Detexta

Can Turnitin Detect ChatGPT? What the Evidence Shows

The short answer

Sometimes, partially, and less reliably than the marketing suggests. Turnitin does run an AI-writing model alongside its similarity check, and it does flag text produced by ChatGPT and other models. But detection is a probability judgement about style, not a fingerprint match, and it degrades quickly the moment the text is edited.

What the vendor says, and what testing found

When Turnitin launched its AI writing indicator, it published accuracy figures and a low false-positive rate, which were widely reported. Independent researchers and journalists then tested those claims, and the picture that emerged was more conditional than the launch material implied:

None of this is a secret to the vendor. It is why every detector positions its score as a signal for a human to review.

Why a high score is not proof

The model is asking one question: does this passage look statistically like generated text? It has no idea who typed it. That is why the false-positive problem is real and documented — formulaic writing, technical prose, and the work of writers composing in a second language all look more machine-like than they are. A number near 100 does not mean someone used ChatGPT; it means this text shares the texture the model associates with generated prose.

What this means if you did use a model

If you used a model for a draft and then rewrote it substantially in your own words, the detector may or may not catch it, and you should not build a plan around the may. What decides the outcome in almost every real case is not the score but whether you can explain the work, and whether your drafts and notes show the process.

If you used a model where it was not allowed, the honest problem is the policy, not the detector. Detection is unreliable in both directions — it misses a lot and it accuses innocent writing — so institutions that rely on it as their only enforcement mechanism are relying on a noisy instrument.

What still works, in order of reliability

  1. Your document version history. A dated trail from outline to final draft is the strongest evidence of authorship, and it is not something a detector can argue with.
  2. Your notes and sources. Highlighted PDFs, reading notes, an outline. These show the argument was built.
  3. An in-person conversation. Most instructors who suspect something will ask you about the work. Being able to explain your choices is worth more than any score.
  4. Consistency with your previous work. A sudden register shift across submissions is what actually triggers suspicion in practice.

What does not work

A note on the numbers you see online

Most "Turnitin detected my ChatGPT essay" posts are people reading a percentage on a screenshot, with no information about threshold settings, document type or model version. Treat them as anecdotes about the tool's texture, not measurements of its accuracy.

Frequently asked

Does Turnitin detect ChatGPT reliably? Unedited output, often yes. Paraphrased and mixed output, much less so. It is a probability judgement, not a match.

Can Turnitin detect a lightly edited essay? That is the weakest case for every detector, and the most common real situation.

Is a 100% AI score proof? No. It means the model reads the text as generated. False positives are documented in formulaic and non-native writing.

What if I only used AI for brainstorming? Policies differ. Keep your notes from the session and ask before you submit, not after.