AI for students who don't want to cheat

It’s ten at night and you have a problem set due at midnight. You have done four of the six problems honestly. The fifth is a wall. You open a chat tab. The cursor blinks. You could paste the question and have a clean answer in eight seconds. You could also paste the question and ask it to help you see what you’re missing — and learn the thing the assignment was supposed to teach you in the first place.

The two prompts are six words apart. The first one is what almost every article about “students and AI” has been panicking about for three years. The second one is the most useful study companion any student has ever had access to, on any night, for the price of a sandwich a month. The difference between them is the entire story.

The line professors actually care about

The discourse around students and AI tends to start in the wrong place. It opens with a stat — a Digital Education Council survey put global student AI use at 86% in 2024 — and then asks whether this is the end of education. It’s a bad question. The honest version is much smaller, and every professor I’ve heard articulate it lands in roughly the same place.

If a professor stopped you in the hallway, pointed at your submitted essay, and asked “did you write this?” and the honest answer was “no, the AI did” — that’s the line. Everything on the other side of it is fine. Studying with AI? Fine. Asking it to explain a confusing paragraph in a textbook? Fine. Having it quiz you on a chapter before an exam? Fine. Asking it to look at your draft and tell you the argument is weak in three places? Fine. Pasting an essay prompt and submitting what it writes back? Not fine.

Notice what that line is and isn’t. It’s not a ban. It isn’t even particularly hard to comply with. It’s a test of whose thinking is in the submitted work. Harvard’s Graduate School of Education says it pretty cleanly: at its best, generative AI “can be like a tutor or thought partner with unlimited time to help you learn — but it should not be used to do the cognitive work for you, or else your own learning will be greatly diminished.” Oxford, Princeton, Cambridge and most of the schools on the 2025 policy survey have converged on roughly the same rule, with transparency added: use AI, say so, don’t pretend its work is yours.

That’s a line you can stay on the right side of and still use AI hard. In fact, that’s the only honest way to talk to a student who knows AI is genuinely useful, because it is. The pretense that the only options are “don’t use it” and “cheat with it” isn’t going to survive contact with anyone who has actually opened the app.

Tutor mode vs. ghostwriter mode

The most useful frame I’ve found is to picture two different people sitting on the other side of the chat. One is a patient tutor who has unlimited time, has read every textbook, and gets paid by the hour to make you understand the material. The other is a stranger on Fiverr who will write your paper for fifty dollars. They live in the same chat box. The prompts you use decide which one shows up.

The split is rhetorical, not technical. The model on the other end is the same model. The exam you have to take next week is the same exam. What changes is what you take away from the interaction — a sturdier understanding of permutations and combinations, or a submitted PDF you don’t understand any better than you did before you opened the chat.

This isn’t a new pedagogy. It’s the Socratic method that every great teacher you remember already used, automated and put on call 24/7. Research on AI tutors trained to ask questions instead of give answers — the SocratiQ work out of Harvard and Google DeepMind’s LearnLM, which in their public technical report outperformed comparable models on every learning-science axis they measured — suggests the tutor mode is not just a vibes-based improvement. It works. The catch is that the user has to ask for it. The default prompt — “solve this” — gets you the ghostwriter.

Five prompts that teach you

Here are the five prompts I’d hand a student who actually wants to learn the material. Memorize them. Adapt them. They work on every major chat model in 2026, and they work on the small ones running offline on a laptop too.

Two things to notice. First, every prompt either constrains the model’s output (don’t solve, don’t rewrite, ask one at a time) or asks it to evaluate yourthinking rather than replace it. Second, you’ll feel the difference immediately. The tutor-mode session takes longer. It’s harder. You’ll be a little frustrated. That’s what learning feels like. The version where you paste “answer this” and copy the result feels great in the moment and leaves nothing behind.

Habits that pair with AI well

Tutor-mode prompts on their own aren’t a study plan. They’re a tool. The habits underneath them are what actually move the needle. A few that compound:

• Try the problem first, alone, for ten minutes before asking for any hint. The struggle is where the learning happens. The AI is the hint-giver, not the substitute for the struggle.
• Active recall before any chat session. Close the textbook, write down what you remember, then ask AI to grade what you got wrong. The mistakes you make from memory are way more informative than the ones you make while reading the answer key.
• Re-derive the explanation without looking.After AI walks you through a proof, close the chat, open a blank page, and try to reproduce the reasoning. If you can’t, you didn’t understand it — you just watched it happen.
• Keep a “things I got wrong” log. A text file, a notes app, anything. At the end of the week, paste it back and ask the AI to make a 10-question quiz out of it. Spaced repetition without an app.
• Time yourself before asking. Ten minutes on a problem before any prompt. Five minutes on a draft paragraph before asking for feedback. The clock is the anti-shortcut.

None of those habits require AI. They were good study habits in 1995. What AI adds is a tireless, infinitely patient, infinitely available partner for the parts where you used to have to flag down a TA, find a study group, or wait for office hours. Bloom’s 1984 paper famously found that one-to-one tutoring produced learning gains of two standard deviations over classroom instruction — the average tutored student outperformed 98% of their classmates. The 2 sigma problem was scalability: tutors are expensive and there aren’t enough of them. AI is the first serious answer the field has had in forty years. It would be a strange thing for a student to opt out of.

Where AI is a cheat code, and it’s fine

There’s a whole category of AI use that isn’t in the gray zone at all. It’s where AI does something that a friend, a sibling, an upperclassman, or a tutor would have done for you ten years ago, except now it’s 11pm and they’re not awake.

• Decoding dense textbook passages.“Explain this paragraph from a 1973 econ textbook like I’m a smart undergrad seeing it for the first time.” The textbook hasn’t changed; your access to a patient explainer has.
• Generating extra practice problems. Most textbooks ship with ten problems per chapter. You want forty. AI makes them all afternoon.
• Explaining a professor’s bad slides.Paste, ask, get a version with examples. You’re not getting around the professor; you’re patching their explanation.
• Debugging code you wrote yourself.The kind of bug that a TA would point at in five seconds, in the small hours of Sunday night when there’s no TA.
• Translating math notation across textbooks.One uses Σ with subscripts, another uses set-builder notation, another uses words. AI translates between them.
• Mock interviews for oral exams or thesis defenses. “I just read this paper. Ask me five questions a committee member would ask.” Free office hours that never close.

None of these submits AI’s thinking as yours. All of them make your own thinking better-equipped. They’re what your friend who always did slightly better than you did in undergrad does on a Wednesday night. That friend used to be rare. Now everyone has one.

Where the line gets blurry

It would be dishonest to pretend the line is always crisp. Some uses live in the gray zone and depend on the class, the professor, and what the assignment is actually trying to measure. Worth naming them out loud rather than pretending they aren’t there.

The honest move in every blurry case is to ask. A two-line email to a professor asking “is X okay for this class?” before you’ve submitted anything turns a gray area into a green light or a clear no. Most professors are visibly relieved when a student asks; they’ve been trying to write a policy for the syllabus all semester and you’re saving them from having to enforce one retroactively.

What professors actually want

The other thing worth saying out loud: most professors are not in a moral panic. The ones writing op-eds about “the death of the essay” are loud; the ones quietly figuring out how to teach with AI in the room are far more numerous. The Harvard provost’s guidance, the UT Austin sample policies, and the menu the Duke Center for Teaching and Learning sends out to faculty all read the same way once you cut through the legalese: tell the students what’s allowed, expect them to disclose it, and grade the thinking, not the polish.

What that translates to, in practice, is a small set of things you can do that essentially never go wrong:

• Read the syllabus AI section first.Most courses now have one. Some are permissive (“use it however helps, just cite it”), some are restrictive (“no AI on assignments, period”), and a few are course-specific (“allowed for studying, banned for the final paper”). Find your class’s rule before you write anything.
• When in doubt, disclose.A one-line acknowledgment at the bottom of the paper — “I used a chat model to quiz me on the readings and to suggest two counterexamples in section 3; the writing is mine.” — turns a potential integrity question into a non-issue. Professors I’ve talked to are universally fine with disclosed use and universally frustrated by suspected-undisclosed use.
• Don’t paste confidential class data into a public chat. Many universities now flag this. Harvard, for example, runs a private ChatGPT Edu pilot specifically because student work and unpublished data shouldn’t go to a vendor’s general training corpus. Tools that run locally on your own laptop sidestep the question entirely — we’ve written about what BYOK and local actually buy you and what they don’t.
• Don’t use AI detectors as moral cover.Both ways. If your professor is using one, know that they false-positive on legitimate student writing surprisingly often. And don’t use the existence of a detector as a reason to feel okay about submitting AI work, because what matters isn’t whether you got caught — it’s whether the work represents your thinking.

The one-line test

If you skip the rest of this post, take the test.

Run it on whatever you’re about to do with the chat box open. “Quiz me on chapter 4” — you’ll be slightly better at the material in twenty minutes. Green light. “Write the introduction to my term paper” — you won’t. Red light. “Give me feedback on my draft introduction” — you’ll probably be a slightly sharper writer. Green light. “Rewrite my draft introduction” — you won’t be a sharper writer; you’ll have a sharper paragraph. Yellow. Ask the professor.

AI is the best private tutor a generation of students has ever had access to. Bloom’s 2 sigma problem — the four-decade-old observation that one-on-one tutoring crushed every other intervention but couldn’t be scaled — finally has a plausible answer, and it’s sitting on your laptop. The version that solves it is the version that asks you questions instead of giving you answers, that grades your reasoning instead of replacing it, that’s patient enough to walk you through the same proof three different ways until the third one clicks. That tutor is already there. The only thing left is to ask for it.