Learning With an AI Tutor: Practice, Not Just Answers

An AI chat tool will answer almost any question instantly and fluently. That is exactly what makes it risky as a tutor: reading a good answer feels like learning, but the effort you skip is often the part that builds memory and skill. Set up differently, the same tool can make you do more of the work that matters. Here is how.

Answers feel like learning; retrieval is learning

One of the best-known experiments on this is Roediger and Karpicke's 2006 study, "Test-enhanced learning". Students studied prose passages and then either restudied them or took recall tests without feedback. A final test came 5 minutes, 2 days or 1 week later. At 5 minutes, restudying did better. On the delayed tests, the students who had been tested remembered substantially more, even though repeated studying had made the other group more confident in their memory.

Reading an AI's explanation is a form of restudying. It is pleasant, it raises your confidence, and it is weak practice. Pulling the answer out of your own head, then checking it, is the strong kind. The rest of this guide is about making the tutor serve the second kind.

What happened when students used a chatbot without guardrails

A field experiment published in PNAS in 2025 by Hamsa Bastani and colleagues tested this with nearly 1,000 high school math students in Turkey. One group practiced with "GPT Base," a standard chat interface. Another used "GPT Tutor," set up with teacher-designed hints and instructions not to hand over answers.

During practice, grades rose 48% with GPT Base and 127% with GPT Tutor. On the later exam, taken without AI, the GPT Base group scored 17% lower than students who never had access. The GPT Tutor group's exam results were not significantly different from the control group.

Two lessons follow. The design of the interaction matters: open access to answers boosted practice scores and hurt learning. And guardrails removed the harm in this study but did not produce exam gains on their own. You have to do the retrieval work either way. The good news is that you can impose your own guardrails with a few instructions.

Ask to be quizzed, not taught

Start sessions with rules that put you on the answering side. Paste something like this, with your topic and source filled in:

I'm learning [topic] from [source: chapter, course notes, official docs].
Quiz me one question at a time and wait for my answer before saying anything else.
After I answer, say whether it's right, what I missed, and which part of the source to reread.
Mix question types: define it, explain why, apply it to a new example, spot the error.
If I say "I don't know," give me a hint, not the answer.
Every fifth question, return to something I got wrong earlier.

A few habits make the quiz work harder for you:

Answer in full sentences. Typing a real answer forces actual recall instead of a vague sense of recognition.
Rate your confidence before you see feedback: sure, unsure or guess. Confident wrong answers are the most valuable ones to find.
Close your notes. An open textbook turns a quiz back into reading.

Use Socratic prompting for understanding

Quizzing works well for facts and procedures. For concepts, try having the tutor ask the questions that lead you to the idea, rather than explaining it.

I want to understand [concept]. Don't explain it to me.
Ask me questions, one at a time, that lead me to work it out myself.
If my reasoning goes wrong, ask a question that exposes the problem instead of correcting me.
When I get there, ask me to state the idea in my own words, then check my statement.

Other moves that keep you doing the thinking:

Teach it back. Explain the concept to the tutor as if it were a beginner, and ask it to find gaps or errors in your explanation.
Fill the missing step. Ask for a worked example with one step removed, and supply it yourself.
Find the bug. Ask for a solution with one deliberate mistake, and locate it.
Get a twin problem. After you miss something, ask for a new problem that tests the same idea with different numbers or context.

Check the tutor against sources

Models can state wrong things in the same confident tone as right things, and a tutor that is wrong teaches you the error. Build checks into the routine:

Anchor to a source you trust. Work from a textbook, course notes or official documentation. Paste the relevant section where its terms allow, and tell the tutor to quiz only from that material and to flag any question that goes beyond it.
Ask where it came from. When the tutor states a fact, ask which passage supports it, then confirm the passage exists and says that.
Do not concede too fast. If the tutor marks you wrong and you think you are right, check the source before accepting the correction. Spot-check its "correct" verdicts too.
Verify anything you plan to memorize. Numbers, dates, formulas and definitions deserve a look at the original before they go into long-term memory.

Avoid copy-paste learning

The failure mode is easy to slip into: paste the problem, read the solution, feel that it makes sense, move on. Warning signs include finishing sessions with polished work you could not reproduce, and freezing when you face a blank page.

Rules that prevent it:

Attempt first. Write your own attempt before asking anything, even a bad one.
Ask for the smallest hint. "Give me a hint, not the next step" keeps the problem yours.
Close the chat and redo it. After getting help, solve a fresh variant with the tool closed.
Separate doing from learning. If you need the output for work, that is fine, but do not count it as practice.
Come back cold. Redo yesterday's hardest problem before starting anything new.

What tutoring research does and does not promise

AI tutoring is often pitched with Benjamin Bloom's 1984 paper, "The 2 Sigma Problem", which reported that students tutored one-to-one performed about two standard deviations better than students taught in a conventional class. Treat that number with care. As Paul von Hippel explains in Education Next in 2024, it came from small experiments by Bloom's doctoral students with 4th, 5th and 8th graders learning probability or cartography, measured on tests of the specific material taught. Tutored students also got mastery learning with corrective feedback, and tutoring replaced their classroom instruction entirely.

Later research, which von Hippel also reviews, found much smaller effects. A 2020 meta-analysis of randomized studies by Nickow, Oreopoulos and Quan found an average tutoring effect of 0.37 standard deviations, and none of the 96 studies they reviewed produced a two-sigma effect.

Kurt VanLehn's 2011 review found an effect size of 0.79 for human tutoring and 0.76 for step-based computer tutoring systems, so well-built software came close to human tutors in the studies he compared. Tutoring helps; "two sigma" is not a realistic expectation.

What made tutoring work in those studies was frequent checks, immediate correction and not moving on until you had it. Those are things you can ask an AI tutor to do. Set the rule at the start of every session: you answer first, and the tutor checks.