How to Use AI to Study Without Skipping the Actual Learning

The short answer

Attempt every problem honestly before asking AI anything. Then use it to hint, quiz and check, never to solve. Keep a tutor mode switched on, spread reviews across several days, and end each session with a short recall test while the chat is closed.

17%

lower scores once unrestricted GPT-4 access was removed (PNAS, 2025)

2x+

median learning gains from a purpose-built AI physics tutor (Harvard, 2025)

61% vs 40%

one-week recall for self-testing versus rereading (2006)

47%

of student chats with Claude sought answers with minimal engagement (Anthropic, 2025)

1. Attempt first, then ask for one hint

Give every problem at least five honest minutes before opening the chat. When stuck, paste the problem and the attempt, and ask for the next nudge only.

Prompt

Act as a tutor, not a solver. Here is the problem and the work so far: [paste]. Do not give the answer or the full method. Point to the first step that goes wrong and give one hint. Wait for a reply before the next hint.

Why it works: This mirrors the GPT Tutor setup in the Bastani study: the same model, prompted to guide rather than hand over answers.

2. Turn class notes into a recall quiz

Rereading feels productive and does little. Pulling facts out of memory feels harder and works. AI removes the chore of writing the questions.

Prompt

Using only the notes below, ask 8 short-answer questions one at a time. Wait for each answer, mark it, and explain any mistake in two sentences. Save the hardest three for last. Notes: [paste]

Why it works: This is the retrieval practice behind the 61% versus 40% result.

3. Explain it back, then get graded

After finishing a topic, write an explanation from memory as if teaching a younger student. Then hand it over for marking, not rewriting.

Prompt

Below is an explanation of [concept] written from memory. Grade it out of 10 like a strict examiner. List what is missing or wrong. Do not rewrite it.

Why it works: Self-explanation earned a moderate rating in the Dunlosky review, and writing it personally sidesteps the recall gap seen in the MIT study.

4. Spread reviews across days

One long night of AI-assisted cramming delivers the five-minute advantage and the one-week collapse. Spacing fixes that. Students who would rather not manage a calendar can use an app built around flashcards that resurface on a spaced schedule, editing the generated cards for accuracy.

Prompt

The exam is on [date]. Topics: [list]. Build a calendar that brings each topic back at growing gaps (1, 3, 7 and 14 days) using short recall quizzes, not rereading.

Why it works: Distributed practice is one of only two techniques rated high utility by Dunlosky's team.

5. Ask for mixed problem sets

Textbook exercise blocks reveal the method before the first problem. Mixed sets force the harder, exam-style decision: which method fits this one?

Prompt

Write 10 problems mixing [chain rule, product rule, quotient rule] in random order. Do not label the method. Give answers only after all 10 are submitted.

Why it works: Interleaved practice received a moderate rating in the same 2013 review.

6. Check finished work, not unfinished work

Comparing a completed attempt with a worked solution is efficient, provided the attempt comes first. A step-by-step homework solver such as Gauth makes line-by-line comparison quick, though its answers can still be wrong on harder problems, so disagreements deserve a second source.

Prompt

Here is a completed solution: [paste]. Do not solve the problem. Say whether the final answer is correct and, if not, name the first step where it goes wrong.

Why it works: Feedback needs an attempt to work on. Without one, there is nothing to learn from.

7. Keep practice tied to the real course

Chatbots drift toward generic textbook questions. Uploading the actual slides keeps quizzes aligned with the lecturer's exam. Apps such as Jungle, which turn lecture slides and videos into practice questions, also link every answer back to its source slide.

Prompt

Using only the attached slides, write 10 exam-style questions. Cite the slide number for every answer. If something is not covered in the slides, say so instead of guessing.

Why it works: Grounding the AI in a source cuts invented content and makes every answer checkable.

After a month of daily use

After a month of studying with AI open in the next tab, the conclusion is calmer than the headlines on either side. The chatbot was never the problem or the miracle. The habit around it was.

Sessions that began with "solve this" produced tidy pages and empty recall by breakfast. Sessions that began with an attempt, then a hint, then a quiz, were slower and noticeably more tiring, and those were the topics that held up once the laptop was shut. None of the three tutor modes stood far above the others; the deciding factor every time was whether it stayed switched on.

The research says the same thing more carefully. The tool that helps and the tool that harms can be the very same model. What changes is who does the thinking. Tutor mode on, real course material uploaded, no answers before an honest attempt, and a closed-chat recall check to finish: that setup costs a little speed and pays it back on exam day.