What AI Detector Does Turnitin Use? How it works.

An in-house engine, not a third party

The question in this article’s title is searched thousands of times a month, usually on the assumption that Turnitin must wrap someone else’s detector. It does not. Turnitin’s AI writing detection is its own classifier, developed internally and run alongside, but separately from, the similarity checking it has sold for two decades. The two reports answer different questions: similarity compares your text against a source database; the AI report estimates whether segments of your text were machine generated, with no database involved.

How the system reads a paper

Turnitin’s published methodology describes segmenting a submission into overlapping windows of sentences, classifying each window, and aggregating the results into the document-level percentage instructors see. The percentage is an estimate of how much of the document is likely AI generated, not a confidence level in an accusation. The system requires a minimum amount of prose (long enough text, in English or supported languages) to produce a score at all, and Turnitin suppresses very low percentages because they proved unreliable. Sentence-level highlighting shows which passages drove the number.

What Turnitin says about its own limits

Turnitin’s guidance to institutions is more careful than campus rumor suggests: the score should not be the sole basis for an integrity action, low percentages should be treated with particular caution, and human review is required. The company has acknowledged elevated false positive risk in specific conditions, and independent reporting has documented universities disabling the feature after their own validation testing. None of this means the tool is useless; it means the tool is an instrument whose reading requires interpretation, which is true of every detector including the one on this site.

Paraphrased AI and the arms race

Turnitin claims coverage of AI text that has been paraphrased by tools built for evasion, with reduced reliability. The claim matches the general physics of detection: paraphrasers disrupt the statistical pattern, human revision disrupts it more, and at some level of rework the text honestly is no longer machine written. If you want the mechanics underneath all of this, read how AI detectors work; for the measured accuracy picture, how accurate is Turnitin AI detection collects the numbers.

A short timeline of Turnitin's AI detection

Context helps when reading the tool's claims. Turnitin shipped AI writing detection in April 2023, weeks after ChatGPT made the problem urgent, and enabled it by default for most institutional customers. Within months, universities running their own validation found more false positives than the launch messaging implied, and several large institutions disabled the feature publicly. Turnitin responded by publishing false positive disclosures, suppressing low-end scores entirely and adding caution language to the instructor report. Since then the system has iterated quietly: thresholds moved, paraphrase coverage was added, and the guidance shifted from detection to conversation starter. The arc matters because it is the whole industry's arc compressed: launch confidence, measured reality, retreat to probabilities. Any school policy written in 2023 and untouched since is calibrated to a tool that no longer behaves the way the policy assumes.

What instructors and students each control

Instructors choose whether the AI report influences grading, whether students are told the feature is on, and whether process evidence is requested before any accusation: three choices that determine most outcomes. Students control their evidence trail: drafting in a tool with version history, keeping notes and outlines, and pre-reading their own prose with a free scanner. Neither side controls the score itself, which is exactly why both sides preparing around it beats either side trusting it.