What AI detector does Turnitin use?
Turnitin describes its AI writing detector as a proprietary model based on an open-source foundation model from Hugging Face, retrained and tuned with Turnitin’s curated academic-writing data. It is a separate prediction from the Similarity score, not a claim that Turnitin simply embeds GPTZero or Copyleaks. Current documentation says coverage changes as the model is updated, and the result is evidence for educator review—not proof of misconduct.
Why — the first-principles explanation
There are two different questions hidden in “what detector does Turnitin use?” First, what model runs the prediction? Turnitin says its detector is proprietary, based on an open-source foundation model, and retrained, evaluated and fine-tuned with its own curated data and long experience with authentic student writing. Second, what writing can that model recognize? Turnitin publishes a changing list of supported model families and says it has expanded beyond its initial GPT-3/GPT-3.5 coverage to newer GPT, Gemini, Claude and LLaMA-based tools. That list is a capability statement, not a guarantee that every output from every model will be detected. The report analyzes qualifying long-form prose in overlapping segments, assigns sentence-level scores from 0 to 1, and averages them into an estimate of the qualifying text likely generated or AI-paraphrased. This is separate from the Similarity score, which measures matching text. A percentage is not a verdict, and it is not necessarily a percentage of the whole file: Turnitin excludes non-prose forms such as code, poetry, tables and bullet lists. Product access, file requirements, language support and the model version used by an institution also matter. Treat the current Turnitin guide as the source of truth rather than an old 2023 whitepaper or a detector-comparison chart.
An example that makes it click
A student submits a 1,200-word English essay through an institution that licenses Turnitin Originality and has AI writing detection enabled. Turnitin’s report may show a percentage for qualifying prose and highlight sections, while the Similarity report shows a separate matching-text percentage. If the essay includes a table, code block or bullet list, that material may not count toward the AI percentage. A low score with an asterisk, a high score, or no report still needs the instructor to review drafts, sources, policy and the student’s explanation; none of those states proves authorship by itself.
How to do it
- Open the current Turnitin AI writing detection FAQ or AI Writing Report guide. Product behavior and supported model coverage change, so do not rely on a 2023 architecture PDF alone.
- Identify the product and license: Turnitin says AI writing detection is available through Turnitin Originality, or as an AI writing add-on for iThenticate 2.0; ask the institution’s administrator if the indicator is missing.
- Separate the two numbers in the report. The Similarity score measures matching text; the AI writing percentage estimates qualifying text likely generated or modified by AI. They are independent.
- Check file requirements before interpreting a missing report: under 100 MB, at least 300 words of long-form prose, no more than 30,000 words, a supported language, and an accepted file type.
- Check the language and capability boundary. English, Spanish and Japanese long-form submissions can receive AI-writing detection; AI paraphrase and AI-bypasser detection are currently English-only in Turnitin’s guide.
- Read the percentage as qualifying prose text, not the whole document. Code, poetry, scripts, tables, bullet lists and other unconventional or short-form text may not be reliably analyzed.
- If a new report shows an asterisk below 20%, do not convert it into an exact score. Turnitin says this range has a higher incidence of false positives and suppresses the exact percentage for newer reports.
- For an academic decision, combine the report with drafts, citations, process evidence, conversation and the institution’s policy. Turnitin explicitly says the indicator should not be the sole basis for adverse action.
- If the detector has changed, remember that old results are not automatically recalculated. Turnitin says a submission must be resubmitted to generate a result under a newer model or newly enabled feature.
- Never upload a student paper to an unrelated free detector just to obtain a second percentage without checking privacy, consent, retention and institutional rules.
Key facts
- Turnitin says its AI writing detector is a proprietary model based on an open-source foundation model from Hugging Face, with retraining, evaluation and fine-tuning by Turnitin.
- Turnitin’s published training description includes AI-generated and authentic academic writing across time, geographies and subjects, with attention to second-language learners and under-represented areas.
- The detector works on overlapping text segments of roughly a few hundred words, assigns sentence scores from 0 to 1, and averages them into an overall prediction.
- Turnitin’s current FAQ lists coverage that has expanded across GPT, Gemini, Claude and LLaMA-based tools; model coverage is updated and is not an exhaustive guarantee.
- The AI writing percentage is independent of the Similarity score and appears in the AI Writing Report within the Similarity Report workflow.
- The percentage refers to qualifying long-form prose, not necessarily the percentage of the entire file.
- A report requires at least 300 words of prose, no more than 30,000 words, a file under 100 MB, a supported language and an accepted file type.
- Turnitin supports long-form English, Spanish and Japanese detection; English is the only one of those languages currently documented with AI paraphrase and bypasser detection.
- Newer reports suppress exact values between 1% and 19% and use an asterisk because false positives are more likely in that range.
- Turnitin warns that the model can misidentify human, AI-generated and AI-paraphrased text and says educators must not use the indicator as the sole basis for adverse action.
Interpret the detector without overclaiming
Separate Turnitin’s model, Similarity score, file limits and institutional policy before choosing an AI-detection tool or making an academic decision.
▶ The 60-second explainer (script)
What AI detector does Turnitin use? Turnitin describes it as its own proprietary AI writing detection model, based on an open-source foundation model and retrained with curated academic-writing data. It is not the same thing as the Similarity score. Turnitin breaks qualifying prose into overlapping segments, gives sentences scores from zero to one, and estimates how much qualifying text may have been generated or changed by AI. Its published coverage has expanded across GPT, Gemini, Claude and LLaMA-based tools, but the list changes and it is not a guarantee for every model or every paper. The report needs at least 300 words of long-form prose, supports English, Spanish and Japanese, and currently provides paraphrase and bypasser detection only for English. Code, poetry, tables and bullet lists may not be reliably analyzed. A score—especially a low score with an asterisk—is a clue, not proof. Turnitin says educators must combine it with human review and policy, never use it alone to punish a student.
What authoritative sources say
People also ask
Does Turnitin use GPTZero or Copyleaks?
Turnitin’s current documentation describes its own proprietary AI writing detection model, based on an open-source foundation model and Turnitin’s retraining and data. It does not document the AI Writing Report as a GPTZero or Copyleaks score.
Is Turnitin’s AI detector GPT-4 or ChatGPT?
No. GPT, Gemini, Claude and other model families are the kinds of outputs Turnitin says its detector can evaluate; they are not the detector itself. Turnitin’s coverage list changes as its model is updated.
What does Turnitin’s AI percentage mean?
It estimates the share of qualifying long-form prose that Turnitin’s model predicts was likely generated by an AI writing tool or likely AI-generated and modified by a paraphraser or bypasser. It is not necessarily a percentage of the entire file.
Is Turnitin’s AI score the same as the plagiarism or Similarity score?
No. Turnitin says the AI writing percentage and Similarity score are independent. Similarity measures matching text; AI Writing estimates likely AI-generated qualifying prose.
How many words does Turnitin need to run AI detection?
The current file requirements say at least 300 words of prose in a long-form format and no more than 30,000 words, plus a supported language, accepted file type and file size under 100 MB.
Does Turnitin detect Spanish or Japanese AI writing?
Yes, Turnitin currently documents long-form English, Spanish and Japanese AI-writing detection. Its AI paraphrase and bypasser capabilities are documented as English-only.
Why does Turnitin show an asterisk instead of a percentage?
For newer reports, Turnitin suppresses exact values above 0% and below 20% and shows an asterisk because it has observed a higher incidence of false positives in that range.
Can Turnitin prove a student used AI?
No. Turnitin says the model can misidentify text and the indicator should not be the sole basis for adverse action. Educators need to review process evidence, the work, the student’s explanation and institutional policy.
Can the Turnitin AI percentage change later?
Yes. Turnitin updates its model and says an existing submission is not automatically recalculated; resubmit it to obtain a result under the newer model or newly enabled feature.
Who can see Turnitin’s AI writing indicator?
Turnitin says the indicator and report are available to instructors and administrators when the institution has the relevant feature enabled. Students may not see the same report view.
The same question, asked other ways
- What AI checker does Turnitin use?