Turnitin AI detection accuracy: vendor claims and independent tests

This record is one of eight in the studies index and the only one built from vendor statements rather than a study. The Turnitin 2023 claims are indexed because Turnitin is the detector most often tested by the other records, and because the vendor's figures and the study figures measure different things.

One detector: Turnitin AI writing detection, the indicator inside the Similarity Report. The help centre describes a transformer-based deep learning model, updated for English in October 2026, with separate models for Spanish, Japanese, Arabic, Turkish, German and Dutch (Turnitin FAQ).

Detector as described on turnitin.com and guides.turnitin.com, read November 8, 2026 (Turnitin FAQ).

The samples behind the vendor claims are human-written corpora only (Turnitin, May 23, 2023).

Samples as stated by Turnitin, Turnitin, May 23, 2023 and Turnitin FAQ.

The metric Turnitin publishes is a false positive rate at document and sentence level. The Turnitin model scores each sentence between 0 and 1, labels sentences over a threshold as AI, and reports the share of sentences above the threshold as the document score (Turnitin FAQ).

The table lists each Turnitin statement with its date, then the three study results in this index that tested Turnitin (Turnitin, June 14, 2023).

Vendor values from turnitin.com and guides.turnitin.com read November 8, 2026 (Turnitin, May 23, 2023); study values as printed in each paper and sourced on the linked record pages.

The vendor claim has stayed constant since 2023. Turnitin's March 16, 2023 post describes "a high accuracy rate accompanied by a less than 1% false positive rate", while admitting the rate "is not zero" (Turnitin, March 16, 2023). The May 23, 2023 Turnitin post narrows the claim to documents with over 20% AI writing and concedes that below 20% "there is a higher incidence of false positives" and that "real-world use is yielding different results from our lab".

The trade-off is stated by Turnitin itself. The current FAQ says that holding the 1% document rate means "there is a chance that we might miss some AI written text", and that a document scored at 50% AI "could contain as much as 65% AI writing" (Turnitin FAQ). The same Turnitin FAQ shows no score in the 1% to 19% range, only an asterisk, and says the indicator "should not be used as the sole basis for action".

The study results sit beside the claim rather than against it, because none measures the same quantity. Weber-Wulff 2023 placed Turnitin first of 14 detectors at 41 of 54 documents (76%) with 0 false accusations in 18 human documents (DOI 10.1007/s40979-023-00146-z). Walters 2023 recorded Turnitin at 100% on 126 documents while warning the human papers may have been in Turnitin's training data (DOI 10.1515/opis-2022-0158). Perkins 2024 recorded Turnitin falling from 50% to 7.9% after adversarial edits, with 0 false accusations in 10 human samples (DOI 10.1186/s41239-024-00487-w).

The model has changed repeatedly since those tests, by Turnitin's own release notes: paraphrase detection on December 6, 2023, bypasser-tool detection on August 27, 2025, recall updates on October 14, 2025 and February 12, 2026, and an English update on October 14, 2026 (Turnitin release notes). None of those notes carries a number. The repeatability question, added 9 November 2026, lists these dates beside the dated model changes of eight other vendors, and the results as evidence question reads the FAQ and the 23 May 2023 post as two of the seven primary documents on acting on a score.

Turnitin lists its own failure cases: content without much structural variation, text that repeats itself, and text paraphrased without new ideas can be flagged. In documents of a few hundred words the prediction is "mostly 'all or nothing'". The model does not reliably detect AI text in non-prose, code or bullet points, and paraphrase and bypass detection exist only for English.

The published record stops at false positives: no true positive rate, recall or overall accuracy figure, no published composition for the 700,000 to 800,000 paper corpus, and no third-party audit. The FAQ says the training sample accounted for second-language learners, but publishes no bias test result (Turnitin FAQ).

A false positive rate measured on papers written before ChatGPT says how often the detector flags old human writing; how often the same detector catches new AI writing is a number Turnitin has never published.

One card: the Turnitin card carries the dated vendor claim and the three study results.

Turnitin appears in the ranking of detectors by independent test record on the strength of the three studies above. How this record is sourced explains why vendor claims are dated and never averaged with study results, and the home page shows which AI detectors exist and who has tested them.

Related records: Weber-Wulff 2023, Walters 2023 and Perkins 2024 are the three tests of Turnitin in this index.