How Turnitin's AI writing detection works

AI indicator vs similarity report

Turnitin is best known for its similarity report, which matches submitted text against a database of sources. That has nothing to do with AI. Its AI writing indicator is separate: a trained classifier that estimates what percentage of a document's prose may have been generated by a model. It is shown to instructors (not always to students) and Turnitin explicitly frames it as a starting point for a conversation, not a determination.

The classifier works on chunks of long-form prose and reports a document-level percentage; Turnitin has stated it requires a minimum amount of text and that very short or heavily formatted submissions are not scored reliably.

Published limits and false positives

Turnitin has acknowledged a false-positive rate at the document level and has adjusted thresholds over time to reduce it, at the cost of missing more AI text. It advises institutions not to act on the indicator alone. Reported weak spots mirror other detectors: non-native writers, formulaic genres, mixed human-and-AI documents, and text that has been paraphrased or lightly edited.

Because the tool is tuned by the vendor and updated without notice, results for the same text can change between terms. No third party, including us, can promise how it will score a given document.

The honest approach for students

Know your institution's AI policy and follow it; many allow assistance with disclosure. Keep drafts and notes so you can show your process. Write in your own voice, with specifics only you would know. AntiGPT helps AI-assisted drafts read naturally and lets you check an estimated score. It is a writing tool, and it does not change what your policy requires you to declare.