Can a teacher check a student’s text for the ChatGPT watermark?

A teacher who reads that ChatGPT now watermarks its text can fairly ask whether the mark settles a doubtful essay. For now it cannot. The signal OpenAI calls textGrain is readable only with a secret key, the detector that holds the key is limited to a few outside parties, and the tools schools already use are documented as scoring writing.

Can a school get OpenAI’s watermark detector?

OpenAI’s watermark detector is not public. OpenAI said on 5 October 2026 that access to its detector will initially be limited to approved researchers and expert organisations, and that, given the risk of missed watermarks and false positives, it is not making the tool publicly available at launch. TechCrunch reported the same limits that day. The technical report explains why the key matters: the detector requires only the generated text and the secret key, so whoever holds the key can test a passage and nobody else can.

Scope comes before access. OpenAI says the watermark is rolling out over the coming weeks to eligible ChatGPT and Codex users on all plans in the EU only, and that API customers can opt in for select models, off by default. So even a perfect detector would say nothing about much AI-written text, and a teacher cannot infer from a missing mark where, or whether, a student used a chatbot. Our article on what textGrain is covers the rollout in full.

What do Turnitin and GPTZero read instead?

Turnitin and GPTZero estimate how a text reads, and neither page mentions watermarks. Turnitin’s feature page says its model gives each sentence a score between 0 and 1, and averages the scores into the percentage a teacher sees. The same page says the figure “should not be used as the sole basis for action”. GPTZero’s FAQ says its model was trained on “a large, diverse corpus of human-written and AI-generated text”.

A classifier trained on examples can say a passage resembles the examples it learned from. Testing a passage against a secret key is another job, and a classifier has no key to test with. Our own AI detector belongs in this group. It returns a reading of register with quoted passages, not a percentage and not a watermark reading, and its limits are the ones above. Our article on what a detector score measures sets out why even a good classifier’s number is not a probability of authorship.

Turnitin has published error rates of its own, in this order. An update of 23 May 2023 said that scores under 20% carry a higher incidence of false positives, hence the asterisk. Its chief product officer wrote on 14 June 2023 that the document-level false positive rate is below 1% for documents with 20% or more AI writing, and that the sentence-level rate is around 4%. The June post advises instructors to use the result “to initiate a conversation, not to draw a conclusion”. The current feature page still states a target of under 1% for documents with over 20% AI writing.

What does a site that reports a watermark score actually show?

A page that reports a textGrain score or a “ChatGPT watermark detected” verdict is not shown to be using OpenAI’s detector. OpenAI limits detector access to approved researchers and expert organisations, and the key is secret. A page that claims a result owes readers an account of how it got access, and OpenAI’s announcement describes none for the public.

What such a page can run is easier to guess than to prove. It may be a style classifier like the ones above, or a scanner for invisible characters. Neither reads textGrain, which the technical report describes as keyed randomness coupled to how each token is chosen. Before trusting a result, ask what the page was given: the key, or only your text.

What would a positive or a missing watermark prove?

Even with a detector in hand, OpenAI’s own evaluation sets modest limits. The announcement reports that at a target false positive rate of 1% its detector found the mark in about 95% of 400-token passages and about 80% of 200-token ones on psychology questions, and that detection is substantially lower for mathematics. In a separate test on 400-token passages, OpenAI reports that replacing 10% of the words with synonyms lowered detection from about 92% to 66%.

OpenAI says so itself: “The absence of a detected watermark does not prove human authorship.” Its post adds that text from OpenAI tools may be too short, edited or translated for detection to work reliably. A positive result is limited too. OpenAI says a watermark can indicate that an OpenAI system generated or processed part of a passage, but not how much human judgment, editing or creativity went into it, and that it does not associate a person, account, prompt or conversation with the text.

The false positive rate matters most to whoever holds the stack of essays. Treat 1%, the rate that evaluation used, as arithmetic only: among 200 essays written without any AI, a detector held to a 1% target rate would be expected to flag about two. That is a target, not a measured error rate for a classroom, and the technical report states that its rate holds under explicit independence assumptions, which real coursework need not meet. A rate describes a population, and a teacher holds one essay.

What can a teacher look at instead of a score?

For most coursework the evidence that exists is process. Google Docs keeps a version history: Google’s help page says to click Last edit at the top right, and that you need permission to edit the file to browse earlier versions. Word keeps one too, with a limit: Microsoft says version history only works for files stored in OneDrive or SharePoint, so a file saved to a local drive has none. Those records show when text appeared and how it changed.

Setting this up before the assignment is the practical part. A task submitted as a shared document leaves a record, and a pasted file leaves none. Beyond records, the general practice that fits Turnitin’s own advice is a conversation: ask the student to walk through the choices in the essay or to explain a paragraph in their own words. We offer that as common practice, not as a tested method.

A student worried about a false flag benefits from the same habit. Keeping drafts and the version history documents how the work was done, and it changes no score. A record of edits shows how a document changed. It does not show where the words came from. A history can also be built by retyping a text, so it informs the conversation and decides nothing. The same is true of every detector named here.