How to Use AI for Studying (Without Undermining Your Learning)

There is a real tension in using AI to study. The techniques that feel most efficient — having material summarised, getting answers explained — are often the ones that produce the least durable learning. Meanwhile the techniques that work are the ones that feel harder.

This is not a moral point about cheating. It is a mechanical one: memory is built by retrieval effort, not by exposure. Reading a good summary feels productive and creates very little retention. Struggling to recall something creates a lot.

So the useful question is: what can AI do that increases retrieval effort rather than removing it? Quite a lot, actually.

The principle in one table

Every workflow below is designed to keep you on the right-hand column.

Workflow 1: active recall testing

The highest-value study use of AI, and the most underused. You are not asking it to teach — you are asking it to test.

The prompt:

Here are my lecture notes on [topic]: [paste notes]

Quiz me using active recall. Rules:
- Ask ONE question at a time, then wait for my answer
- Start with recall questions, then move to application questions
- After each answer: tell me what I got right, what I missed, and what
  I was vague about — vagueness counts as a gap
- Do not give me the answer if I say "I don't know" — give a hint first
- After 10 questions, list the specific concepts I should revisit

Begin with question 1.

Why it works. Three reasons. First, retrieval practice is one of the most robustly supported findings in learning research — testing yourself outperforms re-reading by a wide margin. Second, "one at a time, then wait" prevents the model from dumping all ten questions with answers, which converts a test into a reading exercise. Third, "vagueness counts as a gap" targets the illusion of knowing: the feeling of familiarity that collapses under exam conditions. A vague answer that gets marked correct teaches you nothing.

The hint rule matters too. Getting the answer immediately after failing to recall it removes the effortful part, which is where the learning was.

Workflow 2: the Feynman check

You explain the concept; the AI finds the holes. This inverts the usual direction and it is where most students discover they understood less than they thought.

The prompt:

I'm going to explain [concept] in my own words. Your job is to find
the gaps, not to praise me.

My explanation: [your explanation, written from memory]

Tell me:
1. What did I get wrong or state imprecisely?
2. What did I leave out that actually matters?
3. Where did I use a term without demonstrating I understand it?
4. What follow-up question would expose a shallow understanding here?

Be blunt. Do not soften it.

Why it works. Writing an explanation from memory is retrieval practice plus organisation — you cannot produce a coherent explanation without an internal structure, so gaps surface as you write. Point 3 is the sharpest: using correct terminology while not understanding it is the most common form of false confidence, and it is very detectable in writing. Point 4 gives you your next test.

The "be blunt" instruction is functional, not stylistic. Models default to encouraging feedback, which is useless here.

Workflow 3: worked-example laddering for problem sets

For maths, physics, chemistry, statistics, engineering — anything procedural. Asking for solutions is the classic trap: you can follow a solution perfectly and still be unable to produce one.

The prompt:

I'm working on this problem: [paste problem]

Do NOT solve it. Instead:
1. Tell me which concept or technique this problem is testing
2. Give me the FIRST step only, and explain why that step
3. Then stop and let me continue

If I get stuck later, give me a hint — not the next step.

Then, after you have finished:

Here's my full solution: [paste your work]

Check it. For each error: mark whether it's a conceptual
misunderstanding, a procedural slip, or an arithmetic mistake.
I want to know which kind of mistake I'm making most.

Why it works. The first prompt preserves the productive struggle while removing the failure mode where you stare at a blank page for forty minutes. The second is the genuinely valuable half: classifying your error type tells you what to fix. Conceptual errors mean going back to the material. Procedural slips mean more practice. Arithmetic mistakes mean slowing down. Students routinely treat all three as "I need to study more," which is the wrong response to two of them.

Important caution: AI assistants are unreliable at arithmetic — they predict what a calculation looks like rather than performing it. Use them to check your method, and verify numbers yourself or with a calculator. Ask for the reasoning steps, not the final number.

Workflow 4: turning a dense text into an interrogation

For reading you must actually understand — papers, primary sources, textbook chapters. The instinct is to ask for a summary. Resist it: the summary replaces the reading.

Use this instead:

Here is a section of a paper I need to understand deeply: [paste]

Don't summarise it. Instead:
1. List the questions this text answers
2. List the claims it makes that depend on evidence not included here
3. Identify the single sentence that carries the main argument
4. Give me three questions to test whether I've understood it

Then wait for my answers to those three questions.

Why it works. Reframing a text as "the set of questions it answers" forces you to engage with its argumentative structure rather than absorb it passively. Point 2 builds critical reading — noticing what a text assumes rather than proves is the core academic skill and the one most students develop last. And because you paste the text, the model is reading your source rather than recalling something about it, which is the most reliable mode available.

Workflow 5: spaced repetition card generation

This is a legitimate efficiency win, because writing cards is mechanical while reviewing them is the learning.

From these notes, create spaced-repetition flashcards: [paste notes]

Rules:
- One fact or relationship per card — never bundle
- Question side must be answerable in under 15 seconds
- No yes/no questions
- Avoid cards answerable from wording alone; test the idea
- Include 3 cards that require applying the concept, not just recalling it
- Format: Q: ... / A: ...

Flag any note that's too vague to make a good card.

Why it works. "One fact per card" and "no yes/no" are the two rules that separate useful cards from unusable ones — bundled cards cannot be scored, and yes/no cards are guessable. The final line is a bonus diagnostic: notes too vague to become cards are usually notes you did not understand when you wrote them.

Three ways students accidentally hurt their learning

1. Summarisation as a substitute for reading. A summary transfers conclusions without the reasoning that makes them memorable or transferable. You will recognise the material on an exam and be unable to use it. Summaries are fine for triage — deciding what to read — and poor as a replacement.

2. Getting answers instead of hints. Every time you accept an answer you were close to producing, you spend the retrieval opportunity and get nothing for it. This is why every workflow above includes a "hint first" rule.

3. Trusting explanations of technical detail without checking. This is the one with real academic consequences. AI assistants fabricate confidently: dates, formulas, definitions, mechanisms, and especially citations. A fabricated reference in submitted work is an academic integrity problem regardless of intent. Verify against your course materials — they are the authority, not the model. See how to fact-check AI output.

On academic integrity

Briefly and without moralising: the workflows above — being quizzed, having your explanation critiqued, getting hints, generating your own flashcards — are study techniques, and the work remains yours. Submitting AI-generated text as your own writing is a different thing, and institutions treat it as misconduct.

The practical point is that policies vary by institution, course, and assignment. Check yours rather than assuming. And note the self-interested argument: the workflows that risk integrity violations are the same ones that produce the weakest learning, so the honest path and the effective path point the same way.

Using this with PhantomAI

These workflows depend on two things PhantomAI supports directly: pasting your own material so answers are grounded in your notes rather than the model's recall, and multi-turn conversation so a quiz can proceed one question at a time. Conversations persist, so a quiz session can be resumed. Prompts like the ones above can be saved to the prompt library rather than retyped each week.

openai/gpt-oss-120b is the better choice for critique and problem-set feedback; openai/gpt-oss-20b is faster for rapid quizzing. And as everywhere: verify technical specifics against your course materials. PhantomAI does not know your syllabus, your marking scheme, or your institution's policies. See limitations.

Continue

• How to fact-check AI output — essential before using AI for coursework
• AI for research — literature work and source handling
• How to write better prompts — the techniques these prompts use