OpenAI is quietly marking ChatGPT's text in Europe

On Monday, October 5, OpenAI said in a company blog post that it will add an invisible watermark to eligible text output from ChatGPT and Codex for users in the European Union, rolling out over the coming weeks across all plans. The technology, called textGrain, is OpenAI's answer to the EU AI Act's transparency rules, which require AI-generated text to carry a machine-readable identifier; the compliance deadline is December 2.

Outside Europe, it's a choice, not a default. API customers anywhere in the world can switch textGrain on for select models starting October 5, but OpenAI says it is not making watermarking a global default at launch. The company is also working with cloud partners, including Microsoft Azure, to offer the option for models served through their platforms in the coming weeks.

There's no visible symbol and no hidden characters. During generation, textGrain uses a secret key to nudge the model's word choices in a statistical pattern — hundreds of tiny biases a reader can't see. A detector holding the key can then test a passage for that pattern. Because the signal lives in the words themselves, it survives copy and paste.

OpenAI says the watermark doesn't identify the user or reveal the prompt, and that it saw no meaningful change in model performance with it switched on. Alongside the announcement, OpenAI published a technical report co-written with researchers from the University of Pennsylvania and Yale, and says it plans to release the technology as open source so others can build on it.

OpenAI's own numbers show how fragile the watermark is. Tuned to a 1% false-positive target, the detector found the signal in about 95% of 400-token psychology passages, but only about 80% at 200 tokens. Swap 10% of the words for synonyms and detection falls from about 92% to 66%; swap a quarter and it collapses to roughly 17%. Math answers and translated text are harder to detect because there are fewer ways to phrase them.

That's why you can't get the detector. Access goes only to approved researchers and expert organizations, with OpenAI warning that a missing watermark "does not prove human authorship" — the tool can miss a watermark that's there or find one that isn't. On X, the company put it plainly: "We're expanding our approach to content provenance to include text in response to EU regulatory requirements, while recognizing the significant limitations of current text watermarking technology."

The driver is regulation, not product instinct. The EU AI Act's transparency obligations took effect August 2, 2026, and OpenAI is one of roughly 190 signatories of the transparency code of practice. Anthropic started shipping its own invisible text watermark in August — globally, with no opt-out — and Google DeepMind's SynthID has existed for years. The US has no equivalent federal rule, so OpenAI's new default stops at the EU border.

For readers: this changes nothing if you use ChatGPT in the US. It does mean any text that came through an EU ChatGPT session in the coming weeks may carry a machine-readable fingerprint — one a determined editor can strip with a thesaurus. Treat the watermark as a hint for researchers, not proof of anything.