AI Detectors for Teachers: What a Score Can and Can't Prove (2026)

According to Turnitin's own help guide, its AI writing percentage "should not be used as the sole basis for adverse actions against a student." Read this again before you open up your next report. The company that makes this tool is telling you not to treat it as proof of anything.

That sentence is the most useful thing in the whole category and most teacher guides to AI detectors skip it. They usually give ratings to the tools. I read the documentation from the vendors, the Stanford bias study and the guidelines from three different universities instead. The answer is the same everywhere: AI detectors can tell you where to look but that's it.

This guide covers what a score can and cannot support, the numbers worth knowing before you rely on one, and a way to handle a flagged essay that holds up if the student pushes back.

What a detector score actually tells you

An AI detector is a classifier. It reads text and returns a number based on patterns its training data associated with machine writing. It cannot see who typed the words or how the draft came together. Turnitin's guide adds one related point that surprises people: the AI percentage is "different from and independent of the similarity score", so a paper can show no matching passages and still carry an AI flag, or the reverse. If you want the mechanics, how Turnitin's detector works goes through them.

Turnitin also marks its own weak spot. Scores from 0 to 19 percent get an asterisk and the guide states: "there is a higher incidence of false positives when the percentage is between 0 and 19." In a post from the chief product officer, Turnitin stated that the document-level false positive rate is less than 1 percent for documents with more than 20 percent AI writing and the sentence-level false positive rate is about 4 percent. The same post says 54 percent of falsely flagged sentences sit right next to genuinely AI-written ones.

So, a low score with an asterisk is the weakest form of evidence that can be obtained from such a tool. I would not act on it at all.

The false positive math nobody does for you

Vanderbilt did the arithmetic in August 2023, when it switched off Turnitin's detector. Turnitin claimed a 1 percent false positive rate. Vanderbilt had submitted 75,000 papers in 2022. One percent of 75,000 is 750, and the university's post says around 750 papers could have been incorrectly labeled as partly AI-written. Its conclusion was blunt: "we do not believe that AI detection software is an effective tool that should be used."

You do not have to agree with Vanderbilt to take the lesson. A small error rate applied to a large pile of papers still produces real people who did nothing wrong. In a single class of 30 the risk looks tiny. Across a department that runs every submission through the same tool, someone gets flagged unfairly sooner or later, and the false-positive problem falls hardest on a predictable group of students.

Who gets flagged more often

The best-known evidence comes from a Stanford-led study of seven GPT detectors. The researchers ran 91 TOEFL essays, written by non-native English speakers, through all seven. The detectors misclassified more than half of them as AI-generated, an average false positive rate of 61.22 percent. All seven agreed that 18 of the 91 essays, 19.78 percent, were AI-written. On essays by US eighth graders, the same detectors were near-perfect.

The authors then asked ChatGPT to enhance the word choices of the TOEFL essays to sound more like a native speaker. The average false positive rate dropped from 61.22 percent to 11.77 percent. That shows what the detectors were reacting to: plain vocabulary and simple sentence construction, which is how many careful second-language writers write. The study tested the detectors available in 2023, so the exact figures will have moved. Vanderbilt's guidance still repeats the concern, and if you teach multilingual students I would give a flag on their work even less weight. For a longer look at the pattern, see why an AI detector says your essay is AI when it isn't.

Why even the makers distrust the output

OpenAI built its own classifier and published the numbers. It correctly flagged 26 percent of AI-written text and wrongly flagged 9 percent of human-written text. The page said the tool should not be used "as a primary decision-making tool" and was "very unreliable on short texts (below 1,000 characters)." As of July 20, 2023, OpenAI said the classifier was "no longer available due to its low rate of accuracy."

The University of Rochester's guidance for instructors goes further. It says the university "does not endorse or recommend any AI Detector software" and that the "probative value, in general, is quite low." It also admits that "AI-humanizers can outwit the detectors", which means a clean score proves nothing either. And it names a bias problem no accuracy figure captures: a positive hit "may sway you to look for and identify things as AI use."

Rochester also adds a warning: do not submit personally identifiable information to detectors. That could be a FERPA violation. Check your school's rules before pasting an essay into a third-party tool.

What each source says

How to handle a flagged essay

Arizona State's toolkit says "a conversation (not a confrontation) is the best first step" and tells instructors to start "from curiosity, not accusation." Rochester, Vanderbilt and Turnitin all land in the same place, and the steps below combine what they describe.

1. Write down what made you look. Rochester suggests noting things like fabricated or unverifiable sources, odd revision history or submission timing, and big gaps in skill between assignments, and it says none of these is conclusive on its own.
2. Compare the essay with the student's earlier work for style, tone and level, as Vanderbilt suggests, and check the sources and facts it cites.
3. Meet one-on-one and start with process questions. Rochester's example is "Walk me through how you got to the conclusion of your paper."
4. Let the student explain before you raise misconduct, and stay non-adversarial. Some students will tell you what they did when they are asked well.
5. Document what you saw in enough detail that a board of non-experts could follow it, then use your institution's process.

Notice where the detector score appears in those guides. It is not step one. At most it is a reason to start looking.

If you still use a detector

Many schools license one but you may not get a vote. There are a few habits to keep the damage down. The whole document should be run through the system instead of just a paragraph, as OpenAI found that short texts were unreliable. Ignore the asterisked scores. Turnitin asks educators to "view the highlighted areas in the document in totality." Treat the report as a starting point for a dialogue with the student, as Turnitin does itself.

Remember the other direction as well. Humanizers exist, including our own, and they rewrite wording without proving anything about who wrote the original. A pass is no more proof than a flag is.

Designing the problem away

Vanderbilt's alternatives are worth copying whatever tool you use. Talk with students early about what AI use is allowed in each course. If you permit it, ask for disclosure and citation. Redesign assignments with in-class writing, questions about class discussions, or current issues that AI tools are not trained on. Rochester adds that you should define what you mean by generative AI in your own class rather than simply saying it is allowed or banned.

The short version

A detector score is a probability from a classifier, not a finding of fact. The vendor says not to use it alone, universities that studied it disagree about whether to use it at all, and the Stanford study shows it can be wrong about the same students again and again. Use it, if you must, as a reason to look closer, then rely on what you can document: the student's earlier work, the sources, the drafting history, and a calm conversation. If you want the longer treatment, these cover the ground:

• Are AI detectors accurate? Every published number, in one table
• What AI detector do colleges actually use?
• How AI detectors work: perplexity, classifiers and watermarks
• Does Turnitin detect ChatGPT, Claude and Gemini?