replacing reasoning produces genuine gains.

Every professional using AI tools daily is making an implicit choice between these two modes every time they open a prompt. The research tells you which one protects the judgment that makes your work worth doing. The rest is execution.