Does ChatGPT Have a Watermark? What Text Watermarking Is and What Is Publicly Known

People search for a ChatGPT watermark for two reasons: to know whether a teacher can prove text came from a chatbot, or because they found an odd character in pasted text and wonder if it is a tag. Both rest on a misunderstanding. A text watermark, as researchers define it, is not a mark added to the text but a pattern in which words the model chooses, detectable only by statistics. This page explains how that works, what the company behind ChatGPT has said on its own pages, how text watermarks differ from image labels, and why invisible characters are not one.

What text watermarking is

The watermark is in the choice of words, not in the characters.

The best-known method comes from "A Watermark for Large Language Models" by Kirchenbauer and colleagues, posted to arXiv in January 2023 and published at ICML 2023. The abstract describes a watermark as signals embedded in generated text that are invisible to humans but algorithmically detectable from a short span of tokens. Tokens are the word pieces a model writes.

The method works like this. Before each token is generated, the previous token is used to seed a random split of the vocabulary into a "green list" and a "red list". The model is then nudged to prefer green tokens. A person writing naturally lands on green about half the time by chance. Watermarked text has far more green words than chance allows, and anyone with the rule for drawing the lists can count them.

The paper turns that count into a standard statistical test. If the detector flags text only when its score, a z-statistic, is above 4, the authors calculate the chance of wrongly flagging human text at 3 in 100,000 (a one-sided p-value of 3 x 10^-5). They write that synthetic text can be detected from as few as 25 tokens. In their opening example, a passage a human might write would be expected to contain 9 green tokens; the watermarked version contained 28.

Is there a ChatGPT watermark? What its maker says

For a long time the public answer was no. In an August 4, 2024 update to a post titled "Understanding the source of what we see and hear online", the company behind ChatGPT said it had developed a text watermarking method that it continued to consider while researching alternatives. It described the method as highly accurate and effective against localized tampering such as paraphrasing, but less robust against global tampering: translation, rewording with another model, or asking the model to insert a special character between every word and then deleting it. It added that the method could stigmatize AI as a writing aid for non-native English speakers.

The company's current help center article on provenance signals in its generated content, read on November 2, 2026, says something different for some users. It lists the signals by content type. For text, it says ChatGPT-generated text in the EU includes an invisible change to the randomness used in the model's word choices, called a textGrain watermark, to comply with the EU AI Act, and that API customers anywhere can turn text watermarks on in their settings.

• No hidden characters. The help article says the watermark is part of the wording and does not add hidden characters, invisible spaces or unusual punctuation.
• Short text is hard. It says short passages usually do not contain enough text for reliable detection, and notes that the EU code of practice does not require watermarks in outputs shorter than 200 tokens, about 150 words in English.
• Heavy edits can erase it. Substantial paraphrasing or translation can make the watermark undetectable, the article says.
• Detection varies by language. In the company's test across the 24 official EU languages, using 500 synthetic English prompts translated into each, the detection rate at a 1% false positive rate ranged from 69.0% for Spanish to 42.2% for Romanian.

Two more points matter for students and teachers. Text detector access is limited to qualifying groups such as academic researchers studying provenance, and academic integrity uses, such as checking student essays, are currently outside the program's scope. And a detected watermark shows that one of the company's models likely generated or processed the text, but not how much of it, or who the author was.

Watermarks are not the same as AI detectors

An AI detector such as Turnitin's guesses from style and statistics whether text looks machine-written, and works on text from any model. A watermark detector looks for one specific planted pattern, works only on text from a model that planted it, and needs the key. That is why a watermark can have a very low false positive rate while a style detector cannot, and why a teacher's detector score is not a watermark check. For how style detectors report results, see how Turnitin's AI indicator works and what AI text detectors can tell you.

How text watermarks differ from image provenance labels

Images are handled differently. The same help article says images generated with ChatGPT carry Content Credentials, built on the open C2PA standard, plus an invisible SynthID watermark. Content Credentials are metadata: a signed record of which tool made the file and how it was edited. The article notes they can be lost through a screenshot or a file conversion, which is why the watermark is added as a second signal. A pasted paragraph carries no file header, so for text only the word-choice pattern can carry a signal. For reading credentials on an image, see how to read Content Credentials on images.

Why invisible characters are not a watermark

Text copied from a chat window, a web page or a word processor can carry non-breaking spaces, zero-width characters or curly quotes. They come from formatting and copying, and the help article states that the company's text watermark adds no hidden characters, invisible spaces or unusual punctuation. Finding one proves nothing about where the words came from, and removing one changes nothing about a real watermark. Punctuation habits are no better; the guide to dashes in AI writing explains why one mark is not evidence either.

A worked example: what a watermark check can and cannot say

Here is an example built for this page. A researcher with detector access checks three passages. Passage one is a 600-word chatbot answer pasted unchanged: the detector reports a watermark, which, per the help article, means one of the company's models likely generated or processed it. Passage two is the same answer after a student rewrote most sentences and translated a quotation: the signal may be too weak to find, and a miss does not mean a person wrote it. Passage three is an 80-word reply: too short for a reliable result either way.

None of the three results says who wrote the text, how much the student contributed, or whether the use was allowed. Drafts, notes and the course policy answer those.

1. Ask which signal is meant

   A watermark, a style detector score and image metadata are different things.

2. Check who can run the test

   Text watermark detection needs the key; the company says it does not offer it for checking student work.

3. Look at the process instead

   Drafts, version history and source notes show how text was written; no signal does.

Pitfalls and myths

• "Delete the hidden characters and the watermark is gone." The company says its text watermark has no hidden characters.
• "No watermark found means a human wrote it." Edits, translation, short length and non-EU output all leave nothing to find.
• "My teacher can check the watermark." The company says academic integrity checks are outside its detector program's scope.

Tips for readers and teachers

• Quote the source, not the rumor. Watermark policies change; cite the company's own page and the date you read it.
• Disclose instead of worrying. If your course allows AI help, a short disclosure answers what no signal can. More in the AI writing guides.

So is there a ChatGPT watermark? For EU users' text and for images, the company says yes; elsewhere, its own pages describe none for ChatGPT text. Either way, a text watermark is a hidden pattern in word choice that only a key holder can test, it adds no strange characters, and it cannot say who wrote what or whether the use was allowed. For those questions, drafts, notes and an honest disclosure still do the work.