Will AI Make Students Dumber or Smarter?

Will AI Make Students Dumber or Smarter? What the Research Shows Now

By |

Will AI make students dumber or smarter? I’ve asked myself that more times than I can count this year, usually right after watching a teenager finish a “500-word essay” in ninety seconds flat. My honest answer, after reading through the actual studies instead of the hot takes, is that both things are true at once, depending on how a student uses the tool sitting in front of them.

That is not a dodge. It is what the data actually says, and it is a far more useful answer than the two camps shouting past each other on social media.

Snapshot: Will AI Make Students Dumber or Smarter

  • Nearly universal use: 95% of UK undergraduates now use AI in some form, and 94% use it for assessed work, according to the Higher Education Policy Institute’s 2026 survey.
  • The strongest evidence against AI comes from an MIT Media Lab study that found ChatGPT users showed the weakest brain connectivity of any group tested while writing essays, a pattern the researchers called “cognitive debt.”
  • The strongest evidence for AI comes from a Harvard physics study where students using a purpose-built AI tutor learned more than twice as much, in less time, than students in an active-learning classroom.
  • A large randomized trial in Turkish high schools found that giving students unrestricted ChatGPT access raised their practice scores but left them 17% worse off on a real test once the tool was taken away. A version with teaching guardrails avoided that damage.
  • The difference is not whether AI is used. It is whether the tool is designed, and used, to make a student think harder or to let them think less.

Why This Question Won’t Go Away

Ask any parent, teacher, or student under twenty-five, and you’ll get a strong opinion on whether AI makes students dumber or smarter.

That is because the technology moved into classrooms faster than anyone could study it properly. Pew Research Center found that 54% of US teens had used an AI chatbot to help with schoolwork by late 2025, and one in ten said they do all or most of their schoolwork with a chatbot’s help.

At the same time, 59% of those same teens believe cheating with AI has become a normal part of student life now, not a rare exception.

Zoom out to higher education and the numbers get even more dramatic. According to HEPI’s Student Generative AI Survey 2026, 95% of full-time UK undergraduates report using AI in at least one way, up from just 66% two years earlier, and 94% now use generative AI to help with assessed work.

In the US, a 2025 Inside Higher Ed survey found 85% of college students had used AI for coursework in the past year, and at Harvard the figure was 87.5% as early as spring 2024. This is not a niche habit anymore. It is close to the default way students study.

So the real question was never “will students use AI.”

They already do, in overwhelming numbers, whether their school allows it or not. The real question is what that usage is doing to their heads, and here the research actually gives us something concrete to work with, not just vibes.

The Case That AI Is Making Students Dumber

Let’s start with the study that set off this entire debate: MIT Media Lab’s “Your Brain on ChatGPT,” led by research scientist Nataliya Kosmyna and published in 2025.

Researchers split 54 participants from five Boston-area universities into three groups and had them write SAT-style essays across several sessions over four months. One group used ChatGPT, one used a search engine, and one used nothing but their own heads.

The EEG results were stark. Brain-only participants showed the strongest, most widely distributed neural connectivity of the three groups. The search engine group came in second.

The ChatGPT group showed the weakest connectivity in the networks tied to memory, attention, and analytical thinking, and they struggled more to recall what they had just written minutes earlier. Two English teachers who graded the essays blind called the ChatGPT group’s writing “soulless.”

The researchers coined a term for this pattern: cognitive debt, meaning the tool spares your brain effort today at a cost you pay later, in weaker recall and shallower thinking.

Here’s a detail that rarely makes the headlines but matters enormously: when researchers switched a subset of participants from ChatGPT to brain-only writing in a fourth session, those students still underperformed the group that had never used AI at all, while students who went the other way (brain-only first, ChatGPT second) actually showed stronger recall and broader brain activity than students who leaned on the tool from the start.

Order matters. Struggling first and outsourcing later is a very different experience for your brain than outsourcing first and never struggling at all.

This isn’t an isolated finding. A separate study by Michael Gerlich at SBS Swiss Business School, surveying 666 people across age groups in early 2025, found a significant negative correlation between how often someone used AI tools and how well they scored on critical-thinking measures, with cognitive offloading, meaning letting the tool do the thinking you’d normally do yourself, as the connecting mechanism.

And a joint Microsoft Research and Carnegie Mellon University study of 319 knowledge workers found that the more confidence someone had in an AI tool’s output, the less critical thinking they applied to checking it.

Self-confidence in your own ability worked the opposite way: it correlated with more critical scrutiny of the AI’s answers, not less.

Then there’s the study that should worry anyone advocating for AI tutoring without guardrails. Hamsa Bastani and Osbert Bastani of the University of Pennsylvania, along with several coauthors, ran a randomized controlled trial with almost 1,000 high school students in Turkey.

One group practiced math with a plain ChatGPT interface (“GPT Base”). A second group used a version built with teaching safeguards that gave hints instead of answers (“GPT Tutor”). A third group, the control, used only their textbooks.

While they had access to the tools, both AI groups crushed the control group on practice problems: 48% better scores for GPT Base, 127% better for GPT Tutor. But when the researchers took the tools away and gave everyone a real test, the story flipped. Students who had used the unrestricted GPT Base scored 17% worse than students who never had AI access at all.

They had learned to lean on the tool, not on the math. The guardrailed GPT Tutor group avoided most of that damage, because it was built to make students do the reasoning themselves.

The Case That AI Is Making Students Smarter

Now for the other side, because it is just as real. At Harvard, physics lecturer Gregory Kestin and senior lecturer Kelly Miller ran a randomized study with 194 undergraduates in an introductory physics course.

Each student experienced one week with a custom AI tutor at home, built specifically around research-backed teaching methods, and one week with a typical active-learning classroom lecture, with the conditions flipped for the second half of the study.

The AI-tutored week produced more than double the learning gains of the classroom week, in less time, with students also reporting higher engagement and motivation. Kestin was careful about the takeaway. He titled the paper “AI Tutoring Outperforms Active Learning,” but he told reporters afterward that this isn’t an argument for replacing human teachers.

It’s evidence that a tutor engineered to withhold full answers, give feedback in short bursts, and force students to attempt problems before explaining anything can outperform even a well-designed lecture, at least for this kind of material.

Khan Academy’s Khanmigo tells a messier but still promising story. Khan Academy has cited internal learning gains of up to 23% in mathematics for students who actually use the tool, though as of early 2026 there was still no published, peer-reviewed study proving Khanmigo itself improves learning outcomes at scale, and only about 15% of students with access were using it regularly, which the company has been trying to fix through redesign.

A separate mixed-methods university study comparing Khanmigo to a plain Google search for physics concepts found real learning gains in every condition, but no statistically significant difference between using Khanmigo and just searching Google, at least over the short exposure window the study measured.

Students liked Khanmigo’s step-by-step guidance, but liking a tool and learning more from it aren’t the same thing.

Beyond individual tools, Brookings Institution’s Center for Universal Education spent over a year interviewing hundreds of educators and parents and reviewing more than 400 research articles for its 2026 report on AI and children’s education.

Their conclusion was that AI can genuinely expand access to learning opportunities, particularly for students with disabilities, neurodivergent learners, and multilingual students, when it’s deployed thoughtfully and paired with adequate school infrastructure. That’s a real, documented upside, not just AI-industry marketing.

The Real Pattern Hiding Inside All These Studies

Once you line these studies up side by side, a pattern jumps out that neither the doom camp nor the hype camp likes to admit. AI does not have one effect on learning. It has at least two, and which one you get depends on how the tool is built and how a student is allowed to use it.

The Turkey study is the clearest illustration I’ve found. Same students, same math curriculum, same underlying AI model. One version acted like a search engine that would happily hand over the answer, and it quietly damaged long-term learning once the training wheels came off. The other version was built to behave like a tutor, one that nudges instead of tells, and it protected students from that same damage while still boosting their practice scores. The tool wasn’t the variable. The design of the interaction was.

You can see the same split in the Harvard result. Kestin’s AI tutor wasn’t ChatGPT with a friendly prompt bolted on.

It was fed the same feedback he planned to give students in class, told to stay brief so it wouldn’t overload anyone with information, and instructed to reveal one step at a time rather than the full solution. In other words, it was built to force the struggle that the MIT Media Lab study found is exactly what protects your brain’s neural engagement in the first place. A generic chatbot, used to skip that struggle, produces cognitive debt.

A structured tutor, used to guide through the struggle, produces double the learning.

This lines up with what the Microsoft and Carnegie Mellon researchers found too. Knowledge workers reported using less mental effort on routine, low-stakes tasks with AI’s help, but more effort on high-stakes tasks where accuracy actually mattered to them, because that’s when they bothered to verify the AI’s work instead of taking it on faith.

Students, especially younger ones without the self-discipline of an experienced professional, don’t reliably make that same trade-off on their own.

That’s precisely why the guardrails in tools like GPT Tutor and PS2 Pal (Kestin’s Harvard tool) made such a measurable difference: they built the verification and the struggle into the system instead of hoping the student would supply it themselves.

What This Actually Looks Like in Classrooms Right Now

The numbers on how students use AI in practice back this pattern up. According to HEPI’s 2025 and 2026 surveys, the most common academic uses of AI among UK undergraduates are explaining concepts, summarizing articles, and suggesting research ideas, not generating finished essays wholesale.

Only 12% of students in the 2026 survey admitted to including AI-generated text directly in assessed work, up from 8% the year before but still a minority.

Turnitin’s own research found that 59% of students are themselves worried AI could weaken their critical thinking, and 49% worry about becoming overly dependent on it. Students, in other words, are more clear-eyed about the risk than a lot of the public debate gives them credit for.

Age matters too. Pew found that older teens lean on AI for schoolwork more than younger ones: about a quarter of 11th and 12th graders who’d heard of ChatGPT said they’d used it for schoolwork, compared with 17% of 9th and 10th graders and 12% of 7th and 8th graders.

And acceptability tracks the type of task closely.

In Pew’s surveys, over half of teens think it’s fine to use AI to research a new topic, while fewer than one in five think it’s acceptable to use it to write an essay outright. Students, it turns out, already sense the difference between using AI to think through something and using AI instead of thinking.

So, Does AI Make Students Smarter or Dumber?

Honestly, I think the framing itself is part of the problem. Asking “does AI make students smarter or dumber” is a bit like asking whether a car makes you fitter or lazier. It depends entirely on whether you’re driving to the gym or driving instead of walking to the mailbox.

The MIT Media Lab team wasn’t wrong that unrestricted, answer-first AI use produces measurably weaker brain engagement and shallower recall.

Kestin’s Harvard team wasn’t wrong that a carefully engineered AI tutor can double how much a student learns in the same amount of time. Both studies are true, and both are about the same underlying technology.

What worries me more than any single statistic is the direction the guardrails are heading. Most students aren’t using purpose-built tutors like PS2 Pal or GPT Tutor. They’re using the same general-purpose chatbot everyone else uses, the one that will happily hand over a finished paragraph if you ask nicely, because as the researchers behind the Khanmigo case studies put it, students don’t care about Socratic method or pedagogical guardrails.

They want answers. That gap between what the best available research says works and what students actually reach for is, to me, the real story here, more than any single number about brain connectivity or test scores.

If you’re a student reading this, my honest, unglamorous advice is the same lesson every one of these studies points to: struggle first, then check your work against AI, not the other way around. Write your own draft before you ask a chatbot to improve it. Attempt the practice problem before you ask for a hint.

The MIT researchers found that students who wrote unaided first and brought in ChatGPT second retained more and showed broader brain activity than students who did it backward. That single habit, struggle first, verify second, is the difference between a tool that sharpens your thinking and one that quietly erodes it.

For readers who want to go further into how AI is reshaping our relationship with our own thinking, I’ve also reviewed Jamie Bartlett’s How to Talk to AI, which digs into what these models are actually doing beneath the fluent surface, and Kai-Fu Lee’s AI Superpowers, which looks at the labor and economic side of the same technology now sitting inside every classroom.

What Parents and Teachers Can Actually Do With This

None of this research points to banning AI outright, and honestly, given that 95% of undergraduates and more than half of US teens are already using it, a ban would mostly just push the behavior underground where nobody can study or guide it.

The more useful lesson from Brookings’ year-long review of over 400 research articles is that AI’s benefits for students, particularly students with disabilities, neurodivergent learners, and multilingual learners, show up when the technology is deployed deliberately, with real infrastructure and teacher training behind it, not when it’s simply left on a laptop and hoped for the best.

The same report found that AI can free up teachers’ time on grading and lesson prep, time that, redirected well, goes toward the individualized attention that no chatbot can replace.

A few things teachers and parents can borrow directly from the studies above, rather than from general AI-optimism:

  • Ask for the process, not just the product. The Turkey study’s damage came from students using AI to skip straight to answers. Assignments that require a visible attempt, a screenshot of a first draft, or a recorded explanation of reasoning make that shortcut harder to take.
  • Push AI toward hints, not solutions. GPT Tutor and PS2 Pal both worked because they were instructed to withhold the full answer and respond in short, guided steps. Any parent or teacher can give a student that same instruction directly: ask the chatbot for a hint, not the finished paragraph.
  • Watch for the “confidence trap.” The Microsoft and Carnegie Mellon researchers found that trusting AI’s output more led people to check it less. Teaching students to treat every AI answer as a first draft that needs verification, not a finished one, is a habit worth building early.
  • Treat AI literacy as its own subject. Turnitin’s survey found most students are already worried about their own critical thinking and over-reliance, which means the appetite for honest guidance is there. What’s missing, in HEPI’s own words, is institutions catching up with clear policy and staff training fast enough to meet it.

Conclusion

While the question will AI make students dumber or smarter?, is imperative and timely one, the evidence does not support a clean verdict that AI makes students smarter, nor does it support the panic that AI is simply making a generation dumber.

What it supports is something less satisfying to tweet but far more useful to act on: an AI tool built to make you struggle before it helps you tends to build real learning, and an AI tool that hands you the finished answer tends to quietly erode it, even while your grades look fine in the short term.

The technology in both cases can be the exact same underlying model. The difference is the design choice sitting between the student and the answer, and, just as much, the habit the student brings to that choice every single time they open the app.

References

Sopranu

About Sopranu

Sopranu is a Bangladeshi-born writer, lifelong bookaholic and devoted cinephile. He writes reviews, opinions and reflections shaped by close reading, careful research and a love of philosophy. Every piece on this site starts with the same question: what is this work really trying to say, and does it succeed?

Related Posts

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top