AI for studying: what helps you learn, and what just hands you answers (a four-part test for any tool)

The best AI for studying passes four tests: it makes you attempt before it answers, it checks your work against something verified rather than its own guess, it cites where its answer came from, and it says plainly when it does not know or has no source. A tool that hands over a finished answer in one message fails the first test, and the research now shows why that costs you marks later. Here is the test, then the six guides in this section.

Why "best AI for studying" is the wrong question

Every page that ranks AI study tools is answering a question that does not matter. Any general assistant can write the answer to your homework. What differs between tools, and what decides your mark, is what happens to your learning while it does.

Bastani and colleagues (2025) measured it. Nearly a thousand high-school students in a maths course practised with one of three set-ups: an unrestricted AI assistant that would answer anything, an assistant built to coach, which gave hints and explanations but not final answers, or no AI. During practice the unrestricted group did best, with practice scores 48 per cent higher than the no-AI group's. Then came the exam, with no tool for anyone, and the unrestricted group scored 17 per cent lower than the students who had never had AI. The coaching group, whose practice scores had been 127 per cent higher, scored about the same on the exam as the no-AI group. The same model sat under both assistants. The difference was the design.

That study is the spine of this section. "Which AI study tool" is the wrong question; "what does this tool make me do" is the right one. The four tests below are that question, made runnable.

The four-part test

Ten minutes, your own homework problem, any tool. Each test is a thing to type and a thing to look for.

Test 1: Does it make you retrieve?

Type a problem and nothing else. A pass asks what you have tried, or gives a hint and waits. A fail writes the solution. Pulling an answer out of your own head is what makes it stick (Roediger and Karpicke 2006): the attempt is the learning, and a tool that removes the attempt removes the learning. What an AI tutor should do takes this test to five questions.

Test 2: Does it check, or guess?

Give it your working with one wrong line you planted. A pass finds the line and names the rule it broke, against a verified solution. A fail agrees with you, or corrects a line that was right. In our reading of 320 one- and two-star public reviews of seven answer and flashcard apps (2026-10-18), the largest complaint theme, 105 of the 320, was "the answer was wrong and I trusted it". A generated answer looks the same whether it is right or wrong. How to get unstuck without losing the learning is the method, with or without a tool.

Test 3: Does it cite?

Ask where a claim came from. A pass names a source you can open: the course's own note, the textbook page it was given, a passage it quotes. A fail says "studies show", or gives a page you cannot find.

Test 4: Does it say when it does not know?

Ask about a page you did not give it. A pass says it was not given that page. A fail summarises a page it has never seen. This test tells you what the other three are worth, and it matters most for your teacher's rule: an invented citation in your work is not your work.

What the research says about AI tutoring that coaches

The Bastani study is the warning. The other side is the same lesson from the other direction: design decides.

Kestin and colleagues (2025) built an AI tutor for an introductory university physics course around teaching principles rather than around answering: one step at a time, the student asked to try, the course's own problems. In a randomized comparison, students who worked with the tutor at home learned more than students who had the same material in a well-run active-learning class, with learning gains the authors report as more than double the in-class group's, in less time. It was one course and one term, measured right after the lesson rather than weeks later, so read it as strong evidence that a coaching design can work, not that any AI tutor does.

The older record agrees. VanLehn (2011) reviewed decades of intelligent tutoring systems, software that gives feedback step by step, and found their average effect close to that of human tutors on the same kinds of measure. They worked not by explaining better but by letting the student do each step and correcting the step that failed.

A tool that coaches can match a human tutor; a tool that answers makes the exam worse. The design is the whole difference.

The honest list: what AI is good for in studying, today

Good for, when you check the output:

• Turning a recording or a PDF into questions you can test yourself on tonight: lecture and video to notes shows how to make the notes test you rather than summarise at you.
• Drafting flashcards you then verify. A flashcard maker can draft fifty cards in a minute, and some of them will be wrong; how to use a flashcard maker is the checking method.
• Explaining a step after your attempt. "Why does the sign flip here?" after you tried is a tutor's job.
• Checking your working line by line, when it checks against something verified.

Bad for: being the answer key (a generated answer is a guess that looks like a fact); grading an essay for a mark (it can say a paragraph is unclear, not what your teacher's rubric will say); replacing the textbook as the source (a tool should quote the source, not be it).

Is it cheating?

Using AI to learn is studying; submitting what it wrote is cheating; and the teacher's rule for the course decides where the line falls in that class. The study above shows why the two lines are one line: the students who were handed answers were the students who did worse on the exam. The guide on where the line is, course by course, has a table of learning-with-it against submitting-it and an email to send when the rule is not written down.

Worked example: running the four tests on one question

A Grade 12 calculus student types our problem, "Differentiate y = x² sin x", into a tool, with their own wrong attempt underneath: y′ = 2x cos x. Each factor was differentiated and the results multiplied; the product rule was skipped.

Test 1, pass: "Two functions are multiplied here. Which rule applies to a product? Try differentiating just the first factor, then just the second." Fail: "y′ = 2x sin x + x² cos x."

Test 2, pass: "Your line y′ = 2x cos x differentiated each factor and multiplied the results, which is not a rule. The product rule: derivative of the first times the second, plus the first times derivative of the second, so y′ = 2x · sin x + x² · cos x." Fail: "That looks right."

Test 3, pass: "The product rule is in your course's own note on derivative rules," with the note named so the student can open it. Fail: "This is a standard result."

Test 4, the student asks what page 143 of their textbook says, without pasting it. Pass: "I wasn't given page 143. Paste it and I can read it with you." Fail: a paragraph about what page 143 probably covers.

The answer, checkable by hand: differentiate x² to get 2x and keep sin x; keep x² and differentiate sin x to get cos x; add. y′ = 2x sin x + x² cos x.

The guides in this section

• Is using AI to study cheating? Where the line is, course by course, with the learning-vs-submitting table and the one-sentence email to a teacher.
• What an AI tutor should do. The four properties and the five questions that expose a tutor that only answers.
• Math solvers and homework help. How to get unstuck without losing the learning: attempt, hint, find the line, redo, log.
• Lecture and video to notes. Turning a recording or a video into notes that test you, and checking what the AI wrote against what was said.
• Flashcard makers. Getting cards drafted from your material and catching the bad ones before you study them.
• How to choose a study app. A ten-point checklist for any study app, from the attempt gate to where your data lives.

The four tests, on one screen

1. Does it make you retrieve? Type a problem; a pass hints or asks, a fail answers.
2. Does it check, or guess? Plant a wrong line; a pass names it against a verified solution.
3. Does it cite? A pass names a source you can open.
4. Does it say when it does not know? Ask about a page you did not give it; a pass says so.

Chapter 7 of the free playbook, "Use AI that teaches", is this test on one page, and the ten-point checklist for any study app extends it to the rest of what a tool does with your work. YSR Studies is built to pass it: the tutor coaches and reveals the course's verified answer only after your attempt, cites only what it was shown, and says when it has no course text.