Does ChatGPT watermark its text?

The short answer

As of the dates on the sources below: OpenAI has built a text watermarking method and has chosen not to release it. The company says so itself, and gives its reasons in writing. Google, by contrast, does watermark text in its Gemini chatbot, using a system called SynthID-Text that was published in Nature in October 2024.

So if someone claims a document "was checked for the ChatGPT watermark" — there is, per OpenAI's own public statements, no released ChatGPT text watermark to check for. And if a document has no watermark at all, that proves nothing in either direction, for reasons both companies state themselves.

What a text watermark is

A text watermark is embedded while the AI writes, not detected afterward. The model's word choices are nudged in a secret pattern — Google describes its approach as "modulating the likelihood of tokens being generated" — that reads naturally to a human but can be verified statistically by someone who knows the pattern. Unlike the AI "detectors" sold to schools, a watermark check has real statistical guarantees, because the evidence was planted on purpose.

That is the strength. The weakness is that the watermark lives in the exact word sequence the model produced. Change the words, lose the mark. Our research review of the watermarking literature goes through the papers; this page is the plain-language version.

OpenAI's own words

In its page "Understanding the source of what we see and hear online," OpenAI describes its text watermarking work as:

"highly accurate and even effective against localized tampering, such as paraphrasing"

— but, continuing in the same sentence, it is:

"less robust against globalized tampering; like using translation systems, rewording with another generative model, or asking the model to insert a special character in between every word and then deleting that character - making it trivial to circumvention by bad actors."

Read that last example again: ask the model to put a marker between every word, then delete the markers. That is the entire attack. No skill required.

OpenAI then gives two reasons for not deploying, both quoted verbatim:

It "could stigmatize use of AI as a useful writing tool for non-native English speakers."

"applying it to large volumes of text would lead to a large number of total false positives."

Those two sentences matter beyond watermarking. They are the maker of ChatGPT saying, in writing, that AI-writing identification at scale produces false accusations and falls hardest on non-native speakers — the same findings the independent research reached about detector accuracy and detector bias.

Worth remembering alongside this: OpenAI once did ship an AI-text detector — the "AI Text Classifier" — and withdrew it in July 2023 with the note, quoted by The Register: "As of July 20, 2023, the AI classifier is no longer available due to its low rate of accuracy."

What Google shipped, and what Google admits

Google is the counterexample: SynthID-Text runs in production. The Nature paper says it was "productionized in the user-facing Gemini and Gemini Advanced chatbots," and DeepMind's blog confirms "watermarking AI-generated text in the Gemini app and web experience."

But Google's own caveats, verbatim from its blog, are the ones to keep:

SynthID's "confidence scores can be greatly reduced when an AI-generated text is thoroughly rewritten or translated to another language."

It is "less effective on responses to factual prompts."

And from the Nature paper itself: the watermark is "weakened by edits to the text, such as through LLM paraphrasing," and — the sentence that should hang over every detection product — "No text detection method is foolproof."

So what does "no watermark found" mean?

Nothing. A document with no detectable watermark could be:

• written by a human;
• written by ChatGPT (no released watermark to find);
• written by Gemini and then edited, paraphrased, or translated (watermark weakened or gone, per Google);
• written by any open-weights model running on someone's laptop (nothing stops the user from sampling without a watermark — the Nature authors call open-source enforcement "difficult").

And a found watermark is strong evidence about one thing only: that a cooperating model produced that exact token sequence. It says nothing about who directed the work, how much of the document it covers, or whether the rest was human.

What to rely on instead

The honest summary of the vendor statements above: the companies that build the models either declined to ship identification technology because of its error profile (OpenAI) or shipped it with written warnings that rewriting defeats it (Google). Neither offers what an integrity decision needs — evidence about how a document was actually made.

That evidence has to be created during the writing, not inferred from the text afterward. A writing record that captures the process as it happens — and can be sealed and checked later — does not care how the final sentences read, which is exactly why paraphrasing cannot fake it and plain prose cannot false-positive it. Start with what writing provenance is, or see how to prove you didn't use AI if you are dealing with an accusation right now.

Related reading: LLM watermarking — where the research stands · how AI detectors work — and why they break · what your Turnitin AI score actually means