Playbooks
Using AI for exam prep without hurting your learning
MIT Media Lab's EEG study found that students who wrote essays with ChatGPT showed the weakest brain connectivity of three tested groups and struggled to recall or quote their own submitted work — evidence for what researchers call "cognitive debt." But the same body of research, including Anthropic's analysis of over half a million student conversations, shows AI can support studying well when it's used for practice questions, explanation, and structured review rather than as a replacement for the recall and reasoning the exam is actually testing.
Last updated Aug 21 · 17 min read
What "cognitive debt" actually is, and what MIT's study found
MIT Media Lab's study "Your Brain on ChatGPT" tracked participants over four months, using EEG to measure brain connectivity as they completed essay-writing tasks under three conditions: writing entirely unassisted ("Brain-only"), using a search engine, or using ChatGPT. The ChatGPT group showed the weakest neural connectivity of the three groups and, in the study's own language, "consistently underperformed at neural, linguistic, and behavioral levels" across the study period (MIT Media Lab; arXiv preprint).
The more striking finding is what happened when the same ChatGPT-reliant participants were later asked to write without the tool. Their brain connectivity in that unassisted condition was still reduced, consistent with an accumulated effect the researchers term "cognitive debt" — a pattern where repeated reliance on an external tool for a cognitive task leaves the brain less engaged even once the tool is taken away, similar to how a muscle that stops being used doesn't instantly regain strength the moment you try to use it again.
There's a second, less-quoted finding that matters just as much for exam prep specifically: ChatGPT users reported the lowest sense of ownership over their own essays, and — this is the part that should worry anyone using AI to prepare for a test — they struggled to accurately quote sentences from the work they had just submitted. If a student can't recall or accurately reproduce material they supposedly just studied, that's close to a direct measurement of what an exam is designed to catch.
How students are actually using AI for coursework and exam prep
Anthropic's Education Report analyzed roughly 575,000 conversations with Claude that were classified as academically relevant, giving a real picture of usage patterns rather than a lab simulation. The two largest use-case categories were content creation (39.3% — things like generating practice questions, editing drafts, and summarizing material) and technical/STEM problem-solving (33.5%) (Anthropic).
The report also broke down conversations by depth of engagement, and the split is worth sitting with: about 47% of student-AI conversations were classified as "Direct" — essentially seeking an answer or finished content with minimal back-and-forth — versus more collaborative, multi-turn conversations involving iterative problem-solving or refinement. That's nearly half of all student AI use falling into the pattern most likely to resemble outsourcing rather than learning, even in a dataset that also contains a lot of genuinely constructive use.
A Bloom's Taxonomy breakdown in the same report adds a specific concern: Claude was disproportionately used for higher-order cognitive tasks — Creating (39.8%) and Analyzing (30.2%) — the exact categories of thinking that exam prep is supposed to build in the student, not delegate to a tool. The report explicitly flags certain patterns as academically risky rather than simply neutral: using AI to directly answer multiple-choice questions, or rewriting text specifically to evade plagiarism detection, are both named as concerning use cases rather than acceptable study techniques.
The line between helpful AI use and academically risky AI use
Put the MIT and Anthropic findings together and a pattern emerges that doesn't require picking a side in the "AI ruins learning" versus "AI is just a tool" debate: the risk isn't AI use itself, it's which cognitive step gets delegated. Using AI to generate five extra practice questions on a topic you already understand, then answering them yourself, keeps the recall and reasoning step — the part an exam actually measures — with the student. Asking AI to write the answer and then reading it is a different action entirely, even though both look like "using AI to study" from the outside.
This distinction shows up directly in a real, community-built tool: AnkiAIUtils, a GitHub project by a medical student, uses ChatGPT specifically to generate flashcard content and explanations while leaving Anki's actual spaced-repetition review mechanism — the part that forces active recall on a schedule — completely untouched (GitHub). The project has drawn attention on Hacker News (253 points, 43 comments), which suggests other students recognize the same principle: let AI help produce study material, but don't let it replace the act of retrieving that material from memory yourself.
Community discussion on this split is genuinely divided rather than settled. In an "Ask HN" thread specifically about whether AI chatbot wrappers are ruining EdTech, one self-identified teacher reported students "learning more and faster than ever," while a formal-methods instructor in the same thread said students using LLMs alone "cannot do formal reasoning" — two credible, opposing first-hand accounts in the same discussion (Hacker News). One commenter in the same thread offered a more cynical institutional read: that administrations have an incentive to keep enrollment revenue flowing "regardless of if the students are learning" — a pointed but anecdotal claim worth flagging as opinion, not verified fact.
Where the science is still unsettled
It's worth being explicit about what isn't nailed down yet. A widely cited meta-analysis on ChatGPT's effect on student learning performance — reporting a large positive effect across 51 studies — was published in the Nature-family journal Humanities & Social Sciences Communications, then itself retracted in 2026 after reviewers flagged discrepancies that undermined confidence in the analysis and its conclusions. A study racking up hundreds of thousands of views and hundreds of citations before being pulled is a concrete reason to treat any single study in this space, including ones that sound authoritative, as provisional rather than final (Nature). Evidence not sufficiently verified: there is no single, replicated, consensus number for "how much AI use reduces exam performance" that this research turned up — the MIT study is small-sample and EEG-based rather than a large exam-score comparison, and no equivalently rigorous large-scale exam-score study surfaced in this research to pair with it.
One opinion piece surfaced in this research argues the practical fix for AI-enabled shortcuts isn't better detection tools at all, but a structural change: more in-class, closed-book testing, on the logic that trying to catch AI use after the fact is a losing game compared to designing assessment that doesn't depend on catching it. That's a policy argument, not a settled research finding, but it's a useful frame for why some institutions are changing how they test rather than how they police.
Practical examples
– Real: AnkiAIUtils. A medical student's open-source tool that has ChatGPT generate flashcard content and mnemonic explanations, while Anki's spaced-repetition scheduling algorithm (FSRS) still controls when and how often the student is quizzed on each card — AI assists content creation, the student still does the recall.
– Real: Anthropic's "Direct" conversation category. Nearly half of academically-relevant Claude conversations analyzed fell into this shallow-engagement pattern, which is a real, measured behavior, not a hypothetical worry.
– Illustrative, not a documented case: picture a student prepping for a certification exam who asks an AI model to explain a concept they got wrong on a practice test, works through two more variations of the same problem type by hand, then re-tests themselves a week later without the tool — a workflow consistent with what the "helpful use" pattern above describes, though we're not citing this as a verified real account.
Data and evidence
– ChatGPT-assisted writers showed the weakest brain connectivity of three groups across a four-month EEG study (MIT Media Lab; arXiv preprint).
– Reduced connectivity persisted even after the tool was removed, consistent with "cognitive debt" (MIT Media Lab).
– ~575,000 academically-relevant Claude conversations analyzed; content creation (39.3%) and STEM problem-solving (33.5%) were the top two use cases (Anthropic Education Report).
– ~47% of student-AI conversations were "Direct" (answer-seeking, low engagement) rather than collaborative or iterative (Anthropic).
– Creating (39.8%) and Analyzing (30.2%) were the Bloom's Taxonomy categories most represented in student AI use, per the same report — both higher-order thinking skills exam prep is meant to build directly in the student.
– Evidence not sufficiently verified: a precise, large-scale, replicated percentage for how much AI reliance changes actual exam scores (as opposed to lab-measured brain connectivity or self-reported usage patterns).
Comparisons: AI chatbots vs. spaced-repetition tools
Approach: General chatbot (ChatGPT, Claude, etc.) used for direct answers · What it's good at: Fast explanations, unblocking a stuck concept · Risk if misused: High — MIT's data ties this pattern to weaker connectivity and lower content ownership
Approach: General chatbot used to generate practice questions/explanations only · What it's good at: Expands practice material quickly, keeps recall with the student · Risk if misused: Low, if the student still self-tests without the tool present
Approach: Anki / spaced repetition (FSRS algorithm) · What it's good at: Long-term retention through scheduled active recall · Risk if misused: Low on its own — the mechanism is designed around forced recall
Approach: AI-augmented Anki (e.g. AnkiAIUtils) · What it's good at: Combines AI-generated content with unchanged spaced-repetition mechanics · Risk if misused: Low, specifically because the recall step isn't touched by the AI layer
Approach: AI used to answer multiple-choice questions directly · What it's good at: None for learning; may complete an assignment faster · Risk if misused: High — explicitly flagged as academically risky in Anthropic's own report
The pattern across every row is the same: risk tracks with whether AI is doing the recalling/reasoning, or whether AI is doing the material-generation while the student still recalls and reasons.
Real-world use cases
The clearest documented real-world use case is Anthropic's own data on how students already use Claude for coursework: heavy use for STEM problem-solving and content creation, split roughly evenly between shallow "give me the answer" interactions and more genuinely collaborative multi-turn sessions. That's not a hypothetical trend — it's the university-student conversation pattern as it exists in production usage data today.
The Hacker News discussion thread is a second real-world use case worth citing directly, precisely because it's not unanimous: it documents at least one educator observing accelerated learning and at least one educator observing degraded formal reasoning ability in the same student population, which is closer to the honest state of the debate than either extreme framing ("AI ruins education" or "AI is purely additive") would suggest.
Common mistakes
– Asking AI for the final answer instead of an explanation, then treating that answer as studied material. This is the exact "Direct" pattern Anthropic flags as the most common, lowest-engagement usage type.
– Using AI to write practice-test answers and reviewing the AI's reasoning instead of writing your own reasoning first. The MIT study's ownership finding — students couldn't accurately quote their own AI-assisted essays — suggests this weakens retention specifically.
– Letting AI rewrite your notes into a polished summary and treating reading that summary as equivalent to studying. Passive reading of AI-condensed material skips the retrieval step that actual learning research (independent of this specific research set) consistently ties to retention.
– Using AI to answer multiple-choice practice questions directly. Anthropic's report specifically names this as an academically risky pattern, not a study technique.
– Assuming more AI-generated practice material automatically means better prep. Volume of material doesn't substitute for spaced, active retrieval of that material — the mechanism, not the content source, is what drives retention.
Best practices for using AI in exam prep
1. Use AI to generate practice questions and explanations, not final answers you passively read. This keeps you in the active-recall role the exam will require.
2. Pair AI-generated content with a spaced-repetition system rather than a one-time read-through. The AnkiAIUtils model — AI for content, unchanged spaced repetition for delivery — is a concrete, real template for this.
3. Self-test without the tool present on a delay, not immediately after using it. The MIT study's most concerning finding is reduced recall precisely when the tool was removed — testing yourself under those same conditions before the actual exam surfaces that gap while you can still fix it.
4. Ask AI to explain why an answer is correct, then re-derive the answer yourself from that explanation, rather than copying the explanation as your final study note.
5. Avoid using AI to directly answer practice multiple-choice or exam-style questions — generate distractors or variations instead, and answer them yourself.
6. Treat "I understood the AI's explanation" and "I can reproduce this without the AI" as two different checkpoints, and don't consider a topic studied until you've cleared the second one.
Key takeaways
– MIT Media Lab's EEG study found ChatGPT-assisted essay writers had the weakest brain connectivity of three tested groups, with reduced connectivity persisting even after the tool was removed — the basis of "cognitive debt."
– Anthropic's analysis of ~575,000 real student conversations found nearly half were low-engagement "Direct" answer-seeking interactions, alongside genuinely productive uses like practice-question generation.
– The dividing line isn't whether you use AI, but which cognitive step you hand to it: generating practice material is lower-risk than having AI do the recalling or reasoning for you.
– Real tools like AnkiAIUtils show a practical, already-adopted pattern: AI for content generation, unchanged spaced repetition for the actual memory work.
– The research in this space is genuinely unsettled in places (note the retraction tied to a widely cited study), so treat strong claims — including some of the ones in this article — as evidence-based but provisional.
Relevant tools.scult.in resources
The exam prep prompt library, student-focused prompt library, and research prompt library are built around the practice-generation-and-explanation pattern this article shows is the lower-risk way to use AI in exam prep — worth a look before you build your own study prompts from scratch.
Frequently asked questions
Does using ChatGPT to study actually hurt learning?
Evidence from MIT Media Lab's EEG study shows reduced brain connectivity and weaker content ownership among ChatGPT-assisted writers compared to unassisted writers, but this is one study, not a universal verdict on every use case.
What is "cognitive debt" in the context of AI and studying?
A term from MIT's research describing reduced brain engagement that persists even after the AI tool is removed, following a period of reliance on it — like an accumulated cost from outsourcing a cognitive task.
Is it cheating to use ChatGPT to study for an exam?
It depends on your institution's specific policy and what you use it for; generating practice material for yourself is generally treated differently from having AI produce graded work directly.
Can AI generate good practice questions for exams?
Yes — this is one of the two largest documented use cases in Anthropic's analysis of real student conversations (content creation, 39.3%).
Does AI hurt memory and critical thinking?
MIT's study specifically found reduced connectivity and weaker recall of one's own AI-assisted work; it didn't test memory/critical thinking as isolated constructs beyond that essay-writing task.
What is active recall?
A study technique where you retrieve information from memory without looking at the source, rather than passively re-reading it — the mechanism spaced-repetition tools like Anki are built around.
What is spaced repetition?
A study method that schedules review of material at increasing intervals timed to when you're about to forget it, maximizing long-term retention per unit of study time.
What is FSRS?
Free Spaced Repetition Scheduler, the modern scheduling algorithm used by Anki to decide when each flashcard should reappear for review, based on your past recall performance.
Is Claude better than ChatGPT for exam prep?
The research reviewed here doesn't provide a head-to-head performance comparison for exam prep specifically; Anthropic's own data describes usage patterns for Claude, not a comparative benchmark against ChatGPT.
Can AI replace a tutor for exam prep?
The evidence suggests AI can supplement explanation and practice generation, but the collaborative, multi-turn engagement associated with better outcomes in Anthropic's data still requires the student to stay actively involved, similar to what a good tutor would expect.
What percentage of student-AI conversations are just "give me the answer" requests?
About 47%, classified as "Direct" conversations in Anthropic's analysis of academically-relevant Claude usage.
What are the top ways university students actually use AI for coursework?
Content creation (39.3%) and technical/STEM problem-solving (33.5%) were the largest categories in Anthropic's dataset.
Does AI use show up in brain activity differently than searching Google?
Yes — MIT's three-way comparison found the ChatGPT group showed the weakest connectivity, distinct from both the search-engine and unassisted groups.
Do students who use AI to write feel the work is still theirs?
MIT's study found ChatGPT users reported the lowest sense of ownership over their essays among the three groups tested.
Is there a retraction related to a major ChatGPT-and-learning study?
Yes — a widely cited meta-analysis claiming ChatGPT had a large positive effect on student learning performance, published in Humanities & Social Sciences Communications (a Nature-family journal), was itself retracted in 2026 after reviewers found discrepancies undermining its conclusions — a reason for caution around any single strong claim in this space.
What specific AI uses do researchers flag as academically risky?
Anthropic's report specifically names directly answering multiple-choice questions and rewriting text to evade plagiarism detection as concerning patterns.
Are higher-order thinking tasks (analysis, creation) being outsourced to AI by students?
Anthropic's Bloom's Taxonomy breakdown found Creating (39.8%) and Analyzing (30.2%) were disproportionately represented in student AI use, which the report frames as a potential offloading concern.
Is AI good at generating flashcards for spaced repetition?
Real-world tooling like AnkiAIUtils uses AI specifically for this purpose, generating flashcard content and explanations while leaving the spaced-repetition scheduling itself unchanged.
Do teachers agree on whether AI is helping or hurting student learning?
No — a Hacker News discussion among educators showed genuinely split first-hand experience, with one teacher reporting faster learning and another reporting degraded formal-reasoning ability in the same discussion thread.
Is AI-assisted studying different for STEM subjects versus humanities?
Anthropic's data shows STEM problem-solving as a top use case (33.5%), suggesting heavy real-world use there, but the research reviewed doesn't provide a direct comparison of learning outcomes by subject area.
How do I use ChatGPT to make flashcards without hurting my retention?
Have it generate the questions and explanations, then manually (or via a spaced-repetition tool) quiz yourself on them later rather than reading the AI's answer once and moving on.
How do I use AI for active recall studying?
Ask AI to generate questions on material you've already learned, cover the answer, attempt it yourself, and only check the AI's explanation after you've committed to your own answer.
How do I study with AI without it counting as cheating?
Check your institution's specific academic integrity policy on AI use, and default to using it for practice/explanation rather than for producing content you submit as your own original work.
How do I know if I'm using AI in a way that builds cognitive debt?
A rough signal from the MIT findings: if you can't accurately reproduce or explain material shortly after using AI to "study" it, you may be accumulating the same pattern MIT observed rather than genuinely learning it.
How do I combine ChatGPT with Anki effectively?
Use ChatGPT to draft card content and mnemonics (as the AnkiAIUtils project does), but let Anki's own spaced-repetition algorithm control review timing — don't let AI substitute for the review sessions themselves.
How do I test myself on material after using AI to study it?
Wait at least a day, then attempt to explain or solve the material from memory without the AI open, mirroring the "unassisted" condition MIT's study used to detect cognitive debt.
How do I ask AI for help on a topic without it just giving me the final answer?
Explicitly prompt it to explain the underlying concept or give you a similar-but-different practice problem rather than solving your exact question.
How do I use AI for exam prep in a way that Anthropic's own report wouldn't flag as risky?
Avoid using it to directly answer multiple-choice questions or to reword content specifically to bypass a plagiarism checker — both are explicitly named as concerning patterns in their research.
How do I know if my study habits with AI are working?
Track your unassisted recall performance over time (via self-testing or practice exams without AI open), not just how confident the AI's explanations made you feel in the moment.
How do I structure a study session that uses AI for the parts it's good at?
Use AI upfront to clarify a confusing concept and generate practice material, then close the tool for the actual practicing, recalling, and self-testing portion of the session.
Is there a measurable, agreed-upon threshold for how much AI use is "too much" for exam prep?
No — the research reviewed doesn't establish a specific usage threshold; the more supported signal is the type of cognitive step being delegated, not a quantity of use.
Does the MIT cognitive debt study generalize beyond essay writing to other academic tasks like math or coding?
Not established by the study itself, which focused specifically on essay writing; extending its findings to other task types (e.g. problem sets) is a reasonable hypothesis but not something this specific research verified.
Is the 47% "Direct" conversation pattern in Anthropic's data getting better or worse over time?
The available report is a snapshot analysis; it doesn't provide a longitudinal trend showing whether this ratio is improving or worsening over successive semesters.
Could AI-assisted studying help some students (e.g. those with learning differences) more than it hurts them?
This research didn't turn up a specific study addressing that question directly; it's a plausible hypothesis but "evidence not sufficiently verified" here.
Is "in-class closed-book testing" actually gaining traction as an institutional response to AI use?
At least one opinion piece argues for it as the practical fix, but this research didn't find data on how widely institutions have actually adopted that shift.
ChatGPT vs. Anki for studying — which is better?
They serve different functions: ChatGPT (or Claude) is strong for generating explanations and practice material on demand; Anki is purpose-built for scheduled active recall over time. The AnkiAIUtils project treats them as complementary rather than competing.
AI tutoring vs. traditional studying — what does the evidence actually support?
The evidence supports AI as a supplement for explanation and practice generation; it doesn't support AI tutoring as a full replacement for the recall and self-testing that traditional (or AI-augmented) active study still requires.
Claude vs. ChatGPT for exam prep — which has better data behind it?
Anthropic has published detailed real-usage data for Claude specifically (the Education Report); this research didn't find an equivalent, directly comparable usage-pattern report from OpenAI for ChatGPT.
AI flashcards vs. manually written flashcards — does it matter who writes them?
The AnkiAIUtils use case suggests AI-generated flashcard content paired with unchanged spaced-repetition review works as a real, adopted workflow; the research reviewed doesn't show a controlled comparison of AI-written versus self-written card content specifically.
Is AI-assisted studying better or worse than search-engine-assisted studying?
MIT's three-way comparison found the ChatGPT group had weaker brain connectivity than the search-engine group, suggesting the two aren't equivalent, though the study didn't test broader academic performance outcomes beyond the essay task itself.
I've been using AI heavily to study and now I can't remember material without it — what's going on?
This matches the "cognitive debt" pattern MIT documented — reduced recall when the tool is removed after a period of reliance; the fix is deliberately practicing recall without the tool, starting now rather than right before the exam.
My grades dropped after I started using ChatGPT to help me study — is that the AI's fault?
It's not possible to attribute a specific grade drop to AI use without more context, but the documented risk pattern (using AI for direct answers rather than practice generation) is worth auditing in your own study habits.
I can't tell if I actually understand a topic or if I just understood the AI's explanation of it — how do I check?
Close the AI tool and try to re-explain or solve the topic from scratch; if you can't, you understood the explanation, not the material itself.
AI keeps giving me the answer even when I ask it to just explain — how do I stop that?
Be explicit in your prompt that you want a concept explanation or a similar practice problem, not the solution to the exact problem you're stuck on — model behavior on this varies, so you may need to restate the instruction if it slips back into solving.
I used AI to write my study notes and now they feel like someone else's notes — is that a problem?
It matches the "reduced ownership" finding from MIT's study; rewriting the AI's summary in your own words, by hand, from memory, is a more reliable way to build the retention that reading a polished summary doesn't.
What's the best AI study app for exam prep?
The research reviewed here doesn't endorse a single specific commercial app; the stronger pattern is using a general AI model for content generation/explanation paired with a dedicated spaced-repetition tool like Anki for the recall mechanism.
Is there a good AI flashcard generator that won't hurt my retention?
Open-source tools like AnkiAIUtils are built around the principle of AI-generated content plus unchanged spaced-repetition mechanics, which the evidence in this article supports as a lower-risk pattern than direct-answer AI use.
Should I pay for a premium AI study tool or just use a free chatbot plus Anki?
The research doesn't provide a cost-benefit comparison between paid study apps and a free-chatbot-plus-Anki combination; the underlying principle (AI for content, spaced repetition for recall) applies either way.
Are there AI exam-prep tools specific to competitive exams like UPSC, JEE, or REET in India?
A distinct AI-prep and coaching-adjacent tool ecosystem exists around India's major competitive and government exams, though this research did not independently verify specific tool names, claims, or outcomes in that ecosystem — treat vendor claims there with the same scrutiny applied above.
What's the single most important habit for using AI in exam prep without hurting learning?
Keep the recall and reasoning step — actually retrieving and re-deriving the answer yourself, without the tool open — as something you do every time, regardless of how AI helped you generate or understand the material beforehand.
Sources
- https://www.media.mit.edu/publications/your-brain-on-chatgpt/
- https://arxiv.org/abs/2506.08872
- https://www.anthropic.com/news/anthropic-education-report-how-university-students-use-claude
- https://www.nature.com/articles/s41599-025-04787-y
- https://github.com/thiswillbeyourgithub/AnkiAIUtils
- https://hn.algolia.com/api/v1/search?query=ChatGPT%20students%20learning
