How Should I Use AI to Study and Learn Effectively?

It is 11 p.m., the problem set is due at nine in the morning, and the chat window is already open. Question four stopped making sense an hour ago. The tool will answer in ten seconds with something that reads like the work of a student who understands it, and you have to decide whether to type the assignment in or type your own mess in first.

The sentence worth carrying off this page: the same model, opened at a different point in your study process, does opposite things to your learning. Open it before you have tried, and it ends the thinking, leaving you a clean file and a thinner memory. Open it after an attempt, to quiz you, to explain what you missed, to show where your reasoning broke, and it works like a tutor. The second thing to hold on to: work that came out fine is not evidence that you learned anything. A polished submission shows what you and the model can do together. The exam, the placement, and the next course in the sequence test what you can do alone, which is what the research calls learning.

The short version

1. Write your own attempt before you open the chat, even a wrong one, and tell the tool to withhold the answer until you have one. This is the move with the strongest evidence behind it.

2. Ask for teaching: hints, explanations, worked steps, quizzes, corrections on what you produced. Keep the arguing, the sourcing, and the sentences for yourself.

3. Stay in the conversation past the first reply. Push back, ask why, ask what you got wrong.

4. Check anything you will be graded on against a source you can name, and treat fluent wording as no evidence at all.

5. End the session with the tool closed: answer from memory, out loud, and see what is missing. You find out tonight instead of in the exam hall.

Start the session with your own attempt

Akgun and Toker (2026) gave 89 undergraduates the same adaptive pretesting session and the same instructions, then seven weeks of different practice. The group whose AI probed misconceptions, demanded elaboration of thin answers, and advanced only on real conceptual engagement scored highest on the delayed posttest (M = 78.19) and put in the most practice effort (M = 0.85). Free chat with the same model scored 67.28 and 0.49, a difference of d = 0.92.

The attempt does the work even when the attempt is bad. Writing an answer activates what you already know, and it shows you the edge of it. Then the correction lands on something.

A policy you can paste into any chatbot tonight: explain this as though I will be tested on it, quiz me one question at a time, do not give me the answer until I have attempted it, tell me what I got wrong and why. That sentence reproduces the arrangement that produced the 78.19, and it costs nothing.

Ask for a tutor and name the job

Vague requests get vague help. Viberg, Feldman-Maggor and Wong (2026) interviewed 20 STEM students about how they sought help and found a four-stage sequence worth copying: deciding whether help was needed, choosing whom to ask, deciding what kind of help to request, and judging what came back. These students went to the model first because it was the lowest-barrier option, in one student's words "first ChatGPT, then classmates, and lastly teachers," and they asked for hints, step-by-step guidance, and concept explanations rather than finished solutions. Another described the chatbot as "a hint, an assisting tool, but not the standalone solution."

What that sounds like typed out: "I keep getting the wrong sign when I integrate by parts; here is my attempt; where does it break?" A named sticking point gets a useful reply, and "help me with this assignment" gets a draft. Ask for a worked example of one problem type, study its shape, then do the next one alone. Ask for the concept explained as though you will be tested on it, then explain it back without looking. Ask for questions, one at a time, and hold your answer until you have committed to one.

The pattern is the tool working on your understanding. Summarizing a chapter you have read, translating a source, reformatting your notes, and building practice cards are all reasonable. Deciding what your essay argues, finding the source, and writing the sentence are yours.

Stay in the conversation past the first reply

The gap between what students intend and what they do is the most consistent finding in this literature. Yan and colleagues (2026) interviewed 38 undergraduates in Japan and China while the students walked through their own chat histories for unsupervised essay assessments. Thirty-one of the 38 said they meant to use AI as a learning assistant, and they still described overreliance, mental complacency, and fast forgetting. One said "the speed at which you forget it is also very fast." A smaller group handed the whole task over, teacher assigns, student passes it to GenAI, GenAI generates, student checks formatting, student submits, and one participant described that as having "completely replaced my brain."

Only 8 of the 38 worked as what the authors call cognitive partners. They did not use less AI. Their total effort held steady and changed shape. One moved effort from searching to quality checking. Another wrote short reflection notes after every AI session to slow the fading of information retrieved instantly.

The habits that separate the two groups, as instructions to yourself:

• Paste the problem and your attempt, not the assignment title. 76.32% of these students followed an ask, get an answer, stop pattern, typically pasting the assessment title without saying what they needed and then resubmitting the same prompt when the reply disappointed them. Sustained, iterative dialogue appeared in 23.68%, almost all of them cognitive partners.
• Keep it inside your own reading and writing. 78.94% used AI either before starting or after drafting, and that detachment is the signature of the group that learned least.
• Go past the first reply. The Help-Seeking research shows the same shape. Amoozadeh and Alipour (2026) classified 830 prompts from 72 students across two programming tasks. The same few moves kept recurring: assertions that reported confusion, verification prompts asking whether something was right, and procedural prompts asking for the next step. Comparison, prediction, and feature specification, the moves that make you think, stayed rare. Your own prompt history will predict your learning better than your intentions will.

Check the output against something you can name

"Critical use" is a phrase that means little until you separate the parts. Wei and Shang (2026) separate seven: epistemic evaluation, whether you start checking, how well you checked, whether the check succeeded, what you did with the output, how you performed on the task, and what you learned independently. Two of those matter more than the rest. Checking is not the same as successful checking, and a careful process can end inconclusive while a sloppy one lands on the right answer by accident. Calibrated reliance means your decision to accept or reject matched the actual quality of the output, which you cannot read off how confident the answer sounded.

Three findings make the problem concrete. Dávila et al. (2025) gave learners advice that was correct about half the time and found that how much weight they gave it varied with their prior knowledge and gender, so trust was not tracking accuracy. Zheng et al. (2025) documented a "failed application" pattern in which a student followed correct guidance correctly and still reached a wrong answer. Confidence is not a guide either: Rheu and Cho (2025) found that understanding how language models work was associated with more self-reported fact-checking, while some forms of confidence and feature knowledge were associated with less of it.

So verify anything assessed against a source you can name. The students in the Viberg study checked outputs against their coursework and their instructors, one explaining "I only go to the TA if we can't tell whether ChatGPT is making things up." When you genuinely cannot tell, mark the point unresolved and ask a person. Adopting a claim to get the page finished is how a small uncertainty becomes a wrong answer you have rehearsed.

Decide your ground rules before you need them. Pérez-Portabella and colleagues (2026) surveyed 151 undergraduates and found that ethical judgments were necessary conditions for intending to use a large language model for exam preparation, with consequentialist and deontological reasoning predicting that intention, and intention strongly predicting actual use. What you decide in advance about what you will and will not do is what your behavior follows at 11 p.m.

Let it handle logistics

Speed has legitimate uses. Kawamura (2026) built systems that adapt spoken playback to the difficulty of each segment, averaging about 1.30x, and generate multimodal video summaries that cut viewing time by 53% with no statistically significant difference in quiz scores. Summarizing, translating, and planning a study schedule sit in the same category. Moving through material faster is fine. Skipping the work that changes you is the part to avoid.

The distinction is where the tool sits relative to the thinking. A summary of a chapter you have already read saves time. A summary you read instead of the chapter replaces the reading.

Finish with the tool closed

Judging a session by how it felt or how polished the output looks will mislead you. Independent learning means retention, transfer, unaided performance, and finding your own errors once the support is gone. Three tests need no researcher:

• Close the tool and answer. If you cannot produce the explanation or the solution with the chat shut, you have not learned it yet.
• Test after a delay. The seven-week posttest is what separated the conditions in the statistics study. A score taken with the conversation still on screen tells you very little.
• Explain it out loud, from memory, with no notes. If you need the tool's phrasing, the understanding is not yours yet.

Retrieval needs a schedule. Zhang (2026) describes Memdora, an AI spaced-repetition system built on the finding that roughly 70% of newly learned material is forgotten within 24 hours without review. It makes cards from whatever you are reading and schedules them with FSRS-6. Be precise about what that is worth: the paper reports better retention than traditional flashcard tools, and its contribution is a design and interaction taxonomy rather than a controlled retention trial. Take the spacing and retrieval as established and the specific interaction design as promising. If the app you use reports outcomes per item, read them, because per-card learning tells you more than a streak count.

What students get wrong about this

Hours in the chat are study time. The 89 undergraduates in the pretesting study spent seven weeks in different conditions with the same model. The condition that scored 78.19 was the one where the tool refused to hand over answers. The difference was the pushback.

A good answer means I understood it. Yan's students judged their own use as effective while reporting fast forgetting, and 31 of 38 said they meant to use it as a learning assistant. The output is evidence about the model. Your closed-book answer is evidence about you.

Asking for the answer is the efficient move. It is efficient at producing the artifact. In the Yan sample, the group that handed the task over produced the work and kept the least; the students who reported forgetting fastest were the much larger group who believed they were using the tool well.

Fluent means accurate. Confidence in the answer tells you nothing about its quality, and the same fluency is what makes a wrong claim easy to adopt. Verification has to be against something outside the chat.

Rereading the answer is checking. Wei and Shang's separation matters here. Whether you checked and whether the check worked are different variables, and only the second one protects your grade.

The objections you will hear

"I'm paying for this, so why not use it?" A subscription buys something better than answers, and the answers are the cheap part. You are paying for a patient tutor that will explain the same idea five ways, quiz you until you can do it cold, and tell you what went wrong with your own attempt. Most students could not have that at 11 p.m.

"Everyone else does it this way." Most of them do it in the way these studies show working badly: 76.32% ask, get an answer, and stop; 78.94% never bring the tool into the part of the work that would have changed them. The 8 of 38 who worked as cognitive partners were not a gifted lane. Their distinguishing feature was a rule you can adopt tonight, which is that effort moves rather than disappears.

"It's faster." It is. That speed is the efficiency paradox: the session finishes sooner and the thing the task was for declines. Free chat in the retrieval study did less practice work (M = 0.49 against 0.85) and scored lower seven weeks later (67.28 against 78.19, d = 0.92). You will be faster tonight and slower in the exam, and you get to pick which one you want.

"I don't know enough to attempt it first." The attempt is how you find the edge of what you know, and pretesting helped even when the first answers were wrong. Write down what you think the question is asking and where you get stuck, then open the chat with that.

"My instructor banned it, so there is nothing to learn here." The 38 students in the Yan study reported that their instructors forbade copying and gave almost no concrete guidance, and that vacuum pushed even well-intentioned students toward outsourcing. A ban gives you nothing to do on the day the tool is allowed, and the habits on this page are what you carry into those courses.

What to do this week

1. Pick the course where you currently paste first. For the next problem set, write a one-paragraph attempt before you open any chat, and paste the attempt along with the question.

2. Put the four-line instruction from the first section into the custom instructions of whichever tool you use, and leave it there.

3. Before your next submitted piece of work, ask the tool to quiz you on the material for ten minutes with the answers withheld. Note what you could not answer.

4. Take one topic you think you know and explain it out loud with the laptop shut. Whatever you stumble on is your revision list.

5. Check every claim you plan to submit against a named source, and write "unresolved" beside the ones you cannot.

Where to go next

How Do I Keep Students from Over-Relying on AI? covers the habits that keep the tool from replacing your judgment, How Do I Teach Students to Verify AI Output? handles the checking side in more detail, and How Is AI Impacting Students? sets out what the wider evidence says about students and these tools. If you also teach, Does Using AI Actually Help My Students Learn? is the evidence on durable learning, and How Does AI Affect Student Anxiety and Well-Being, and What Can We Do? covers what the pace of this change does to students who are trying to keep up.