AI Study Tool Statistics: What the Data Shows

The headline numbers

Most writing about AI and studying is either breathless or panicked. The AI study tool statistics below are more interesting than either mood. Here are the ones worth remembering, each unpacked in the sections that follow.

Two things jump out. Adoption is nearly universal among students, and the evidence on outcomes is genuinely split: the same technology produces two-year learning gains in one study and measurable cognitive decline in another. The difference, as you'll see, is almost entirely about how the tool gets used.

How many students actually use AI to study

Start with teenagers, because they moved first. Pew Research has tracked the same question since 2023, and the trend line is steep. About 13% of U.S. teens who'd heard of ChatGPT had used it for schoolwork in 2023. By the 2024 survey that had doubled to 26%. In Pew's most recent report, released in February 2026 and based on interviews with teens ages 13 to 17 conducted in late 2025, 54% said they'd used a chatbot of some kind to help with schoolwork. One in ten now say they do all or most of their schoolwork with a chatbot's help.

Older students lean on it more. Among 11th and 12th graders, 31% report using ChatGPT for schoolwork, compared with 20% of 7th and 8th graders. Teens also draw clear lines around what feels fair: 54% say using ChatGPT to research a new topic is acceptable, but only 29% say the same about solving math problems and just 18% about writing essays.

In higher education the numbers are no longer a trend, they're the baseline. The Digital Education Council's Global AI Student Survey, which polled 3,839 bachelor's, master's, and doctoral students across 16 countries in mid-2024, found 86% use AI in their studies. More than half use it weekly and 24% use it daily. The average student juggles 2.1 different AI tools, with ChatGPT the clear leader at 66% adoption, followed by Grammarly and Microsoft Copilot.

There's a catch worth flagging in that same survey: 58% of students said they don't feel they have enough AI knowledge and skills, and about half don't feel prepared for an AI-driven workplace. High usage and low confidence are sitting side by side.

What they use it for, and the scale of ChatGPT

To understand study behavior you have to understand the tool most students reach for. In September 2025 OpenAI published "How People Use ChatGPT," its first large public analysis of usage, drawn from more than a million sampled conversations. The scale is hard to overstate: around 700 million weekly active users when that study ran in mid-2025, a figure OpenAI said had passed 900 million by early 2026, roughly a tenth of the world's adults.

The report broke messages into three intents. Just under half (49%) were "Asking," seeking information or advice. Forty percent were "Doing," asking the model to produce something. The rest were "Expressing." By topic, three categories cover about 80% of everything people ask: practical guidance and writing each around 28%, and seeking information about 21%. Practical guidance is the one that includes tutoring, how-to help, and step-by-step explanations. OpenAI called out that bucket directly, noting about 10% of all messages are requests for tutoring or teaching and naming education "a key use case."

Two other figures give away who's driving this. Nearly half of all messages come from users under the age of 26, and ChatGPT's U.S. traffic dips every summer when school lets out. Similarweb measured U.S. visits falling through May, June, and July 2025, then recovering as students returned in August. The usage curve literally follows the academic calendar.

Practical guidance is the bucket that matters most for learning, and it's one of the two largest categories of everything people ask ChatGPT to do. When students describe using AI to "study," this is usually what they mean: not a specialized flashcard app, but a general chatbot acting as an on-demand explainer.

That distinction matters because a general chatbot has no idea what you already know, no record of what you highlighted last week, and no reason to make you do the cognitive work yourself. The purpose-built tutors we'll look at next are engineered to withhold answers. A default chatbot is engineered to give them. If you want a chatbot to think with rather than for you, prompt design does a lot of the work, which is the subject of Claude vs ChatGPT for learning.

The tools and the money behind them

Beyond the general chatbots, a generation of specialized study apps has bolted AI onto older learning products, and their user numbers show how mainstream this has become.

• Duolingo reported 56.5 million daily active users, 137.8 million monthly active users, and 12.5 million paid subscribers in its Q1 2026 filing, with daily users up 21% year over year. Duolingo Max, its GPT-4-powered tier, is the premium layer on top of that base.
• Quizlet, the flashcard mainstay, reports around 60 million monthly learners and has folded AI into its study sets.
• Coursera passed 197 million registered learners at the end of 2025, and 94% of learners who used its generative-AI "Coursera Coach" said it improved their experience.
• Khan Academy's Khanmigo, built on GPT-4, rolled out to school districts as a guided tutor rather than an answer machine.

The money following these products is substantial, though you should treat any single market-size figure with skepticism because the research firms disagree wildly. Grand View Research projects the AI-in-education market will reach about $32.27 billion by 2030, at a 31.2% compound annual growth rate. Other firms publish very different numbers on different end dates, with some projecting well over $100 billion by the mid-2030s. The precise figure is guesswork. The direction, 30%-plus annual growth, is the consensus.

The same wave has been brutal to the old model. Chegg, the homework-help service, watched its subscriber base fall 31% year over year to 3.2 million in early 2025 and its revenue drop about 30%, declines it blamed directly on free AI tools. Its stock is down roughly 99% from its 2021 peak. Students didn't stop wanting help; they stopped paying for the version a chatbot now gives away.

Here's the uncomfortable part for the specialized apps: OpenAI's hundreds of millions of weekly users dwarf all of them combined. The center of gravity for AI studying is a general-purpose chatbot, not a purpose-built learning tool, which is exactly why the outcome data is so mixed.

Do AI study tools actually work?

This is the question that matters, and to answer it fairly you need a yardstick. The classic one comes from educational psychologist Benjamin Bloom, who reported in 1984 that students given one-to-one mastery tutoring performed two standard deviations better than students in a normal classroom, meaning the average tutored student outscored 98% of the conventionally taught group. Bloom called it the "2 sigma problem" because individual human tutoring is far too expensive to give everyone. AI's implicit promise is to solve that problem cheaply. So does it?

The strongest evidence says yes, when the tool is built well. A randomized study led by Greg Kestin at Harvard, published in Scientific Reports in June 2025, gave about 194 physics students either a normal, highly refined active-learning class or a purpose-built AI tutor called "PS2 Pal." The tutor ran on GPT-4 but was instructed to reveal only one step at a time and never hand over a full solution. Students using the AI tutor scored roughly 30% higher on the follow-up test, learned more than twice as much for the time spent, and reported higher engagement.

The most striking result comes from the developing world. In a World Bank randomized controlled trial in Benin City, Nigeria, about 800 secondary students took six weeks of after-school English sessions where teachers guided them through Microsoft Copilot. The gains were equal to roughly two years of ordinary schooling. The overall effect measured around 0.3 standard deviations, and the program outperformed 80% of education interventions ever rigorously tested in the developing world.

Notice what these winning cases share. The AI was configured to tutor, not to answer. It asked questions, revealed one step at a time, and kept the student doing the thinking. That design choice, not the underlying model, is what produced the gains. Strip it away and the picture changes fast.

The cognitive debt problem

The same technology that drives two-year gains can also quietly make you learn less. The most rigorous evidence is a field experiment by Hamsa Bastani and colleagues at Wharton, published in PNAS in 2025. Nearly 1,000 high-school math students practiced with one of two AI helpers or none. During practice, students using an unrestricted "GPT Base" scored 48% higher and those using a safeguarded, hint-based "GPT Tutor" scored 127% higher. Then the researchers took the AI away for the real exam. The GPT Base group scored 17% worse than students who'd never used AI at all, and the Tutor group landed back at the no-AI baseline. The practice gains were an illusion; the unguarded chatbot had done the learning for them.

A widely discussed 2025 MIT Media Lab preprint led by Nataliya Kos'myna, "Your Brain on ChatGPT," reached the same conclusion through neuroscience. Researchers used EEG to monitor 54 people writing essays with ChatGPT, with a search engine, or with no tools. The brain-only writers showed the strongest, most distributed neural connectivity; ChatGPT users showed the weakest, up to 55% lower. The researchers named the effect "cognitive debt," and it lingered after the AI was taken away.

The preprint's most memorable number: after the first session, 83% of the ChatGPT group could not quote a single sentence from the essay they had just produced, while nearly everyone in the other groups could. That study is small and not yet peer-reviewed, and its own authors warned against "brain rot" headlines, but it points the same direction as the peer-reviewed Wharton result. When the AI does the composing, the words never really pass through your memory. This is the same passive-versus-active split that shows up throughout learning science, and it's why simply reading or generating output feels productive while teaching you little. We dug into the underlying mechanism in the statistics on highlighting and note-taking statistics.

The lesson isn't "AI makes you dumber." It's that offloading the effortful part of thinking, whether to a chatbot or a highlighter you never revisit, removes the very struggle that builds memory. AI study tools work when they add difficulty in the right place and fail when they remove it.

Teachers, the guidance gap, and cheating

Students didn't adopt AI in a vacuum. Their teachers are adopting it too, just more cautiously. In a Gallup survey run with the Walton Family Foundation in spring 2025, 60% of K-12 public school teachers said they'd used an AI tool during the 2024 to 2025 school year. About 30% use AI at least weekly, and those weekly users estimated they save around six weeks of work a year on tasks like lesson prep and grading.

The gap is in guidance, not usage. Gallup found most teachers get no formal direction on how AI should be used for learning, and the same is true for students. That vacuum is where the anxiety about cheating lives. In a College Board survey of more than 3,000 faculty in 2025, 84% agreed that AI reduces students' critical thinking and deep engagement, and 92% worried about AI-facilitated dishonesty. Turnitin, which scans student writing, reported that by late 2025 nearly 15% of English-language submissions were at least 80% AI-generated.

Students feel the tension too. In one survey, 51% of college students said using a tool like ChatGPT on schoolwork counts as cheating, even as large shares admitted using it. The line between "the AI tutored me" and "the AI did it for me" is exactly the line the Wharton and MIT studies measured, and right now most students are drawing it for themselves.

The honest read is that AI in education is running ahead of the norms meant to govern it. The tools are in nearly every backpack. The instructions for using them well are not.

What the statistics mean for how you study

Line the numbers up and a single pattern emerges. Every case where AI clearly helped learning shared one feature: the student stayed the one doing the thinking. The tutor asked instead of told. The student generated, retrieved, and struggled a little. Every case where AI hurt shared the opposite: the tool did the work and the student watched.

That maps precisely onto decades of learning science. Active recall beats rereading. Generating your own words beats copying them. Desirable difficulty builds durable memory. AI doesn't repeal any of that. It just makes it dangerously easy to skip the effortful step, because the answer is always one prompt away.

So the practical move is to use AI to surface material and yourself to process it. Read the source, mark what matters, then put it in your own words. When you highlight an article or PDF with Glasp, you're forcing the encoding step the MIT essay-writers skipped. When you summarize a YouTube lecture, the timestamps and transcript keep you anchored to the real material instead of a hallucinated paraphrase. And when you use Glasp's AI chat to question your own highlights rather than the open web, the AI works from what you actually read, so it deepens your notes instead of replacing them. The same logic applies to Kindle highlights you want to revisit.

For a fuller playbook on turning AI into a thinking partner rather than a crutch, see how AI is reshaping the way we learn and the case for active recall. The statistics all point one way: the tool matters far less than whether you're still the one doing the learning.

Frequently Asked Questions

What percentage of students use AI to study?

It depends on age. Among U.S. teens ages 13 to 17, 54% say they've used a chatbot for schoolwork, per Pew Research in 2026, and 26% use ChatGPT specifically. Among college and graduate students the figure is far higher: the Digital Education Council's 2024 survey of nearly 3,900 students across 16 countries found 86% use AI in their studies, with more than half using it weekly.

Do AI study tools actually improve learning outcomes?

They can, substantially, when the tool is designed to tutor rather than answer. A 2025 Harvard study found students scored about 30% higher with a purpose-built AI tutor, and a World Bank trial in Nigeria measured six weeks of AI tutoring worth roughly two years of schooling. But a 2025 MIT study found that using ChatGPT to write essays reduced brain engagement and left 83% of users unable to quote their own work. The design and the way you use it decide the outcome.

Is using AI to study considered cheating?

There's no universal rule, and that's the problem. Most teachers report getting no formal guidance on acceptable AI use. As a working line: using AI to explain a concept, quiz you, or check your reasoning supports learning, while using it to produce work you submit as your own crosses into academic dishonesty at most institutions. Check your school's specific policy.

How big is the AI education market?

Estimates vary widely because research firms use different definitions. Grand View Research projects roughly $32 billion by 2030 at a 31% annual growth rate. Other firms publish figures ranging into the hundreds of billions on longer horizons. Treat any single number cautiously; the reliable signal is the growth rate, which sits above 30% a year across most estimates.

Which AI tools do students use most?

ChatGPT dominates, used by about two-thirds of surveyed higher-ed students and by roughly 900 million people weekly worldwide as of early 2026. Grammarly and Microsoft Copilot follow. Specialized study apps like Duolingo (56.5 million daily users) and Quizlet (around 60 million monthly learners) have large audiences too, but general-purpose chatbots are where most "studying with AI" actually happens.

Conclusion

The statistics tell a coherent story once you stop treating "AI study tools" as one thing. Adoption is effectively universal: a majority of teens and the overwhelming majority of college students now study with AI, and the market is racing to meet them. The outcome data is split, but not randomly. Tools that keep you thinking produce some of the largest learning gains ever measured. Tools that think for you produce measurable cognitive debt.

Your job is to stay on the right side of that line. Use AI to gather, explain, and quiz, then do the encoding yourself by reading closely, marking what matters, and rewriting it in your own words. Glasp is built for exactly that active layer: highlight the real source, summarize the real lecture, and keep your notes tied to what you actually read. The AI can help you find the material. Learning it is still, thankfully, your move.