{"course":{"slug":"ai-for-students","title":"AI for Students","url":"https://addaly.com/learn/ai-for-students","level":"beginner","track":"work"},"lesson":{"slug":"what-happens-to-your-memory","title":"What Happens To Your Memory","url":"https://addaly.com/learn/ai-for-students/what-happens-to-your-memory","position":2,"of":73,"minutes":8,"answer":"The tutor that gave answers raised practice scores and lowered exam scores; the one that gave hints did neither harm.","keyIdea":"The tutor that gave answers raised practice scores and lowered exam scores; the one that gave hints did neither harm."},"body":{"format":"markdown","text":"## The study you should know about\n\nIn 2024, researchers ran something close to the experiment every student privately wonders about. Bastani and colleagues worked with roughly a thousand high school students in Turkey, across grades 9 to 11, in maths.\n\nStudents were split three ways for their practice sessions:\n\n- **Control**: practice as normal, no AI.\n- **GPT Base**: a standard GPT-4 chat assistant. Ask anything, get an answer.\n- **GPT Tutor**: the same underlying model, but constrained: it gave hints, asked questions, and would not hand over the solution.\n\nDuring practice, both AI groups did better than the control group. Not slightly better. GPT Base students got about **48% more practice problems right**. GPT Tutor students got about **127% more right**. If you stopped the study there, you would write a press release.\n\nThen came an exam. Closed. No AI for anyone.\n\nThe GPT Base students scored about **17% worse than students who had never had the tool at all**. The GPT Tutor students came out roughly level with the control group: the safeguards removed the harm, though they did not produce a lasting gain either.\n\n> Figure: The tool was taken away, and the groups separated. Bars (closed-exam score, control group = 100): Control, no AI in practice 100; GPT Tutor, hints and questions only 100; GPT Base, answers on request 83. Bastani and colleagues, about a thousand Turkish high-school students, mathematics, 2024, with the control group set to 100. During practice both assisted groups did better, and the gap was not small: the answering assistant raised practice accuracy by roughly half, the hinting one by more than double. Then came a closed exam with no tool for anyone. The variable that moved the exam score was not whether AI was present. It was whether it gave answers or gave hints, and that is the part you set in the prompt.\n\n## Read that carefully, because the obvious reading is wrong\n\nThe intuitive story is \"the AI group was lazy.\" That is not what happened. The AI group *worked and performed better* during practice. They were not doing less. They were doing something else.\n\nThe intuitive story two is \"AI is bad for learning.\" That is not what happened either: the two arms used the same model. The variable that moved the exam score was not whether AI was present. It was **whether it gave answers or gave hints**.\n\nThat is a much more useful finding, because it is something you control. You cannot change what model your school has access to. You can change what you ask it for.\n\n## Why practice performance and learning came apart\n\nThis is not a strange new effect of AI. It is a well-documented feature of how memory works, and it has a name: **desirable difficulties**, from Robert Bjork's work in the 1990s.\n\nThe short version: the conditions that make performance *during* study feel good and look good are frequently the conditions that produce the least durable learning. Conditions that feel slow and error-strewn (recalling from memory, spacing sessions out, mixing topics) feel worse and work better.\n\nAn AI that answers is a machine for making study feel good. Every step arrives clean. You never sit in the gap. And the sitting in the gap, it turns out, was not an obstacle to the learning. It was the learning.\n\nThere is a related result worth knowing: Roediger and Karpicke (2006) had students either reread a passage repeatedly or read it once and then try to recall it. Five minutes later, the rereaders won. A week later, the recall group remembered substantially more; and, notably, the rereaders had *predicted* they would do better. The feeling of readiness pointed the wrong way.\n\n## What is not settled\n\nBe honest about the state of the evidence, because you will meet people who overclaim in both directions.\n\nThe Turkey study is one study. One subject, one country, one model version, one age group. Maths is unusually well-suited to the design; nobody has shown the same numbers for essay writing or lab work. A 17% drop is large, and single large effects sometimes shrink when repeated.\n\nA widely shared 2025 MIT Media Lab preprint (\"Your Brain on ChatGPT\") reported weaker EEG connectivity in essay-writers using an LLM, and that most of that group could not quote a sentence from an essay they had finished minutes earlier. It is suggestive. It is also 54 participants and, at release, not peer-reviewed. Do not treat brain scans as proof of anything about your degree.\n\nWhat is solid is the underlying mechanism, and it has forty years behind it: **you remember what you generate, not what you read.** The Turkey study is a clean demonstration of that principle meeting a new tool.\n\n## The practical conclusion\n\nYou do not have to give up the tool. You have to change what you ask it to be.\n\nEvery time you are about to type a question, you are choosing between GPT Base and GPT Tutor. The model does not choose. You do, in the prompt. The next lesson is about how."},"check":{"question":"In the Turkey study, unguarded-tutor students solved 48% more practice problems correctly than the control group and then scored 17% worse on the unassisted exam. What does the hint-only arm add to that picture?","options":[{"text":"That practice scores were an unreliable measure, so neither group's practice gain tells you anything useful.","correct":false,"tempting":"Reads the result as discrediting the practice measure, when the point is that practice performance and durable learning genuinely came apart."},{"text":"That the harm came from how the tool answered rather than from AI use as such: same model, hints instead of solutions, large practice gains, no exam penalty.","correct":true},{"text":"That the unguarded group simply did less work, and the volume difference explains the exam gap.","correct":false,"tempting":"Assumes the AI group was lazier, but they performed better during practice, not worse: they were doing something else, not less."},{"text":"That the effect belongs to that particular model version, so a more capable model would remove the penalty.","correct":false,"tempting":"Treats model capability as the variable when the manipulated variable was answer-giving versus hint-giving, which a stronger model does not change."}]},"prev":{"slug":"two-kinds-of-help","title":"Two Kinds of Help","url":"https://addaly.com/learn/ai-for-students/two-kinds-of-help"},"next":{"slug":"what-a-model-actually-does","title":"What it is actually doing","url":"https://addaly.com/learn/ai-for-students/what-a-model-actually-does"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'What Happens To Your Memory', in AI for Students. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-for-students/what-happens-to-your-memory (Accessed: 31 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"What Happens To Your Memory","courseTitle":"AI for Students","module":{"index":1,"slug":"what-the-machine-is-doing","title":"What the machine is doing when it helps you"},"answer":null,"keyIdea":"The tutor that gave answers raised practice scores and lowered exam scores; the one that gave hints did neither harm.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure","figure":{"type":"bars","unit":"closed-exam score, control group = 100","title":"The tool was taken away, and the groups separated","series":[{"label":"Control, no AI in practice","value":100},{"label":"GPT Tutor, hints and questions only","value":100},{"label":"GPT Base, answers on request","value":83,"highlight":true}],"caption":"Bastani and colleagues, about a thousand Turkish high-school students, mathematics, 2024, with the control group set to 100. During practice both assisted groups did better, and the gap was not small: the answering assistant raised practice accuracy by roughly half, the hinting one by more than double. Then came a closed exam with no tool for anyone. The variable that moved the exam score was not whether AI was present. It was whether it gave answers or gave hints, and that is the part you set in the prompt."},"alt":"Figure: The tool was taken away, and the groups separated. Bars (closed-exam score, control group = 100): Control, no AI in practice 100; GPT Tutor, hints and questions only 100; GPT Base, answers on request 83. Bastani and colleagues, about a thousand Turkish high-school students, mathematics, 2024, with the control group set to 100. During practice both assisted groups did better, and the gap was not small: the answering assistant raised practice accuracy by roughly half, the hinting one by more than double. Then came a closed exam with no tool for anyone. The variable that moved the exam score was not whether AI was present. It was whether it gave answers or gave hints, and that is the part you set in the prompt."}},"blocks":[{"t":"h2","id":"the-study-you-should-know-about","runs":[{"text":"The study you should know about"}]},{"t":"p","runs":[{"text":"In 2024, researchers ran something close to the experiment every student privately wonders about. Bastani and colleagues worked with roughly a thousand high school students in Turkey, across grades 9 to 11, in maths."}]},{"t":"p","runs":[{"text":"Students were split three ways for their practice sessions:"}]},{"t":"ul","items":[[{"text":"Control","b":true},{"text":": practice as normal, no AI."}],[{"text":"GPT Base","b":true},{"text":": a standard GPT-4 chat assistant. Ask anything, get an answer."}],[{"text":"GPT Tutor","b":true},{"text":": the same underlying model, but constrained: it gave hints, asked questions, and would not hand over the solution."}]]},{"t":"p","runs":[{"text":"During practice, both AI groups did better than the control group. Not slightly better. GPT Base students got about "},{"text":"48% more practice problems right","b":true},{"text":". GPT Tutor students got about "},{"text":"127% more right","b":true},{"text":". If you stopped the study there, you would write a press release."}]},{"t":"p","runs":[{"text":"Then came an exam. Closed. No AI for anyone."}]},{"t":"p","runs":[{"text":"The GPT Base students scored about "},{"text":"17% worse than students who had never had the tool at all","b":true},{"text":". The GPT Tutor students came out roughly level with the control group: the safeguards removed the harm, though they did not produce a lasting gain either."}]},{"t":"h2","id":"read-that-carefully-because-the-obvious-reading-is-wrong","runs":[{"text":"Read that carefully, because the obvious reading is wrong"}]},{"t":"p","runs":[{"text":"The intuitive story is “the AI group was lazy.” That is not what happened. The AI group "},{"text":"worked and performed better","i":true},{"text":" during practice. They were not doing less. They were doing something else."}]},{"t":"p","runs":[{"text":"The intuitive story two is “AI is bad for learning.” That is not what happened either: the two arms used the same model. The variable that moved the exam score was not whether AI was present. It was "},{"text":"whether it gave answers or gave hints","b":true},{"text":"."}]},{"t":"p","runs":[{"text":"That is a much more useful finding, because it is something you control. You cannot change what model your school has access to. You can change what you ask it for."}]},{"t":"h2","id":"why-practice-performance-and-learning-came-apart","runs":[{"text":"Why practice performance and learning came apart"}]},{"t":"p","runs":[{"text":"This is not a strange new effect of AI. It is a well-documented feature of how memory works, and it has a name: "},{"text":"desirable difficulties","b":true},{"text":", from Robert Bjork’s work in the 1990s."}]},{"t":"p","runs":[{"text":"The short version: the conditions that make performance "},{"text":"during","i":true},{"text":" study feel good and look good are frequently the conditions that produce the least durable learning. Conditions that feel slow and error-strewn (recalling from memory, spacing sessions out, mixing topics) feel worse and work better."}]},{"t":"p","runs":[{"text":"An AI that answers is a machine for making study feel good. Every step arrives clean. You never sit in the gap. And the sitting in the gap, it turns out, was not an obstacle to the learning. It was the learning."}]},{"t":"p","runs":[{"text":"There is a related result worth knowing: Roediger and Karpicke (2006) had students either reread a passage repeatedly or read it once and then try to recall it. Five minutes later, the rereaders won. A week later, the recall group remembered substantially more; and, notably, the rereaders had "},{"text":"predicted","i":true},{"text":" they would do better. The feeling of readiness pointed the wrong way."}]},{"t":"h2","id":"what-is-not-settled","runs":[{"text":"What is not settled"}]},{"t":"p","runs":[{"text":"Be honest about the state of the evidence, because you will meet people who overclaim in both directions."}]},{"t":"p","runs":[{"text":"The Turkey study is one study. One subject, one country, one model version, one age group. Maths is unusually well-suited to the design; nobody has shown the same numbers for essay writing or lab work. A 17% drop is large, and single large effects sometimes shrink when repeated."}]},{"t":"p","runs":[{"text":"A widely shared 2025 MIT Media Lab preprint (“Your Brain on ChatGPT”) reported weaker EEG connectivity in essay-writers using an LLM, and that most of that group could not quote a sentence from an essay they had finished minutes earlier. It is suggestive. It is also 54 participants and, at release, not peer-reviewed. Do not treat brain scans as proof of anything about your degree."}]},{"t":"p","runs":[{"text":"What is solid is the underlying mechanism, and it has forty years behind it: "},{"text":"you remember what you generate, not what you read.","b":true},{"text":" The Turkey study is a clean demonstration of that principle meeting a new tool."}]},{"t":"h2","id":"the-practical-conclusion","runs":[{"text":"The practical conclusion"}]},{"t":"p","runs":[{"text":"You do not have to give up the tool. You have to change what you ask it to be."}]},{"t":"p","runs":[{"text":"Every time you are about to type a question, you are choosing between GPT Base and GPT Tutor. The model does not choose. You do, in the prompt. The next lesson is about how."}]}],"check":{"question":[{"text":"In the Turkey study, unguarded-tutor students solved 48% more practice problems correctly than the control group and then scored 17% worse on the unassisted exam. What does the hint-only arm add to that picture?"}],"options":[{"text":[{"text":"That practice scores were an unreliable measure, so neither group’s practice gain tells you anything useful."}],"correct":false,"tempting":[{"text":"Reads the result as discrediting the practice measure, when the point is that practice performance and durable learning genuinely came apart."}]},{"text":[{"text":"That the harm came from how the tool answered rather than from AI use as such: same model, hints instead of solutions, large practice gains, no exam penalty."}],"correct":true,"tempting":null},{"text":[{"text":"That the unguarded group simply did less work, and the volume difference explains the exam gap."}],"correct":false,"tempting":[{"text":"Assumes the AI group was lazier, but they performed better during practice, not worse: they were doing something else, not less."}]},{"text":[{"text":"That the effect belongs to that particular model version, so a more capable model would remove the penalty."}],"correct":false,"tempting":[{"text":"Treats model capability as the variable when the manipulated variable was answer-giving versus hint-giving, which a stronger model does not change."}]}],"order":[2,3,0,1],"why":null},"prev":{"slug":"two-kinds-of-help","title":"Two Kinds of Help","minutes":7},"next":{"slug":"what-a-model-actually-does","title":"What it is actually doing","minutes":9}}}