Technology and artificial intelligence have changed a lot about how students study, from the tools they use for homework to the way universities track progress. But the more interesting — and more contested — question isn't which apps students use. It's what those tools are doing to the actual mechanics of learning: how well information sticks, whether students can still reason through a problem without help, and whether their attention holds up under constant digital interruption. Researchers are now producing real evidence on this, and it's more mixed than either the hype or the panic suggests. For students planning to study at a university — especially internationally, where support systems and study habits at home may not translate directly — it's worth understanding what the research says before building a study routine around AI tools.
The Core Tension: Cognitive Offloading
Most of the debate about AI and learning comes down to one concept: cognitive offloading, the use of an external tool to do mental work you'd otherwise have to do yourself. Writing a phone number down instead of memorizing it, using a calculator instead of doing long division by hand — humans have offloaded cognition for as long as we've had external tools. AI just does it faster and for a much wider range of tasks, including the ones that used to be considered "the learning itself," like working through a problem or drafting an argument.
A 2025 review in Frontiers in Psychology frames this directly as a paradox: AI-based adaptive learning and intelligent tutoring systems can genuinely personalize instruction to a student's level, but the same convenience "reduce[s] the opportunity for active recall and problem-solving, which are essential components of cognitive development" (Jose et al., 2025, Frontiers in Psychology, via PMC). Whether AI helps or hurts a given student's learning seems to depend heavily on how it's used, not just whether it's used — a theme that shows up across nearly every study on the subject.
What the Evidence Says About Memory and Retention
The research on memory specifically points toward a real risk of over-reliance. The same Frontiers in Psychology review cites a study of information science undergraduates who were split into two groups: one did a "pretest" — attempting the material before turning to AI — while a control group went straight to AI for answers. Pretesting improved retention and engagement, but the group with prolonged, undirected AI exposure showed measurable memory decline over time (Jose et al., 2025).
A more striking version of this came out of a 2025 MIT Media Lab study (cited in the same review) that used EEG to measure brain activity while participants wrote essays with an LLM, with a search engine, or with no tool at all. Cognitive activity was lowest in the LLM group, and self-reported "ownership" of the essay — how much participants felt the ideas were actually theirs — was lowest for LLM users and highest for the no-tool group. Across a four-month follow-up, the LLM group consistently underperformed the other two on neural, linguistic, and behavioral measures (Kosmyna et al., 2025, cited via Impact of AI Tools on Learning Outcomes, arXiv).
A randomized classroom experiment tells a similar story from a different angle. At Corvinus University of Budapest, researchers split an operations research class into an AI-permitted group and an AI-restricted group for both coursework and exams. The point of the study was to isolate what unrestricted AI access does to learning outcomes when it's genuinely optional rather than banned outright. The authors' conclusion: "uncontrolled use of AI tools leads to disengaged students and low understanding of material" (Benedek & Sziklai, 2025, Corvinus University of Budapest, arXiv). Worth noting: the experiment was disrupted by student pushback partway through (some escalated concerns to Hungary's State Secretary for Higher Education), which the authors themselves treat as a data point — a sign of how quickly AI use has become something students feel entitled to, independent of whether it helps them learn.
None of this means AI universally damages memory. It means unstructured, answer-seeking use — asking AI for the finished product and moving on — is the pattern associated with weaker retention. Studies that build in a "try first" step before AI assistance tend to find smaller downsides or even benefits.
Critical Thinking, Problem-Solving, and Passive Acceptance
A second strand of research looks at reasoning skills rather than memory. A 2024 survey by Gerlich, cited in the Frontiers in Psychology review, found a statistically significant negative correlation between frequent AI tool use and critical thinking ability, and identified cognitive offloading as the mechanism connecting the two (Jose et al., 2025). A separate randomized study measuring cognitive engagement during academic writing found significantly lower engagement scores in an AI-assisted group compared to a non-assisted control — summarized bluntly by the study's author as evidence that "ChatGPT produces more lazy thinkers" (Georgiou, 2025, cited via Jose et al.).
A January 2026 Brookings Institution report — based on focus groups and interviews with students, parents, and educators across 50 countries, plus a review of hundreds of studies — goes further, describing a "doom loop" pattern of AI dependence in which students increasingly offload thinking onto the technology. Brookings senior fellow Rebecca Winthrop put it this way: "When kids use generative AI that tells them what the answer is … they are not thinking for themselves. They're not learning to parse truth from fiction. They're not learning to understand what makes a good argument" (Brookings Institution, via NPR, January 2026). The report is explicit that this isn't a new phenomenon — keyboards reduced handwriting practice, calculators automated arithmetic — but argues AI has "turbocharged" the offloading effect because it can now do the reasoning, not just the calculation.
It's worth flagging that this literature is young and contested. A separate meta-analysis of 51 experimental studies on ChatGPT found a large positive effect on learning performance overall, and one 2025 experiment found students learned significantly more, in less time, using an AI tutor compared with in-class active learning (cited in Benedek & Sziklai, 2025, arXiv). The arXiv paper's own authors note the field is "highly contradictory" and offer a plausible reconciliation: outcomes look positive when instructors actively structure how students use AI, and outcomes look negative when use is unstructured and answer-seeking. The tool isn't the variable that matters most — the usage pattern is.
How Students Are Actually Using AI Right Now
A 2025 USC Center for Generative AI and Society survey of 1,000 U.S. college students sheds light on what "unstructured use" looks like in practice. Researchers distinguished between "executive help" — using AI to get a fast answer with minimal effort — and "instrumental help" — using it to clarify a concept or build a skill. Most students defaulted to executive use. The one factor that reliably shifted students toward instrumental, learning-oriented use was active encouragement from instructors to use AI thoughtfully rather than as a shortcut (USC Center for Generative AI and Society, via USC Today, September 2025).
The same research team piloted a writing tool called ABE (AI for Brainstorming and Editing), designed to prompt reflection and revision rather than generate finished drafts. Students who used it reported treating it as a companion for strengthening arguments and exploring counterarguments, not a replacement for drafting — evidence that tool design, not just willpower, shapes whether AI use ends up instrumental or executive.
Where AI Genuinely Adds to Learning
It isn't all downside. Harvard Graduate School of Education researcher Ying Xu argues the framing of "what AI takes away" often misses what it can add, particularly for tasks that were never going to get deep human attention anyway. Her research with PBS Kids embedded conversational AI into children's TV programming so kids could ask science questions of on-screen characters mid-episode — turning a one-way, passive medium into something interactive. The result, she reports, was measurable improvement in children's scientific reasoning and engagement compared with passive viewing (Ying Xu, Harvard Graduate School of Education, April 2025). Her framing is useful for older students too: the relevant comparison usually isn't "AI versus deep, focused human study" — it's "AI versus whatever a student would otherwise be doing with that time," which for a lot of everyday studying is passive rereading or skimming, not rigorous self-testing.
The U.S. Department of Education's 2023 report on AI in education makes a related point about accessibility: AI's ability to personalize pacing and adapt content difficulty has real value for students with disabilities, multilingual learners, and students catching up on unfamiliar material — the same populations for whom cognitive offloading concerns are sometimes overstated, because the alternative isn't unaided deep work, it's being unable to access the material at all (U.S. Department of Education, Office of Educational Technology, 2023). The Brookings report echoes this from the equity side: one program uses AI to deliver curriculum in Dari, Pashto, and English to Afghan girls barred from formal schooling — a case where AI access is not competing with deep, unaided study, because no alternative currently exists (Brookings, via NPR).
A Practical Framework: Instrumental, Not Executive
Pulling the research together, a rough rule emerges for using AI without eroding the learning it's meant to support:
- Attempt first, then check. The pretesting study above found this single sequencing change — try before you look — improved retention even when AI was available afterward. Treat AI as a way to check and refine your own attempt, not a first resort.
- Ask AI to explain, not to answer. Instrumental use (clarifying a concept, working through why an answer is correct) tracked with better outcomes than executive use (getting the finished answer) in the USC survey.
- Use retrieval practice AI can't replace. Self-testing without notes is more effortful than rereading or having AI summarize — and that effort is exactly what research on desirable difficulty says drives retention (cited via Benedek & Sziklai, arXiv).
- Notice if you can't explain it without the tool. If you can't reconstruct an explanation in your own words a day later, that's a signal the material didn't transfer to memory — regardless of how confident the AI-assisted version felt in the moment.
- Treat institutional guidance as useful, not just restrictive. The USC research found students who got direct instructor guidance on how to use AI shifted toward learning-oriented use far more than students left to figure it out alone — so if a program or professor offers AI-use guidelines, they're worth taking seriously rather than working around.
What This Means If You're Choosing Where to Study
AI policy and support vary a lot by institution — some universities have built explicit AI-literacy guidance into coursework, others are still working it out ad hoc, and a few restrict it heavily in exams. None of this is usually advertised prominently, so it's worth asking directly during admissions conversations or orientation how a program expects students to use AI tools, and whether there's structured guidance versus a blanket ban. If you're comparing programs abroad, factor this in alongside the more familiar variables like curriculum and cost — a program that treats AI literacy as a taught skill is arguably preparing students for how research and knowledge work actually happen now, compared with one that either ignores the issue or treats all AI use as misconduct.
FAQ
Does using AI to study make you worse at learning? The evidence is genuinely mixed and depends heavily on how it's used. Unstructured, answer-seeking use is associated with weaker memory retention and lower critical-thinking scores in several 2025 studies. Structured use — attempting material first, then using AI to check or clarify — shows smaller downsides or even benefits.
Is cognitive offloading always bad? No. Offloading routine tasks (like using a calculator for arithmetic) frees up mental effort for higher-level thinking, which is generally a good trade. The research concern is specifically about offloading the reasoning and problem-solving steps that are the actual point of an assignment — not offloading everything, all the time.
What's the difference between "executive" and "instrumental" AI use? Executive use means getting a fast answer with minimal effort — essentially skipping the learning step. Instrumental use means using AI to clarify a concept, get feedback, or build a skill you still have to apply yourself. USC's 2025 survey found most students default to executive use unless an instructor actively steers them toward instrumental use.
Should international students worry about this more than others? Not inherently more, but the stakes can feel higher: international students are often adapting to a new academic culture, a second language, and unfamiliar norms around plagiarism and AI-assisted work simultaneously. It's worth clarifying a program's specific AI policy early rather than assuming norms from home will transfer directly.
Are universities actually restricting AI use? Practices vary widely and change quickly. Some programs have adopted explicit AI-literacy components; others restrict AI in assessed work; enforcement mechanisms (like AI-detection software) are themselves imperfect and contested. Confirm current policy directly with any program you're applying to rather than assuming a default.
The Bottom Line
None of these tools replace good instruction, a supportive institution, or your own study discipline — and the emerging research suggests that used carelessly, they can actively work against retention and critical thinking. Used deliberately — attempt first, use AI to clarify rather than replace effort, and treat instructor guidance as signal rather than noise — the same tools can extend support outside class hours without the downsides the research flags. If you're weighing where to study next and want to see how programs and universities stack up, browse universities or explore funding options to help plan your next step.
Sources
- U.S. Department of Education, Office of Educational Technology — Artificial Intelligence and the Future of Teaching and Learning (2023)
- Jose, Cherian, Verghis, Varghise, Mumthas & Joseph — The cognitive paradox of AI in education: between enhancement and erosion, Frontiers in Psychology (2025)
- Benedek & Sziklai, Corvinus University of Budapest — Impact of AI Tools on Learning Outcomes: Decreasing Knowledge and Over-Reliance, arXiv (2025)
- USC Center for Generative AI and Society, via USC Today — AI is changing how students learn — or avoid learning (2025)
- Brookings Institution Center for Universal Education, via NPR — The risks of AI in schools outweigh the benefits, report says (2026)
- Ying Xu, Harvard Graduate School of Education — AI Can Add, Not Just Subtract, From Learning (2025)