OpenAI announced on October 5 that it will add an invisible watermark to eligible text produced by ChatGPT and Codex in the European Union over the coming weeks. The rollout covers all plans in the EU, and the company is not making text watermarking a global default at launch. Developers who use the API can opt in worldwide for supported models, but the feature is off unless they turn it on.
Why OpenAI is doing it
The move is a response to the EU AI Act, whose transparency rules for AI-generated content took effect on August 2, 2026. The law expects providers to make AI-generated text identifiable by machines, and a watermark is one way to do that without changing what readers see.
How the watermark works
OpenAI calls its watermark textGrain. It is not a visible label or a note at the end of a message. The signal is a statistical pattern in the model’s word choices, so nothing changes when you read, copy or paste the text. A detector can later analyze the wording and report whether it finds that pattern.
OpenAI says the watermark contains nothing that identifies the person, account, prompt or conversation behind the text. It also says the mark does not meaningfully reduce output quality, with benchmark results staying broadly similar when it is enabled.
What it cannot do
OpenAI is unusually open about the limits. In its own tests, the detector found the watermark in about 80% of short responses of around 200 tokens and about 95% of longer ones of around 400 tokens, at a 1% false-positive target. Editing weakens the signal quickly. Replacing or changing even a modest share of the words, such as with synonyms, lowers detection sharply, and the more a text is rewritten or translated, the less reliable the check becomes.
For that reason, OpenAI warns that a missing watermark does not prove a human wrote the text. Short passages, edited text, translations and unsupported models can all prevent detection. The opposite also holds: a detected watermark shows that OpenAI systems generated or processed some of the text, but it cannot say how much of the final work a person wrote or edited.
Who can check for it
At launch, the detector is not a public tool. Researchers and expert organizations can apply for access, initially case by case. OpenAI’s reasoning is that missed detections and false positives could cause real harm if anyone could use the tool, for example by wrongly accusing a student or a writer.
OpenAI also says it intends to release the technology as open source and is working with cloud partners to offer watermarking through their services. Two things remain unclear: which API models are covered, and whether ChatGPT users in the EU will be able to turn the watermark off.
How it compares
OpenAI is not alone. Anthropic announced in August that newer Claude models would watermark the text and files they generate. Google has its own SynthID system, which this week gained a public detector website for images, video and audio. Text is the hardest case for all of them, because short and edited passages carry very little signal.
What this means for you
If you use ChatGPT or Codex in the EU, nothing changes on screen, and the watermark does not carry your identity. Students, writers and developers should be cautious about treating a watermark check as a verdict. A positive result is a clue, and a negative result proves nothing.
If you build products on the API, the opt-in gives you a way to show that content came from OpenAI’s models, which may help with disclosure duties under EU rules. For everyone else, the practical lesson is that machine-readable provenance for text is arriving, but it is early, easy to weaken, and not a reliable way to tell human writing from AI writing.
