Where the line is with AI and coursework.
A three-column split of what's fine, what needs asking about, and what isn't. Plus two questions that settle the cases the columns don't cover. Free to copy, print, adapt, or paste into a syllabus.
The three columns
Your institution's own policy always wins. This is for the cases it doesn't cover, which is most of them.
- Making practice questions from your own material
- Having a concept explained to you
- Quizzing yourself and building flashcards
- Critiquing a draft you wrote
- Organising notes you took
- Getting deadlines onto a calendar
- AI-assisted outlining
- Grammar and clarity passes
- Brainstorming thesis options
- Debugging code in CS courses
- Translating your own writing
- Transcribing or summarising a lecture
- Turning in AI text as your own
- Having it write the paper
- Answering graded work with it
- Citations you haven't opened and checked
- Anything you'd hide from your instructor
The two questions
Most cases don't fit a column cleanly. These settle them, and a student can apply both in about five seconds.
- The disclosure test. Could you tell your instructor exactly what you did, in detail, without being asked?
- The deletion test. If every AI tool vanished tomorrow, would you still know this material?
Fail either one and you didn't study. You outsourced.
Read the syllabus first, then keep your receipts
This is not a page about which AI use is okay. Nobody can tell you that in general, and I won't. It's about where to look, and what to hold on to. Your course's written rule decides. Then you need proof the work is yours.
What binds
- Your course syllabus and your school's academic-integrity code are the rule, and they differ course to course. Penalties run from a zero to a failed course, suspension, expulsion or a note on your transcript. S
- AP has its own rules: Art and Design bans generative AI for the final artwork, Capstone allows optional AI aids with the analysis yours, skipped checkpoints score 0, a teammate's violation can zero the group, and portfolios go through a similarity checker. P
- Test day has device rules. The SAT bans smart glasses; the ACT bans recording or connected devices. S
- If you're in a public school in Tennessee or Ohio, your district has to adopt a written AI policy, so one exists for you to read. Ohio P, Tennessee S, and whether it reaches colleges is open.
- Applications and professional exams have their own rules. The AAMC lets applicants use AI to brainstorm and edit if the work is theirs. Berkeley Law bans AI drafting and AI on exams. S
- Every tool has an age gate. Claude is 18 and over. ChatGPT is 13 and over with a parent's permission under 18. Copilot and Gemini are 13 and over, with regional differences. Claude and Copilot P, ChatGPT and Gemini S.
- Federal law has no student-facing AI rule. FERPA, which governs school records, was not read for this page, so I make no claim about it. S
Three things you can do
Where your course says yes, and with disclosure when it asks. S
- Have a concept explained, make practice quizzes, or build a study guide from your own notes.
- Run a grammar or clarity pass on writing you did, or brainstorm, if the syllabus allows it.
- Use AI for an accessibility need your school's disability office has set up with you.
If the syllabus is silent, ask your instructor in writing. Keep the reply.
The pattern that gets students sanctioned
Using AI past what the course says, with nothing of your own to show for it. Yang v. University of Minnesota is the one that lands hardest: the school's process was upheld, because the evidence went well beyond a detector score and the exam instructions banned AI. S
Now the other side. Newby v. Adelphi: a court called the cheating finding without valid basis and ordered it removed, where the school had a detector score and little else. S Detector-only accusations lose. Cases where the school held more than a detector went against the student. So you don't need luck. You need a record.
That means version history on, drafts saved, outlines and notes kept, sources opened. Detector scores are unreliable, worst for people writing in a second language, and a flag is not proof. If you're ever accused, ask in writing for the evidence and the policy being cited. None of this is a way to get past a detector. It's a way to show you did the work.
Use AI only as far as your course's written rule lets you, keep proof of your own work, and never paste anyone else's personal information.
Never paste these
- A student ID number, a school email, or a classmate's full name with a grade or a discipline record.
- A question or prompt from a proctored or open-note exam.
- A minor's home address or phone number.
- Anyone's face into an image generator. AI images of classmates bring school discipline and possible criminal exposure.
And when a course asks you to disclose, say three things: the tool, what you used it for, and what you changed.
Two-minute exercise
Take one thing you did with AI this week.
- Find the line in your syllabus or assignment sheet that covers it. If there is none, write the instructor two sentences asking.
- List the proof you hold for that piece: version history, outline, notes, sources. Is anything missing?
- Did you paste in a classmate's name with a grade, a student ID, or an exam question? If yes, take it out next time.
If you can answer all three in two minutes, you have a record to point to. If you can't, that's your finding.
A receipt you can keep
When a course asks you to disclose, you say the tool, what you used it for, and what you changed. This builds those lines for you in your browser, with the self-checks you have done on this site. Nothing is stored or sent. Fill in the blanks truthfully; it only writes what you type.
Draft copy, not legal advice. P primary source read, S secondary report, M not confirmed. Dates, prices and case details live on the current notes, checked on a stated day. Reviewed by: nobody yet.
Audit the answer before you use it
For a student, invented means a source. A journal article with a real-sounding title, two authors, a year and page numbers, and no paper behind it. A quote from a book that never says it. A statistic with a link that goes nowhere. It slips into a draft at 1am because it fits the sentence you were stuck on, and you were tired. The school doesn't care that you didn't mean it. A source you haven't opened is a source you can't defend. Open it, find the line, or cut it. A fake source is the thing a grader can verify in thirty seconds.
There are two kinds of wrong, and they need different fixes. The first is wrong about the world: it states something false, or something that changed after its training stopped. That stopping point is its knowledge cutoff, so recent rules, prices and papers are the riskiest. The fix is to hand it the real source. The second is wrong about what you gave it: you pasted your notes and it quotes a line that isn't there, or says Tuesday when your page says Thursday. The fix is to make it quote the exact line back, then check that line against your page. Before you blame the AI for a bad answer, work out which kind you got.
So before any AI answer goes anywhere, run it through this. Same six-part skeleton as the prompts page.
[PASTE: the AI answer I'm about to rely on, plus the sources it cites, if any] ——— ROLE: A skeptical TA checking an answer before it goes anywhere. CONTEXT: [course + level]. This answer came from an AI. I haven't checked it yet. TASK: Go through it claim by claim. For each one, say whether I can check it from my own course materials, from a source I'd have to open, or not at all. CONSTRAINTS: Don't fix the answer. Don't add new facts. Every citation, quote, statistic, case name and date stays flagged until I've opened it myself. OUTPUT: A list, one line per claim: claim | how I check it | what happens if it's wrong. Then the one claim most likely to be invented, and why. INTEGRITY: Mark anything you can't verify [UNVERIFIED]. Never fill a gap from general knowledge.
Then open the flagged ones yourself. The prompt finds the weak spots; it can't open the library for you.
One good run proves nothing. Ask the same question twice and you can get two different answers, because the AI builds each answer fresh instead of looking it up. So the audit is not a check you do once and then skip. Run it on every answer that matters, and if two runs disagree on a fact, that fact goes to the top of your list to check by hand.
The rule underneath all of it
Every column above comes from one line: AI generates the questions. You generate the answers.
That isn't a moral posture, it's what the evidence supports. Retrieving something from your own memory is the act that installs it. Outsource the retrieval and you outsource the learning — and you don't find out until the exam, which is why researchers named the effect cognitive debt.
The nuance matters too. Offloading work to a machine isn't automatically harmful; Lodge and Loble argued in 2026 that what matters is what you do with the freed capacity. Spend it on something harder and you come out ahead. Spend it on nothing and you've bought debt.
For a syllabus
Paste this in and edit it to fit. No attribution needed.
AI tools are permitted in this course for learning, and not for producing work you submit. You may use them to generate practice questions, explain concepts, quiz yourself, build flashcards, organise your own notes, and critique a draft you wrote yourself. You may not use them to write, rewrite or draft any part of an assignment, to answer graded work, or to supply citations you have not opened and verified. If you are unsure, apply two tests. Could you describe exactly what you did, unprompted, to me? And if every AI tool disappeared tomorrow, would you still know this material? If the answer to either is no, you have crossed the line. Ask me before using AI in any way not listed here. Asking is never penalised.
Where this comes from
Kosmyna et al., MIT Media Lab (2025). 54 people wrote essays under EEG across three sessions. The group using an AI assistant showed the weakest neural engagement, underperformed at neural, linguistic and behavioural levels, and struggled to quote essays they had finished minutes earlier. The authors named the effect cognitive debt. Published as a preprint and not peer reviewed at release.
Gerlich, Societies (2025). A survey of 666 people found frequent AI use correlated negatively with critical thinking, with cognitive offloading as the mediating mechanism. The youngest participants scored worst.
Dunlosky et al., Psychological Science in the Public Interest (2013). Graded ten study techniques. Practice testing and distributed practice rated high utility. Highlighting, rereading and summarising rated low.
The cases named above (Yang, Newby, and the others on the current notes) are reported from secondary sources, court filings not opened. Each carries its date and source on the current notes.
Lodge and Loble (2026). Argued that cognitive offloading is not inherently harmful and that the outcome depends on what the learner does with the freed capacity.