Is Turnitin's AI detector accurate? The short answer
The honest answer to "is Turnitin's AI detector accurate?" depends on which job you mean. At one job, mostly yes: in published tests it rarely labels long, fully human academic writing as AI. At the other job, catching AI writing, it misses a fair amount, and it missed more as AI models got better. Turnitin built it that way on purpose. It would rather let some AI text through than accuse a student who wrote every word.
We haven't tested Turnitin ourselves, so everything below comes from Turnitin's own documentation and from independent studies, with dates, because the model keeps changing. Our free AI detector runs a different model and reports the share of a text that reads human, so its score won't match Turnitin's. For false positives across all detectors, see are AI detectors accurate.
Turnitin AI detection accuracy: what Turnitin claims, and how that changed
Turnitin's current claim is in its AI writing detection FAQ: it aims to keep its false positive rate under 1% for documents with more than 20% AI writing. In its own words, it "might flag a human-written document as AI-written for one out of every 100 fully-human written documents." Before every update, it says, it re-tests the model on more than 700,000 academic papers written before ChatGPT existed.
The same FAQ is open about the cost. To keep false positives that low, the detector lets some AI text through: if it reports that 50% of a document is AI, the document "could contain as much as 65% AI writing."
| When | What Turnitin said or changed |
|---|---|
| February 2023 | Before launch, it said its detector caught 97% of ChatGPT and GPT-3 writing in its lab, with a false positive rate under 1 in 100. |
| April 4, 2023 | Detection switched on for existing customers, described as having "high accuracy and low false positive rates." |
| May 2023 | Turnitin wrote that real-world use was "yielding different results from our lab". It found more false positives below 20% AI and marked those scores with an asterisk, raised the minimum length from 150 to 300 words, and changed how it scores the first and last sentences of a document. |
| July 2024 | Scores from 1% to 19% stopped being shown at all: the report displays *% with no highlights, "to avoid potential incidence of false positives." |
| October 2025 and February 2026 | Model updates "to improve recall while maintaining a low false positive rate", meaning to catch more AI text. |
| July 2026 | Several models merged into a single model, which Turnitin says keeps the false positive rate under 1%. |
One detail stands out. In May 2023 Turnitin considered hiding scores under 20%, then decided against it because educators wanted to see them. Fourteen months later it hid them anyway. Between those dates it also started detecting AI text run through paraphrasers and "bypasser" tools; does Turnitin detect ChatGPT covers what it says it can catch.
Note what the headline number covers: whole documents scored above 20%. It says nothing about the 1% to 19% range, the one Turnitin found less reliable and now hides, or about single sentences: in 2023 Turnitin put its sentence-level false positive rate at about 4%.
How accurate is Turnitin's AI detector in independent tests?
Three published tests measured Turnitin directly, on texts where the researchers knew who or what wrote each one.
| Study | What was tested | What it found |
|---|---|---|
| Kramer and Javorková, Open Information Science, August 2026 | 100 psychology journal introductions published in 2016, plus new versions of each written by GPT-4o and by GPT-5 | 1 of 100 human texts flagged (it scored 32%). Missed 26% of the GPT-4o texts and 47% of the GPT-5 texts at Turnitin's 20% line. |
| Perkins and colleagues, Journal of Academic Ethics, October 2023 | 22 assignments written with GPT-4, using prompts meant to avoid detection, marked by 15 staff alongside real student work | Flagged 91% of the submissions as containing AI, but identified only 54.8% of the AI-written content. |
| Temple University's teaching centers, early version of the detector | 120 texts: 30 human, 30 ChatGPT, 30 ChatGPT run through free paraphrasers, 30 mixed | 28 of 30 human texts scored correctly. 77% of the ChatGPT texts and 63% of the paraphrased ones scored 100% AI. Only 13 of 30 mixed texts scored between 0% and 100%. |
Read together, they point the same way. Turnitin's low false positive rate holds up on long, polished human writing: the 1 in 100 from the 2026 study matches the company's own claim. Catching AI is the weak side. Nearly half of the GPT-5 introductions passed as human, and the authors note that their simple prompt may make even that an underestimate.
Mixed drafts are the hardest case. When the Temple team compared highlighted sentences with the parts AI actually wrote, they "found no relationship at all." Turnitin's FAQ says something similar in milder words: in a document that mixes both, "it can be difficult to exactly determine where the AI writing begins and original writing ends."
Two caveats. Turnitin has updated its model since each of these tests, so the exact numbers may have moved. And none of them focused on non-native English writers. Turnitin's own test found no significant gap in false positives between English learners and native speakers on texts of 300 words or more, mostly short school essays, but a larger gap on texts under 300 words.
Turnitin false positives: when a score deserves less trust
Turnitin's own documents list the situations where its score is weakest. If a flagged paper fits one of them, say so.
- Short papers. Turnitin needs 300 words of prose. For a document of only a few hundred words, its FAQ says the prediction is "mostly all or nothing", so a draft mixing your writing and AI can be flagged as entirely AI.
- Repetitive or formulaic writing. Turnitin says false positives can include text without much structural variation, text that repeats itself, and paraphrasing that doesn't develop new ideas.
- Generic openings and closings. In 2023 it found more false positives in the first and last few sentences of a document, often stock introductions and conclusions.
- Lists, tables and other non-prose. The score covers only prose sentences in long-form writing. Bullet points, tables, poetry, code and annotated bibliographies aren't reliably assessed, so the percentage and the highlights can disagree.
- Heavy rewriting with AI features. Turnitin says spelling and grammar fixes from grammar checkers were mostly not flagged in its tests, but text produced by their drafting, paraphrasing or summarizing features likely will be.
Low rates still add up across thousands of papers. ABC News reported in October 2025 that Australian Catholic University logged nearly 6,000 alleged misconduct cases in 2024, about 90% of them about AI. Students said they waited months to be cleared, in some cases with a Turnitin AI report as the main evidence; one nursing student's results were withheld for six months. The university called the figures "substantially overstated" and said cases resting only on the AI report were dismissed. It stopped using Turnitin's AI tool in March 2025.
What a Turnitin AI score means, including the asterisk
The percentage is the share of qualifying text, meaning prose sentences in long-form writing, that Turnitin's model thinks was likely written by AI, or written by AI and then run through a paraphraser or bypasser. It isn't the share of the whole file. It's also separate from the similarity score, which measures text matching other sources. Here's what each state in the AI writing report means.
| What the report shows | What it means |
|---|---|
| 0% | No qualifying text was identified as likely AI. |
| *% | Turnitin found something between 1% and 19% but shows no number and no highlights, because that range has more false positives. Reports made before the July 2024 change may still show a number. |
| 20% to 100% | The share of qualifying prose identified as likely AI-generated or AI-paraphrased, with highlights showing where. |
| -- (gray, no score) | Not processed: fewer than 300 words of prose, more than 30,000 words, an unsupported language or file type, or submitted before detection was enabled. |
Scores don't update on their own. A report keeps the result of the model version that produced it, and Turnitin only rescores a paper that is submitted again. Students can't open the AI report at all, so if you're asked about a score, ask your instructor for a copy.
Flagged by Turnitin? What to do
Start with Turnitin's own position: its score should not be used as the sole basis for adverse actions against a student. Reply calmly and in writing.
- Ask for the AI writing report. Instructors can download it as a PDF and share it. You need the score, the highlighted passages and the date it was generated.
- Check it against the weak spots above. A score just over 20%, a short paper, highlights on your introduction and conclusion or on formulaic sections such as methods are all worth pointing out.
- Show your process. Version history, drafts, notes and sources say more about authorship than any score. Offer to talk through your argument in person.
- Use the formal process if needed. Quote Turnitin's guidance, ask what other evidence exists, and follow your school's appeal route.
Our guide on why writing gets flagged as AI walks through the evidence and the conversation step by step.
You can't run Turnitin on your own drafts. Another checker, ours included, uses a different model, so it can show which sentences read like AI but can't predict your Turnitin score; checking a draft before you submit explains the difference. And if your course doesn't allow AI, no checker changes that: follow your school's rules.