AI for Students: Where the Line Is, and What Detectors Actually Prove

The question students ask isn’t “which AI tool should I use”. It’s “how do I use this without getting into trouble, and what happens if a detector flags my own writing”. This page answers those two, gives the working method for each of the main study jobs, and links on to the full guide for each one.

One thing to say plainly at the top: this is general information, not advice on your case. Procedures differ by institution and the consequences are real, so where something specific is happening to you, the people to talk to are at your own university.

The line, and where it’s drawn

The useful distinction is between using AI to understand something and using AI to produce something you submit. Explaining a concept back to you, arguing against your thesis, generating practice questions, checking whether an argument holds: that’s a tutor doing tutor things. Handing in text you didn’t write isn’t.

What matters more than any general principle is that the rule binding you is your own institution’s written policy, and those differ enormously. Some departments ban AI outright. Some require a declaration. Some permit it for planning and outlining but not drafting, and a few now build it into the assignment. Find the actual document and read the actual sentences. If the wording is vague, ask your instructor in writing and keep the reply.

What detectors do and don’t prove

This is the part that causes the most anxiety, and it deserves specifics.

Start with how hard the problem is. OpenAI built a classifier to detect AI-written text and then withdrew it. On its own evaluation, run on what OpenAI calls a “challenge set” of English texts, the tool correctly identified 26 percent of AI-written text while incorrectly labeling human-written text as AI-written 9 percent of the time. A notice now sits above the original announcement: “As of July 20, 2023, the AI classifier is no longer available due to its low rate of accuracy.” A challenge set is a deliberately hard sample, so 26 percent isn’t a general accuracy rate. What it does show is that the company with the deepest access to how these models write could not make detection good enough to keep offering it.

Turnitin, whose detector most students will actually meet, does better, and has been unusually open about where it doesn’t. Three of its published statements are worth knowing.

Turnitin restated the under-one-percent figure in September 2024 and added the framing it wants educators to use: the AI writing report “is not an absolute proof or disproof of AI writing,” and detection is “a signaling tool and one piece of an investigative puzzle”. In its 2023 guidance the instruction was blunter still, telling educators to “use the information to initiate a conversation, not to draw a conclusion”.

There’s one more finding that belongs here, and it matters most to the students least likely to hear about it. In a study of several widely used GPT detectors, Liang and colleagues found that the detectors “consistently misclassify non-native English writing samples as AI-generated, whereas native writing samples are accurately identified”, and concluded by cautioning against using such tools in educational settings. If you write English as a second language, your baseline risk of being wrongly flagged is not the same as your classmate’s, and that’s a documented finding you can point to.

If you’re the one flagged

A score is a signal that starts a process, and the case is rarely built on the score alone. Institutions look at drafts, timestamps, submission metadata and, most of all, how you talk about your own work.

Some things are worth doing regardless of what happened. Ask for the allegation and the evidence in writing. Find out what the procedure is and what the deadlines are, because they’re usually short. Ask your students’ union or student advice service what support you’re entitled to, including whether you can bring someone with you. Those services exist for exactly this and are used far less than they should be.

Writing somewhere that keeps version history helps, and costs nothing: a document with weeks of genuine edit history is useful corroboration. It’s corroboration and not proof, and it cuts both ways, since a history showing one large paste says something too.

And if you did use AI in a way the policy didn’t allow, most integrity offices will tell you the same thing: an early, honest account tends to end better than a denial that doesn’t hold. That’s a decision only you can make, with someone who knows your institution’s process.

The four jobs worth using it for

Each of these has a full guide behind it. Here’s the version that fits in a paragraph.

Learning something you find hard

The pattern that works isn’t “explain X to me”. It’s asking for an explanation, then explaining it back in your own words and asking where you got it wrong. That second step is the technique. It turns a passive read into a test, and models are good at spotting the specific place an understanding breaks. Full guide: using AI as a tutor.

Turning a syllabus into a study plan

Most students plan badly because they estimate optimistically and never schedule review. Give a model the topic list, the exam date, and an honest number of hours per week, and ask it to work backward with spaced repetition built in. Then push back on the first version, which will be too ambitious. Full guide: building a study plan that survives contact with reality.

Notes and recordings

Automatic lecture summaries are genuinely useful and genuinely lossy. They compress well, and they drop the thing your lecturer said once, in passing, that turns out to be the exam question. Use them as a second pass over your own notes. Full guide: what the note-taking apps do well and badly.

Citing it properly

If your institution permits AI use, it almost certainly requires you to say so, and the style guides disagree with each other about how. MLA advises against treating the tool as an author; Chicago does the opposite and puts ChatGPT in the author position. Full guide: how to cite AI in APA, MLA and Chicago.

The failure mode that catches people out

Invented sources. Ask for references and you can get citations that look perfect: plausible authors, a real journal, a page range, sometimes a working DOI. The formatting is right because formatting is a pattern.

The scale has been measured. Walters and Wilder, writing in Scientific Reports in 2023, found that 55 percent of the citations produced by GPT-3.5 and 18 percent of those from GPT-4 were fabricated. The part most people miss is what they found in the citations that were real: 43 percent of the genuine GPT-3.5 citations and 24 percent of the genuine GPT-4 ones still contained substantive errors, most often in volume, issue, page numbers or year.

That second number is why the obvious check isn’t enough. “Does the link open” passes a citation whose DOI resolves to a real paper that isn’t the one you were given. What you have to do is open each reference and match the author, the title, the year and the journal against what appears. If any of the four is off, you don’t have that source.

Those figures come from models working without retrieval. Modes that search the web and return links you can click are a different situation, and a better one, though a link that exists still isn’t a link that says what the sentence claims. The check stays the same either way.