Does Turnitin detect ChatGPT?

Turnitin added an AI writing indicator in April 2023. It reports a percentage of the submission it believes was generated. That is a different measurement from the similarity score it has always produced — similarity compares your text against a corpus of existing documents, while the AI indicator looks at statistical properties of the writing itself.

Like every text detector, it is measuring predictability rather than authorship. Generated prose tends to choose likely next words and to hold a steady sentence rhythm. Human writing usually varies more. Those are real measurable differences, but they are properties of a writing style, not a record of who produced it.

Turnitin initially described a false-positive rate below 1%. In mid-2023 it revised that upward, acknowledging roughly 4% at the sentence level, and added a caveat for shorter documents where the estimate is less stable.

Vanderbilt University disabled the AI detector in August 2023. Its stated reasoning was arithmetic rather than ideological: processing around 75,000 submissions a year, even the optimistic 1% document-level figure implies roughly 750 papers a year incorrectly flagged. Pittsburgh, Boston University and Georgetown subsequently turned it off as well.

The point that matters is not that the tool is useless. It is that a small percentage applied to a large number of students produces a lot of wrongly accused people, and an institution has to decide whether it can absorb that. Several decided it could not.

The errors are not randomly distributed. Research has repeatedly found detectors flag writing by non-native English speakers at substantially higher rates, because second-language writing tends toward more common vocabulary and simpler sentence construction — the same features detectors read as machine-like.

The same applies to anyone writing in a deliberately plain register: technical and scientific writing, students taught to write in clear simple sentences, and first-generation students who have been coached toward formal, careful prose. A false positive lands hardest on people with the least standing to contest it.

Generally yes, and this is the awkward part. Substantive editing disrupts exactly the statistical patterns detectors rely on, and purpose-built rewriting tools automate that.

So the tool tends to fail against deliberate evasion while flagging people who wrote honestly in a consistent style. That inversion is the strongest practical argument against treating a score as evidence: it is least reliable exactly where the stakes are highest.

Ask what evidence exists beyond the score. Turnitin's own guidance is that the indicator is not proof and should not be the sole basis of an allegation, and most institutional policies say something similar even where the tool is still enabled.

Your process is your defence. Draft history, browser or document version history, notes, sources consulted, and the ability to discuss your argument in detail all demonstrate authorship in a way a percentage cannot rebut. If your institution uses Google Docs or Word online, the version history is often decisive.

Write in your natural voice rather than trying to sound less machine-like. Deliberately adding irregularity to avoid a detector is both unreliable and looks worse if it comes up.

Treat a score as a trigger for a conversation, never as its conclusion. Ask the student about their process — what they discarded, why they structured the argument as they did, which sources they rejected. Someone who did the work can discuss it.

Where practical, design the assessment so the question rarely arises: staged drafts, in-class or supervised components, oral defence, or work requiring specific personal or local context. These take more effort than running a checker and are the only approaches that hold up when challenged.