OpenAI to add invisible watermark to ChatGPT text in the EU

OpenAI announced on Oct. 5 that it will put an invisible watermark on written output from ChatGPT and Codex for eligible users in the European Union, "in the coming weeks," according to OpenAI. It will cover all plans, but only inside the EU.

For people using ChatGPT in Mexico or the United States, nothing changes for now: the watermark applies only in Europe and readers cannot see it. In OpenAI's API, the tool for developers, any organization in the world can switch it on starting Oct. 5, but it ships turned off and is enabled under "Text provenance" in the settings, according to ActuIA.

The system is called textGrain and does not hide characters in the text. It statistically adjusts which words the model picks, and a detector holding the secret key can later recognize the pattern. The method is described in a 20-page report, "textGrain: Entropy-Calibrated Watermarking for Language Model Text," with authors from OpenAI, the University of Pennsylvania and Yale, according to Unite.AI. The report does not appear on arXiv under that title.

How well it detects

The figures come from OpenAI's own report, and no independent evaluation has confirmed them. At a 1% false positive rate:

• The watermark is detected in about 80% of 200-token passages (a token is a word fragment; 200 is roughly a short paragraph) and about 95% of 400-token passages.
• With synonyms swapped into 10% of the words, the rate falls from 92% to 66%. At 25% replaced, it drops to 17%.
• In math it is around 60% at 400 tokens. Across EU languages it runs from 42.2% (Romanian) to 69.0% (Spanish), according to ActuIA, a figure that could not be checked against the original.

OpenAI acknowledges that substantial paraphrasing or translation can make the watermark undetectable. It also says the mark does not prove authorship, does not measure how much a person contributed and does not verify accuracy: it only indicates whether its systems contributed to the text.

Who can detect it

For now, almost no one outside a small circle. The detector is limited to approved researchers and expert organizations, among them Cornell, ETH Zurich and KInIT, which must apply for access. A teacher, a newsroom or a company doubting who wrote a text cannot use this tool yet.

Behind it is a legal obligation. Since Aug. 2, 2026, the EU's AI Act (Regulation 2024/1689) asks in its Article 50 that synthetic content be marked in a machine-readable way. Systems already on the market before then, such as ChatGPT, have until Dec. 2 to comply.

OpenAI has not given an exact date for ChatGPT in Europe. According to ActuIA, it plans to release textGrain as open source, also without a date.