What AI detectors do schools and colleges use?
Most AI detection in education runs through Turnitin: the work is already there for similarity checking, and the AI writing indicator sits in the same report. Then come standalone checkers opened in a browser, GPTZero and Copyleaks among them. The platform detects nothing itself: Canvas, Moodle and Blackboard pass the submission to whichever tool the institution licensed and show the number that comes back. Some universities have switched AI detection off, including Yale and the University of Pittsburgh. Whichever tool is running, the output is a likelihood, not a finding: it can say a passage looks machine written, not who wrote it.
The detectors that turn up most
Turnitin's AI writing indicator needs at least 300 words of prose and reports the share of qualifying sentences it thinks were machine generated. Scores from 1% to 19% appear as an asterisk with no number, which Turnitin says is deliberate: false positives cluster at the low end. Its own guidance calls the tool data, not a verdict, and leaves misconduct decisions to the instructor.
GPTZero is free to try and takes pasted text. It lists integrations for Canvas, Google Classroom, Google Docs and Chrome and advertises 99% accuracy from its own testing, while its home page says results "should not be used to punish or as the final verdict."
Copyleaks turns up where detection is wired into the platform, not opened in a tab. It publishes integrations for Canvas, Moodle, Blackboard, D2L, Schoology and Sakai, advertises over 99% accuracy on its detector page and a 0.026% false positive rate at default sensitivity in testing it publishes itself.
Every one of those figures comes from the company selling the tool.
Does Canvas have an AI detector?
No. Canvas is a learning management system, and Instructure does not ship AI detection in it. Canvas provides a plagiarism platform, and Instructure's own developer documentation is explicit that the outside tool processes and scores: the assignment goes out over LTI, the tool sends an originality report back, and Canvas displays the score next to the submission. So the AI detector Canvas uses is whichever one your institution licensed, most often Turnitin, whose documentation notes that no AI score sits in the LMS interface: somebody has to open the report on Turnitin's side.
A course with no tool connected is checking nothing, whatever the submission page looks like. AI writing detection is licensed separately, so not every Turnitin similarity customer has it, and an administrator can switch it on or off for a whole institution. That is why two students can submit identical work and only one is scored.
What a teacher actually sees
A percentage and some highlighted sentences. The report puts likely AI text in cyan, and text it thinks went through a paraphraser or bypasser gets the same highlight. The percentage is not the share of the document, which is the most misread thing on the screen. Students see none of it: the indicator and report go to instructors and administrators, though an instructor can download and share them.
How reliable are AI detectors in education?
Not reliable enough to decide a case on their own. A 2023 study of fourteen detection tools found them neither accurate nor reliable, and another found detectors consistently misclassified writing by non-native English speakers as AI generated. A 2026 paper in English Teaching: Practice & Critique argues they are ineffective, unreliable and harmful in the classroom. MIT Sloan tells instructors not to rely on them, and notes that OpenAI withdrew its own classifier for poor accuracy. Every vendor publishes a false positive rate, and not one is zero.
A rate only sounds small until you multiply it. When Vanderbilt disabled Turnitin's AI detector in 2023 it did that arithmetic in public: at 1%, the 75,000 papers it submitted in 2022 would have produced roughly 750 wrongly labelled ones. Pitt's Teaching Center, in guidance updated in February 2026, does not endorse or support any AI detection tool and has disabled the one in Turnitin. A 2025 NBER working paper reduces the choice to two numbers, the false negative rate and the false positive rate, and the tools it tested were far apart on both.
The false positives land on writing that is formulaic by design, which in a school is a great deal of it: the lab report that should read like every other lab report, the essay written to the shape a rubric asked for, the second language writer whom detectors misclassify most often. None of that is misconduct. All of it scores like it. Turnitin is the one vendor that documents its own failure cases in any detail, and our page on Turnitin's AI detection goes through them with the false positive rates it publishes.
That gap between the vendor page and the classroom is why these searches end on Reddit, where teachers and students comparing notes have nothing to sell and nothing but anecdote: the results they report vary because the texts they tested vary.
Can a school prove I used AI?
Not with a detector alone. A detector returns a likelihood that a passage resembles machine output, a claim about the writing, not the person. It has no access to who typed what, and it cannot tell your plain, hurried writing from a model's. Turnitin says outright that it does not determine misconduct, and GPTZero says its results should not be the final verdict. A school acts on its own academic integrity process, where a score is one input beside your drafts, version history, earlier work, whether the citations exist, and what you can say about your argument when asked.
If you have been wrongly flagged
The number comes without explanation, so move the conversation to the record of how the work was made.
- Ask which tool produced the score, and what it was. An asterisk, a 38% and a 96% are three different conversations.
- Ask what else the concern rests on. If it is only the detector, ask for the institution's policy on detector evidence in writing.
- Ask to see the report. You are entitled to know which sentences were highlighted.
- Hand over the drafting history. Google Docs keeps version history and lets you name versions, Word keeps its own, and both show a document built rather than pasted. Grammarly Authorship records what was typed, pasted or generated as you write, and produces a report you can hand over.
- Offer to talk it through. Explaining your argument, your sources and your choices is evidence a detector cannot produce or contradict.
None of this guarantees an outcome. What it does is move the ground from a percentage nobody can interpret to a record you can show.
For teachers deciding what to use
Search "AI detector for teachers" and you get a shortlist with no way to choose between the items: published evaluations use their own corpora and cover different tools, so none compares that shortlist on the work you actually mark. Treat a score as triage, telling you which of forty scripts to read properly. It is not an adjudication, and no vendor claims it is.
Two things are worth more than a better detector. One is evidence you can read: a tool showing the passages and signals behind a number lets you judge it, and our page on the signs of AI writing sets out those signals and where each is wrong often enough to matter. The other is process, where the market has moved: Turnitin Clarity, released in 2025, gives instructors revision timelines and pasted versus typed text instead of a verdict, and Instructure frames its own position on integrity in Canvas the same way. Drafts, in-class writing and two minutes on the argument do the same job without a licence. Check what your institution has enabled, because its policy may already say a score cannot stand alone.
This site gives a second reading: paste a passage, get a likelihood and the parts that produced it. Treat it as this page treats every other detector, as something that tells you where to look. A student can run a draft through it before submitting, which an indicator students never see cannot do. It judges where the words came from, not whether they are any good, and it cannot tell you who typed them.
Last updated 2026-08-16.