Academic checks

Can Turnitin detect ChatGPT watermarks in 2026?

No. Turnitin’s AI writing report and OpenAI’s ChatGPT text watermark are different instruments. Turnitin classifies how a passage is written. textGrain is a pattern OpenAI can look for only with its own key. A percentage in one system does not confirm or clear the other.

That split matters in 2026 because the two systems are both in the news at once. Turnitin’s AI writing model reports likely AI-generated text and likely AI-paraphrased text inside Originality. OpenAI, separately, is adding textGrain watermarks to eligible ChatGPT text in the EU, and offering an opt-in for API customers worldwide. Students, instructors, and integrity offices search as if those were one switch. They are not.

Two questions people collapse into one

“Can Turnitin detect ChatGPT?” and “Can Turnitin detect the ChatGPT watermark?” look like the same search. The first is about a writing classifier. The second is about a provenance signal embedded in token choice. An institution can have a confident answer to the first and no access at all to the second.

OpenAI’s own help center draws the line. Third-party tools, it says, typically use classifiers on patterns such as word choice after the text already exists. The EU AI Act’s marking rule asks for a signal embedded in the generated text. textGrain is OpenAI’s version of that embedded signal. Turnitin’s published description of its model is the classifier kind: a transformer trained to separate likely machine prose, likely machine prose that was then spun, and the rest of a submission.

What Turnitin actually scores

Turnitin’s similarity report and its AI writing report are different products that happen to sit in the same workflow. Similarity compares a submission with other sources. The AI writing report is a separate indicator. Turnitin’s guide describes highlights for text the model treats as likely AI-generated, and a further breakdown when that text looks likely to have been revised by an AI paraphraser or word spinner.

The report is a percentage of qualifying text, not a courtroom finding. Turnitin has also changed how very low scores are shown: scores it treats as below a 20 percent threshold are indicated with an asterisk rather than a precise percentage on reports generated after that change. A star is not a watermark miss. It is Turnitin declining to put a fine-grained number on a weak classifier signal.

Language coverage is not one global model. Turnitin documents separate Japanese and Spanish AI writing models trained on writing in those languages and on named GPT-family models. An English report and a Spanish report are not interchangeable, and neither one is a textGrain decode. If a department policy treats the percentage as proof of a specific product, the policy is ahead of what the vendor page claims.

Institutional tools are still evolving, and a public page cannot reproduce an Originality report. Where you are allowed to inspect a draft yourself, a ChatGPT watermark detector can list invisible Unicode and, for a file, C2PA or AI metadata. It will not print a Turnitin percentage, and it should not be used to shop for a score before a submission.

What a ChatGPT watermark is

textGrain changes which token the model samples when several wordings are plausible. The pattern is in the wording. It does not add a hidden character, a comment, or a tag that a plagiarism database can match. The longer explanation, including where the EU rollout applies and where the API switch stays off, is in how the OpenAI text watermark works.

Detection needs the same key and the same tokenizer settings used at generation time. OpenAI says enabling the API watermark does not give the customer the detector. Access is reviewed for research and academic organizations studying the mark. A university that licenses Turnitin has not, by that contract, received OpenAI’s key.

The public checker at openai.com/verify looks at supported image and audio files for SynthID or a trusted C2PA manifest. Pasting an essay there is the wrong surface. Instructors who upload a PDF of a paper are checking a document container, not the statistical pattern in the sentences, unless the file happens to carry some other provenance metadata.

Why the two results diverge

A classifier and a keyed watermark fail in different places. Turnitin can flag prose that never passed through ChatGPT: another model, a heavy template, or in some cases unusual but human writing. textGrain can miss ChatGPT text that was generated outside the marked path, that is shorter than the reliable window, that is code, that is a near-quote, or that was rewritten enough to break the pattern. Both misses are expected. Neither miss proves the other system wrong.

Length is the clearest split. OpenAI points at the EU Code of Practice, which does not require a watermark on outputs shorter than about 200 tokens, roughly 150 English words, or on code snippets. A discussion-board post can be long enough for a style model to score and too short, or too constrained, for textGrain to be detectable. A 2,000-word methods section is the opposite problem: enough text for both systems, and still no reason to treat their answers as the same fact.

Editing sits in the middle. Light proofreading is what a sampling watermark is built to survive. A full paraphrase is what a style model is increasingly trained to notice, which is why Turnitin added an AI-paraphrase category. The same edit can weaken a watermark and strengthen a paraphrase flag. Reading only one of those outcomes produces a false story about “removal.”

Paraphrasers and “AI bypass” flags

Turnitin’s model notes say the AI-generated portion of a report can now include text that looks modified by an AI bypasser: a spinner or paraphraser whose job is to keep the claims and change the surface. That feature is still a classifier. It does not decode textGrain, and it does not know whether a human editor or a second model made the change.

Cleaning invisible Unicode is not in that category. Zero-width spaces do not alter a single visible word, so they do not move a wording score. They also do not carry textGrain. The separate guide to ChatGPT hidden characters is about paste hygiene. It is not a way to change an integrity report, and this site does not offer a rewrite button that would try.

If the institutional rule is “disclose assistance,” neither score is the disclosure. A watermark hit would be evidence that an OpenAI model likely processed text. A Turnitin highlight is evidence that a classifier is suspicious. A student or researcher who used a model and then relies on either miss is betting on a measurement error. That is not what either vendor describes as the point of the tool.

What the 2026 rollout changes

Before textGrain, “ChatGPT watermark” in student searches usually meant invisible characters or a rumor about a secret detector. In 2026 OpenAI’s public note is more specific. Eligible ChatGPT and Codex text generated in the EU is being marked to meet the EU AI Act. API watermarking is opt-in worldwide and is not the default at launch. The details of who is covered are in does ChatGPT watermark text in 2026.

For a campus, the practical change is interpretive, not a new Turnitin toggle. An integrity office still cannot paste a paper into a free site and receive OpenAI’s verdict. If the office later receives detector access through OpenAI’s research program, that result belongs next to the writing report, not instead of it. One can be positive while the other is quiet: marked text that was heavily edited, or unmarked text that still looks machine-written because it came from a different model.

Geographic assumptions fail quickly. A student in a course may use ChatGPT outside the EU path, an API wrapper a writing tool enabled, or another vendor entirely. Claude’s text watermark, which Anthropic describes as a SynthID-Text variant applied broadly, is a different key again. Turnitin does not become a universal reader of every lab’s mark because one lab turned a mark on.

What you can check on a draft you have

On text you are allowed to process, three checks stay available without pretending to be Turnitin:

• Read the institutional report if you are the instructor or the student who received it. Note whether the highlight is “AI-generated” or “AI-paraphrased,” and whether a low band is starred rather than numbered.
• Scan the paste for invisible characters with the ChatGPT watermark detector. A hit explains a broken diff or a weird search. It does not explain a Turnitin percentage.
• If the submission is a file export, check the container for C2PA or other AI metadata. A PDF or DOCX can carry document properties that the sentences do not. That is the file layer in the C2PA guide, not a text watermark.

Do not use those checks to tune a submission until a classifier goes quiet. The removal guide explains why deleting characters, stripping metadata, and rewriting are three different operations, and why none of them is a certificate that a person wrote the page.

How to read a real case

Take four outcomes that already show up in appeals. They are not a ranking of guilt. They are a map of what each tool was able to see.

• High AI-writing score, no invisible characters. The ordinary case. The classifier is looking at sentences. The character scan is looking at codepoints. Both results can be accurate.
• Low or starred score, text known to be from ChatGPT. Possible when the prose was short, heavily edited, outside the model’s training neighborhood, or in a language configuration the report was not built for. This does not mean a watermark is absent. It means this classifier was quiet.
• AI-paraphrase highlight after a “humanizer.” Turnitin is scoring the revision pattern it was updated to catch. The humanizer did not remove a key it never had. If the source was marked with textGrain, a heavy rewrite is also how that pattern gets harder to detect. The highlight is not proof the watermark survived.
• File metadata says “trained algorithmic media,” prose score is low. Someone exported or edited a document that still carries a provenance tag. The sentences may have been replaced. Check the container and the prose separately.

A missing watermark, in particular, is easy to over-read. OpenAI lists the boring reasons a detector stays quiet: the text predates the rollout, the path was unsupported, the passage was short or factual, or the wording moved. That argument is spelled out in what a missing watermark means.

The comparison across free checkers, including which ones are classifiers and which ones read a file manifest, is in free AI watermark checkers compared. None of the free paste boxes in that list is Turnitin, and none of them is OpenAI’s text detector.

FAQ

Related guides: how textGrain works, watermark versus an AI detector, and the full article index.

← All AI watermark guides