How Do Professors Check for AI in College

Professors check for AI two ways, and only one of them is software. The automated half is a detector, almost always Turnitin, that scores a submission and shows the result to the instructor and nobody else. The manual half is a person noticing that a paper does not sound like your previous work, cites sources that do not exist, or cannot be explained when asked about it. In practice the second half is what starts most conversations, because a percentage on its own is not treated as proof.

The reason this question is hard to answer well is that almost every page ranking for it quotes the same statistic, and the statistic is about a different group of educators than the one you are asking about.

Do colleges check for AI?

Some do and some do not, and the honest answer is that no reliable national figure exists for United States higher education. The number you will meet on nearly every page is that educator use of AI detection tools jumped from 38 percent to 68 percent. That is real, and it comes from the Center for Democracy and Technology, which found that "68% of teachers report using an AI content detection tool regularly". CDT also states, in the same release, that its research "was developed based on nationally representative surveys of K-12 public school teachers".

K-12. High school teachers. Not professors, not colleges, not four-year institutions. Several pages currently ranking for college search terms present that figure as the share of universities using detection, and one of the most-cited restatements of it attributes it to "higher education institutions" outright. It is a survey of public school teachers, and the same report found that 64 percent of those teachers said a student at their school had faced consequences over generative AI, a 16 percentage point rise in one year.

What can be sourced at university scale is volume rather than adoption. Turnitin, reporting on the first year of its AI writing detector, said it had reviewed over 200 million papers, of which "over 22 million (approximately 11 percent)" contained at least 20 percent AI writing and "over six million (approximately three percent)" contained at least 80 percent. Alongside it Turnitin cited Tyton Partners research finding that nearly half of surveyed students used generative AI monthly, weekly or daily, and that 75 percent said they would keep using it even if faculty or their institution banned it.

So detection is widespread enough to be worth assuming. But whether it runs on your specific assignment is a local decision, made in four steps by four different people, and none of them is visible to you.

What AI detector do schools use?

There is no single answer, because the school picks, not the platform. Turnitin has the deepest footprint in United States higher education and the widest documentation for the learning management systems universities run, which is why it accounts for most of the scores students ever hear about. Others appear alongside it, usually chosen by individual instructors rather than institutionally.

Two things follow from that table. First, if your instructor pastes your essay into a free web detector, they are using a different model than your institution's, with a different error profile, on text that has been stripped of formatting. Second, none of these numbers are comparable, because each vendor measures on its own private test set. We traced every published accuracy figure in this niche back to its source, and how many of them have no source at all, in the AI detector false positive rate.

How do professors detect AI in papers without a detector?

This is the half that actually gets students called in, and it needs no software.

Your writing changed. An instructor who has read three of your assignments has a rough model of how you write. A sudden shift in vocabulary, sentence length or confidence is noticeable in a way a percentage is not. This is why the pattern of using AI on one paper out of five is riskier than it feels.

The citations do not exist. Language models invent plausible references, and a fake source is verifiable in a way that a style judgment never is. Checking one suspicious citation takes thirty seconds and produces something concrete. This is the single most common way a suspicion becomes a case.

The essay is fluent and empty. AI drafts are strong on structure and weak on specifics. Assignments come back with no engagement with the seminar reading, no reference to anything said in class, and no argument that could only have been made by someone who was there.

You cannot talk about it. Several institutions now resolve suspicions with a short conversation about the paper rather than a disciplinary process. It is faster, it is fairer, and a student who wrote their own work can do it easily.

The assignment was designed to make AI useless. The quieter response across a lot of departments has been to change what gets assigned: in-class writing, oral components, work that builds on a specific lecture or lab. An instructor who can turn the lecture recording itself into a set of questions does not need to guess whether a take-home essay was drafted by a model, because the graded moment happens in the room.

Can professors tell if you used ChatGPT?

Not with certainty from the text alone, and the tools do not claim they can. Turnitin says its score "does not make a determination of misconduct" and that the percentage "should not be used as the sole basis for action or a definitive grading measure by instructors". It also concedes that its reported figure is a floor rather than an estimate: "if we identify that 50% of a document is likely written by an AI tool, it could contain as much as 65% AI writing."

There is also no watermark to find. OpenAI's provenance work covers images and audio, and its own update dated July 31, 2026 describes the expansion as reaching audio, with text absent from the page entirely. The Wall Street Journal reported in 2024 that OpenAI had built a text watermarking method and shelved it. Nothing in a ChatGPT output identifies itself.

What a professor can tell is that a piece of writing has the statistical shape of model output. That is a real signal and it is often right. It is also wrong often enough that it should never be the whole case, which is roughly the position Turnitin itself takes. Liang and colleagues, publishing in Patterns in 2023, ran seven detectors over TOEFL essays written by humans and found an average misclassification rate of 61.3 percent, so the students most likely to be wrongly accused are the ones writing in a second language.

What happens if a professor thinks you used AI

Usually a conversation before anything formal. The score itself is invisible to you throughout: Turnitin states that "the AI writing detection indicator and report are not visible to students", though instructors can download the report as a PDF and share it if they choose. So the first you hear of a number is generally when someone tells you about it.

The evidence that helps you at that point is not an argument about detector accuracy. It is a record of how the document was written: version history showing the essay developing over days, drafts, notes, the outline you started from. Detectors infer from finished text. A version history records the process. If you write in something that saves revisions and avoid pasting a completed draft into an empty file, you have that record for free. We walk through exactly what to keep and what it is worth in how to prove you did not use AI.

Where a rewriting tool fits, and where it does not

If you drafted with AI and the prose came out flat and uniform, that uniformity is both what a detector responds to and what makes the paper worse to read. Rebuilding sentence variance and register fixes both at once. That is what an AI humanizer for students does, and it is worth being precise that this is a style operation on writing you have already thought through.

It does not help with the manual half of this page. A rewritten essay still cites the sources it cited, still fails to engage with the seminar reading, and still cannot be discussed in a meeting if you did not do the thinking. And it cannot undo a false flag on work that was yours to begin with, because the remedy there is evidence about process, not more editing. Nobody can honestly promise you a score against Turnitin either: it sells only to institutions, publishes no public checker, and deliberately refuses to name the rewriting tools it has trained against, on the grounds that a list "would make it easier for students to evade our system". Anyone quoting you a guaranteed percentage is guessing. The full picture of what does and does not exist is in the Turnitin AI checker.

The short version

Professors check for AI with a detector their institution licensed, and with their own reading of your work. The 68 percent figure everyone cites is K-12 teachers, not colleges. Turnitin runs on most of the higher education submissions that get scored at all, but only where the school licensed AI detection, an administrator left it on, and the instructor enabled it on that assignment, and you cannot see any of those. The detector output is not treated as proof by the company that sells it. Fake citations and a paper you cannot talk about are what turn a suspicion into a case, and a drafting history is the only thing that reliably answers one.