Can Turnitin detect ChatGPT?

Sort of, but not the way most people assume. Turnitin gives your instructor a percentage — its estimate of how much of your prose reads like a language model wrote it. It isn't reading a watermark, and it can't tell ChatGPT from Claude, Gemini, or anything else. Its model was trained on how LLM prose behaves in general, so a flag means "this reads like machine output", not "this came from ChatGPT".

What the percentage actually means

It isn't a confidence level. A 40% score means roughly 40% of the qualifying prose got classified as AI-like. It does not mean Turnitin is 40% sure you cheated.

Turnitin's own published figures are worth knowing:

• Below 1% false positives at the document level — but only for documents it scores at 20% or above.
• Around 4% at the sentence level. That is one wrongly highlighted sentence in twenty-five.
• Under 20%, it shows an asterisk instead of a number, because false positives are more common down there.
• It is deliberately tuned to miss AI writing rather than flag people. Turnitin's chief product officer has said it accepts missing around 15%.

What it can't show

Who wrote it. Turnitin says so itself: the score should not be the sole basis for action against a student. Vanderbilt switched the feature off in August 2023, pointing out that a 1% error rate across the 75,000 papers it submitted in a year works out to roughly 750 students flagged for nothing.

The reverse holds too. A 0% score is not evidence a person wrote something, and no detector can give you that.

None of this involves anything hidden in your file, either. The indicator is about how the writing reads — it is not checking for invisible characters or metadata. If you want to see what is literally in your text, paste it into the detector, though a zero-width space tells you which software touched the text, not who wrote it. And if you are worried a watermark gave you away, there isn't one to find.