ChatGPT watermark: is your ChatGPT text marked now?

If you use ChatGPT in the United States, your text is not being watermarked by default. In the European Union, OpenAI is adding an invisible watermark called textGrain to ChatGPT and Codex output over the coming weeks, and API developers anywhere can switch it on for select models.

What the watermark can prove is narrower than most people assume, so it's worth knowing how it works before you worry about it.

What OpenAI posted

On October 5, @OpenAI announced it was extending content provenance to text "in response to EU regulatory requirements." The same post made a point of "recognizing the significant limitations of current text watermarking technology." Companies rarely launch a feature and flag its weaknesses in the same sentence. I read it as OpenAI saying this is a compliance step, and that nobody should treat it as an AI detector.

The same day, @weijie444, Weijie Su, one of the University of Pennsylvania researchers behind the technical report, posted the research side. He called textGrain OpenAI's answer to the EU AI Act, said detector access is opening to researchers, and added: "We'll open-source it so the community can build on top."

How textGrain works

According to TechCrunch and BleepingComputer, the watermark lives in the words themselves. A secret key nudges the model's choice of next word, so the text carries a statistical pattern spread across hundreds of tiny choices. You can't see it, and it survives copy and paste because nothing is hidden in metadata.

OpenAI's own numbers, as reported:

• At a 1% false-positive target, it caught about 95% of 400-token passages and about 80% of 200-token ones.
• Swapping 10% of the words for synonyms dropped detection from about 92% to 66%. Swapping 25% dropped it to 17%.
• Math answers, short passages and translated text are harder to catch, and OpenAI says rewriting or translating can remove the mark entirely.
• A detected mark does not reveal the user, the account, the prompt or the conversation.
• OpenAI says model quality stayed broadly the same with it on.

The detector is not public. OpenAI is taking applications, and early access goes to approved researchers and expert organizations.

Is ChatGPT watermarked in the US?

Not by default. ChatGPT and Codex consumer apps outside the EU are unchanged for now. The one exception is the API, where developers worldwide can opt in, and it ships turned off.

That is a different choice from Anthropic's. When Claude started marking its text in August, Anthropic applied it everywhere, which we covered in Claude is watermarking its text. OpenAI went the minimum route: the EU, where the law asks for it, and nowhere else. TechCrunch notes OpenAI built a text watermark years ago and shelved it, partly over worry that users would leave for rivals without one. My guess is that a Europe-only rollout is how it keeps that risk small.

My read

The most honest line in the whole launch is OpenAI's own: a missing watermark "does not prove human authorship." Text can be too short, too edited, translated, or written by another company's model. So textGrain answers one narrow question, whether an OpenAI system produced or processed part of a passage, and only for people who have the detector.

For a small business in Illinois, the practical effect today is close to zero. What I'd watch is where this is heading. Anthropic marks everything, OpenAI now marks Europe, and Su says textGrain will be open-sourced. Within a year or two, a client or a hiring manager may be able to run a tool that checks whether unedited text came from AI.

What this means for a business using AI to write

Keep using AI to draft, but treat raw output as a first draft. If a proposal, bid or client email goes out word for word from a chat window, assume it could one day be traced back. If a person actually edits it, it's more likely to read as yours, which it should anyway.

The harder problem is usually upstream. Most teams don't know where AI already shows up in their writing, who reviews it, or which documents carry their name without a person reading them first. That is exactly what our free process audit looks at: where AI sits in your workflow today, and where a review step would save you trouble later.