# How to Learn With AI Without Fooling Yourself > AI explanations feel like learning even when they aren't. See what 2025–2026 trials found, plus seven copyable study prompts and a weekly loop to follow. - Canonical page: https://learnhubplay.nl/blog/learning-with-ai-without-fooling-yourself - Published: 2026-10-09 - By: LearnHubPlay, an independent software studio in Amsterdam - Topic: AI at Work - Tags: Learning, Retrieval practice, Spaced review, Prompting, Verifying AI output Say you ask an AI assistant how database indexes work. The answer is tidy, the analogy fits, and you nod through every paragraph. You close the tab sure that you understand them. On Friday a colleague asks why a query ignores one, and you can't produce a sentence without reopening the chat. The gap between feeling you understand and being able to do the thing is the central hazard of studying with AI. Researchers have documented it for decades with textbooks and lectures. Recent randomized trials of AI in learning, including two 2026 working papers, point in different directions: one found practice scores rising while exam scores fell, one found a purpose-built tutor beat in-class active learning on an immediate test, and one found that offering a course tutor went with lower grades. You'll leave with a rule for what to hand to the AI, seven copyable prompts, and a weekly loop that measures what you actually kept. ## Why a smooth explanation feels like understanding Elizabeth and Robert Bjork separate *performance*, what you can do now, from *learning*, a lasting change in what you know. They argue that [current performance can be "a highly unreliable index of whether learning has occurred"](https://bjorklab.psych.ucla.edu/wp-content/uploads/sites/13/2016/04/EBjork_RBjork_2011.pdf). Robert Bjork's term for the helpful opposite is *desirable difficulties*: spacing sessions, mixing topics instead of blocking them, varying conditions, and using tests rather than re-presentation as study events. Three findings give the idea its weight: - **Testing beats restudying after a delay.** In two experiments, [Roediger and Karpicke (2006)](https://profiles.wustl.edu/en/publications/test-enhanced-learning-taking-memory-tests-improves-long-term-ret/) had students read passages, then restudy them or take recall tests. After five minutes restudying won; after two days and a week the tested students remembered substantially more, though restudying had made them more confident. - **The feeling of learning can point the wrong way.** [Deslauriers and colleagues (2019)](https://doi.org/10.1073/pnas.1821936116) randomly assigned 149 Harvard physics students, in a crossover design, to active-learning sessions or lectures by highly rated instructors. The active group scored about 0.46 standard deviations (SD) higher on independent tests and reported about 0.56 SD lower feeling of learning. - **Producing beats reading.** [Slamecka and Graf (1978)](https://doi.org/10.1037/0278-7393.4.6.592) compared words people generated ("fast" from the cue "rapid–f") with the same words they only read ("rapid–fast"). Across five experiments, generated words were remembered better. Single words, so a principle, not a forecast. Our reading: an AI explanation resembles rereading with better production values, because it supplies the retrieval and generation you would otherwise do. That is an analogy, not a trial result: in the Contractor and Reyes trial below, explaining was the most common use of AI, and students who used it that way kept a gain at one week. The risk is not explanation; it is stopping there. The Bjorks add a caveat: a difficulty is desirable only if you have the background to respond to it, so for first contact AI legitimately helps. ## What the 2025–2026 trials of AI in learning found Four studies are worth knowing; they differ in age group, tool and outcome, so read the last column first. | Study | Who and how | What happened | Limits | |---|---|---|---| | [Bastani et al., PNAS 2025](https://pmc.ncbi.nlm.nih.gov/articles/PMC12232635/) | Nearly 1,000 students, grades 9–11, one Turkish school: textbook only; GPT-4 chat ("GPT Base"); or GPT-4 told to give hints, with teacher solutions ("GPT Tutor") | Practice scores rose 48% (Base) and 127% (Tutor) over control. On a closed-book exam Base scored 17% below control; Tutor did not differ significantly | One school, one subject, fall 2023 | | [Kestin et al., Scientific Reports 2025](https://pmc.ncbi.nlm.nih.gov/articles/PMC12179260/) | 194 eligible Harvard physics students, crossover: a purpose-built AI tutor versus an in-class active-learning lesson | Median gains over double; effect about 0.63 SD, an underestimate per the authors (ceiling effects) | Versus an active-learning lesson, not no help. Immediate post-test; relied on videos and expert-written prompts | | [Contractor and Reyes, working paper 2026](https://arxiv.org/abs/2607.08849) | 211 undergraduates, one college, proctored: studied an unfamiliar topic and wrote an essay with or without AI, then took unaided tests immediately and a week later | Scores 0.28 SD (7.2 percentage points) higher immediately, 0.21 SD (3.9 points) a week later. Gains held for students who used AI to explain, faded for those who had it write | One college, one week. Students chose their use, so that comparison is descriptive and imprecise. AI-allowed students broke the test's no-resources rule more often | | [Liu et al., working paper 2026](https://edworkingpapers.com/sites/default/files/ai26-1598.pdf) | 2,379 undergraduates, 30 instructors, a large U.S. public university; instructors randomly given a GPT-4o-based course tutor or not | Same-course subsample (1,353 students): final grades 0.37 SD lower (about 4 points of 100), recorded participation in online course activities 0.90 SD lower; full-sample grade drop about 0.27 SD, marginally significant. About 15% of students offered the tutor used it at all | Course grades, not independent tests. Most instructors kept the default direct-instruction mode (two used guided mode by term end). Both groups had free ChatGPT and Gemini | Three patterns run through them. **Help during practice and learning afterwards can diverge.** In the Turkish study, unrestricted GPT-4 lifted practice scores, then left students behind those who never had it; the authors call it a crutch. Surveys echo the fluency problem: unrestricted-arm students did not think they had learned less, and guarded-arm students believed they had done significantly better though their exam scores did not differ significantly from control's. Asked each of 57 practice problems ten times, GPT-4 was also right only about half the time; the authors note newer models have improved, so read 51% as a 2023 figure, not a current error rate. **Design seems to matter.** In the Turkish trial, the arm told to give hints, with teacher-written solutions and common mistakes in the prompt, avoided the plain chatbot's exam penalty. The Harvard tutor took students through each problem part in order, with step-by-step solutions in its prompts; a system prompt alone could not keep multi-part problems in sequence. The U.S. university tutor defaulted to a direct-instruction mode, few instructors switched to its guided mode, and the authors say their results do not speak to more scaffolded tutoring. Only the Turkish study randomized the design itself, so treat the shared ingredients (reference solutions in the prompt, structure that keeps the AI on one step, hints) as a hypothesis, not a proven recipe. **Use may matter as much as access.** In the Liu request logs, 73.8% of requests sought information, explanations or solutions, 11.3% involved practice or test preparation, and 0.7% asked for feedback on the student's own work, though the authors caution this doesn't measure offloading directly. [Fan and colleagues (British Journal of Educational Technology, 2025)](https://doi.org/10.1111/bjet.13544), with 117 students on a writing task, name the risk "metacognitive laziness": their ChatGPT group improved essays most, yet its knowledge gain and transfer did not differ significantly from the other groups'. Neither use-type comparison was randomized. > **Note:** This evidence is young: two of the four studies are working papers, outcomes were short-term (days to one semester), and the Turkish study used 2023's GPT-4. None tested the prompts below, which apply the mechanisms the research points to (retrieval, generation, hints before answers, feedback on your own attempt). Treat them as a reasoned default and check them against your own unaided score. ## One rule for deciding what to hand to the AI The AI can supply material, questions and feedback on your attempts. You supply the retrieval and the generation. | You ask the AI to… | Who retrieves or generates? | Verdict | |---|---|---| | Quiz you on your own notes | You | Do it: retrieval practice | | Critique an explanation you wrote | You, then the AI | Do it: generation plus feedback | | Give a hint after your attempt | You, with a nudge | Do it: the struggle survives | | Explain something new | The AI | Fine for first contact; then close the chat and recall it | | Summarize a chapter you haven't read | The AI | Skip; read first, then compare | | Solve the problem set, or rewrite your essay or code | The AI | Avoid: the "GPT Base" pattern; the output improves, you may not | Order matters too. An attempt first, even a weak three-minute one, gives the AI something to critique, the use almost nobody picked in the Liu logs. Before each prompt ask: will I have done something I couldn't do before, or only read something? ## Seven prompts that keep you doing the thinking Each prompt uses `{{double-brace}}` variables, with `{{name|default}}` for defaults, so you can keep it in a plain text file and fill the variables by hand or in your prompt tool. Our guide to [reusable prompt templates](https://learnhubplay.nl/blog/prompt-templates-with-variables) covers structuring and testing them. > **Tip:** Paste your own reference into every prompt where you can: lecture notes, the textbook page, an answer key. Check your school's or employer's policy before pasting copyrighted or confidential material. ### 1. Quiz me from my own material One question at a time, no hints, no answers until you reply: the testing effect, with your notes as ground truth. ```text You are my quiz master for {{topic}}. Treat only the text in as ground truth. {{my_notes_or_chapter}} Ask {{question_count|8}} questions, ONE at a time, mixing recall, "why" and "apply" questions. Do not hint, and do not reveal answers before I reply. After each reply, say correct, partly correct, or incorrect, give the right answer from the material in two sentences, then ask the next question. At the end, list the questions I missed. ``` ### 2. A Socratic tutor for when you're stuck Write your understanding first; the tutor questions it. The "I don't know" escape hatch matters: without background, a difficulty is just an obstacle. ```text Act as a Socratic tutor for {{topic}}. Here is my current understanding, in my own words: {{my_attempt}} Do not explain. Ask one short question at a time that exposes a gap or contradiction in what I wrote. If I answer "I don't know" twice in a row, give a three-sentence explanation of that one point and ask me to restate it. Stop when I have stated the idea correctly, and tell me what I got right. ``` ### 3. Explain it back, then critique me You explain in writing, then get feedback on your own work, the 0.7% category in the Liu logs. ```text I will explain {{concept}} to a {{audience|first-year student}}. {{source_material}} {{my_explanation}} Compare my explanation with the reference. List: (1) what I got right; (2) errors or misconceptions, quoting my words; (3) important points I left out; (4) the one question a sharp listener would ask next. Do not rewrite my explanation. ``` ### 4. Worked examples that fade Study a complete example, then take over a step at a time. A [2000 report by Renkl, Atkinson and Maier](https://www.psycharchives.org/handle/20.500.11780/196), published with Staley as [a 2002 journal article](https://doi.org/10.1080/00220970209599510), found that this fading beat conventional example-problem pairs in a field and a lab experiment, at least for near transfer. If you can already solve such problems, skip to prompt 3. ```text Teach me {{skill}} with faded worked examples. Example 1: fully worked, with the reason for each step. Example 2: same problem type, new numbers; leave the last step for me. Example 3: leave the last two steps for me. Continue until I solve a whole problem. Wait for my answer before showing anything. If a step of mine is wrong, say which step, but don't fix it. Vary the surface details between examples. ``` ### 5. Find the error Reviewing a flawed solution makes you check each step, a habit that may also help you catch AI mistakes. The evidence is thinner than for quizzing: an [Institute of Education Sciences project summary](https://ies.ed.gov/use-work/awards/improving-students-skill-solving-equations-through-better-encoding-algebraic-concepts) describes a growing body of research on incorrect examples, and one algebra study is [Booth and colleagues (2013)](https://doi.org/10.1016/j.learninstruc.2012.11.002). Promising, not proven. ```text Here is a problem and its correct solution: {{problem}} {{correct_solution}} Write a plausible-looking worked solution that contains exactly {{error_count|1}} mistake of a kind learners commonly make. Do not say where it is. I will name the faulty step and explain why it is wrong; then reveal the error and tell me whether I was right. ``` > **Warning:** Supply the correct solution yourself. If the model invents both the problem and the "error", it may plant the wrong mistake or call a correct step wrong, and you will memorize the confusion. ### 6. A question bank for spaced review End each session by turning notes into questions you can answer later without AI. The key comes last so you don't see it mid-attempt. Delete trivia and check the key against your source. ```text From the material below, write {{card_count|10}} questions I can answer from memory, ordered from basic to hard, each tagged with the concept it tests. Put the answer key in a separate list after all the questions. Skip trivia and yes/no questions. {{session_notes}} ``` ### 7. Hints, not answers For problem sets. It borrows the guarded tutor from the Turkish study: hints rather than answers, with a reference solution behind them. That tutor removed the exam penalty but showed no exam gain over control, so use it to protect effort, not as a shortcut. ```text I'm working on this problem: {{problem}} My attempt so far: {{my_attempt|none yet}} Reference solution (do not show unless I ask): {{reference_solution|none}} Give hints in three levels, one level per reply, and only when I type "next": (1) the concept involved; (2) the first step; (3) a similar solved example. Never skip a level and never solve the problem for me. Only if I type "show solution", give the full solution. ``` ### What it looks like: prompt 2 on the index question Back to the opening. Say your attempt was: "An index is a sorted copy of a column, so the database can jump to rows instead of scanning the table." An illustrative exchange (not a recorded transcript): ```text Tutor: A table has an index on email. Your query says WHERE lower(email) = 'ana@example.com'. What is the index sorted by? You: The original email values, I think. Tutor: And what does the query compare? You: lower(email). So the sort order doesn't match, and the database scans every row? Tutor: Test that idea. What would you build so the index matches the query? ``` PostgreSQL's [documentation](https://www.postgresql.org/docs/current/indexes-expressional.html) says a query on `lower(col1)` can use an index defined on that function's result; its [index-usage page](https://www.postgresql.org/docs/current/indexes-examine.html) names other reasons a planner skips an index, such as tables small enough that scanning is cheaper. If a prompt drifts (the Harvard team found a system prompt alone could not keep multi-part problems in sequence), restate its rule in the last line, such as "Ask only question 1, then wait", or open a fresh chat and re-paste your material and last attempt. ## A weekly study loop that measures what you keep Say you're learning SQL window functions in three 45-minute sessions a week plus a short Saturday check (a hypothetical schedule). Every session has the same shape: 1. **Retrieve without AI (10 minutes).** Answer last session's questions closed-book and mark yourself. Skip this in week one. 2. **Attempt cold (10 minutes).** Pick a source in advance: a chapter, tutorial or lecture. Read or watch it, close it, then work its exercises or write what you recall on a blank page. If you can't write one correct line after two minutes, you lack the base: borrow up to 10 minutes from step 3 for an AI explanation of the source's key terms, close the chat, and recall them. 3. **Assist (20 minutes).** Use prompts 2, 3, 4 or 7 to close the gaps you found. 4. **Log (5 minutes).** Run prompt 6 to save ten new questions, and note your score and one gap. Silence notifications for the session (our [30-minute notification audit](https://learnhubplay.nl/blog/notification-audit-mac-chrome-phone) shows how); switching away mid-attempt leaves [attention residue](https://learnhubplay.nl/blog/attention-residue-task-switching) behind, so an interrupted attempt tells you little. | Day | What you do | AI allowed? | |---|---|---| | Mon | New material A (no step 1 in week one) | Steps 3 and 4 | | Wed | Step 1 on Monday's questions, then material B | Steps 3 and 4 | | Fri | Step 1 on a mixed set (five questions each from A and B), then material C | Steps 3 and 4 | | Sat | **Unaided check (20 min):** ten questions across A to C, timed, no notes | Only after you've scored it | Mixed-order questions apply interleaving, another of the Bjorks' desirable difficulties; the spacing is a practical default, not a researched optimum. On Sunday, record the score and carry misses over, or fold it into a [20-minute weekly review](https://learnhubplay.nl/blog/weekly-review-template). ### The unaided score is the only number that counts Your **unaided score** is the share of Saturday's questions answered correctly with no AI and no notes. Score strictly: an answer counts only if you wrote it before looking, and a partly right answer is wrong. Track new questions and carried-over misses separately, because carried-over items test what you just re-studied and will rise faster. Say four Saturdays of new-question scores read 4, 5, 7 and 7 out of 10 (hypothetical). Ten questions is a coarse ruler, and a swing of one or two is noise, so judge the trend over three or four Saturdays. It is a rough but honest signal; how the chat felt is a poor guide, because in the chat everything feels fluent. A rule of thumb of ours, not from a study: if sessions feel easier but the new-question score is flat for two Saturdays, shift minutes from step 3 to steps 1 and 2: more retrieval and cold attempts, fewer explanations. ## When not to use AI for learning | Situation | Why hold back | Instead | |---|---|---| | You can't check the answer and a mistake is costly: dosages, legal deadlines | Models still make errors; in 2023 GPT-4 was correct only about half the time on the maths problems in the Turkish study | Paste your source, use only it, and verify anything you will memorize or act on | | You'll perform the skill without a tool: an exam, an interview | Assisted practice hides what you can't do alone | Run the Saturday check under real conditions | | You're reaching for the answer to escape discomfort | The discomfort may be the learning | Make two real attempts or work ten minutes (our rule of thumb, not a research finding), then ask for a level-1 hint (the concept only) | ## Where to start this week 1. Pick one topic and save prompts 1 and 7 as templates in a text file. 2. Run one session: retrieve, attempt, assist, log. 3. On Saturday, do the unaided check and write the score down. If the number climbs while your AI sessions get shorter, the tool is doing its job. ## Frequently asked questions ### Can I use this for coding or a language? Yes, in the same order: you generate, the AI gives feedback. Write the function or sentence first, run it or say it aloud, then ask for a critique, not a rewrite. For maths, prompts 4, 5 and 7 fit best; for reading-heavy subjects, use prompts 1, 3 and 6 on pasted text. ### What if I have no notes or textbook to paste? Use a syllabus, chapter outline or official documentation as the source. A quiz written from the model's memory has no ground truth, so check its answer key against a trusted source before you memorize it. ### My assistant has a study mode. Do I still need these prompts? Several assistants added one in 2025: [Gemini's Guided Learning (August)](https://blog.google/products-and-platforms/products/education/guided-learning/), [ChatGPT's Study mode (July)](https://campustechnology.com/articles/2025/07/30/new-chatgpt-study-mode-guides-students-through-questions.aspx) and [Claude's learning mode, opened to all users in August](https://engadget.com/ai/anthropic-brings-claudes-learning-mode-to-regular-users-and-devs-170018471.html). They push toward questions and hints, a good start. Menu names and availability change, so check your assistant's help page, turn the mode on and paste your source. None measures what you kept, so still run the Saturday check. ### Is it fine to let AI write my flashcards? Reasonable, with checks. What matters most is answering the cards from memory, which is the testing effect above. Slamecka and Graf compared generated words with read words; they did not test whether authoring the card yourself beats using one an AI wrote. ## Sources and further reading - [Generative AI without guardrails can harm learning: Evidence from high school mathematics](https://pmc.ncbi.nlm.nih.gov/articles/PMC12232635/) — Bastani et al., PNAS 122(26), 2025. The Turkish field experiment. - [AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting](https://pmc.ncbi.nlm.nih.gov/articles/PMC12179260/) — Kestin et al., Scientific Reports 15:17458, 2025. The Harvard crossover trial. - [The Effects of Course-Integrated AI Tutoring on Student Performance and Engagement: A Randomized University Trial](https://edworkingpapers.com/sites/default/files/ai26-1598.pdf) — Liu et al., EdWorkingPaper 26-1598, 2026. Working paper; grades and request types. - [Experimental Evidence on the Learning Impact of Generative AI](https://arxiv.org/abs/2607.08849) — Contractor and Reyes, arXiv 2607.08849, 2026. Working paper; explain versus write. - [Making Things Hard on Yourself, But in a Good Way: Creating Desirable Difficulties to Enhance Learning](https://bjorklab.psych.ucla.edu/wp-content/uploads/sites/13/2016/04/EBjork_RBjork_2011.pdf) — E. L. Bjork and R. A. Bjork, in *Psychology and the Real World*, 2011. Performance versus learning; desirable difficulties. - [Measuring actual learning versus feeling of learning in response to being actively engaged in the classroom](https://doi.org/10.1073/pnas.1821936116) — Deslauriers et al., PNAS 116(39), 2019. The 149-student crossover study. - [Test-enhanced learning: Taking memory tests improves long-term retention](https://profiles.wustl.edu/en/publications/test-enhanced-learning-taking-memory-tests-improves-long-term-ret/) — Roediger and Karpicke, Psychological Science 17(3), 2006. Testing versus restudying. - [The generation effect: Delineation of a phenomenon](https://doi.org/10.1037/0278-7393.4.6.592) — Slamecka and Graf, Journal of Experimental Psychology: Human Learning and Memory 4(6), 1978. Generated versus read words.