OpenAI Stamps ChatGPT Text With an Invisible Watermark: EU Only, and Swapping a Quarter of the Words Erases It
On October 5 OpenAI announced a statistical watermark for text generated by ChatGPT and Codex. It is switched on only in the EU, and OpenAI itself opened with a disclaimer: the technology has significant limitations.
Three key facts
The watermark is called textGrain. During generation it uses a secret key to nudge the statistical pattern of word choice. No hidden characters, no invisible spaces. Readers see nothing; OpenAI's detector reads the signal. All ChatGPT and Codex plans inside the EU get it by default over the coming weeks. API customers worldwide can turn on a Text provenance setting, but it stays off by default. The detector is not public. First access goes to research groups such as Cornell, ETH Zurich and KInIT, plus regulators, law enforcement and fact-checkers.
Detection depends on length: roughly 95% for a 400-token passage, about 80% at 200 tokens, and around 60% for math. Editing collapses it. Replacing 10% of words with synonyms drops detection from 92% to 66%. Replacing 25% leaves 17%. Languages vary widely too: Spanish 69%, Romanian 42.2%. The false-positive rate is tuned to 1%. OpenAI's own words: it is expanding provenance to text in response to EU regulatory requirements, while recognizing the significant limitations of current text watermarking technology.
The driver is Article 50 of the EU AI Act, in force since August 2, 2026, which requires machine-readable marking of generative AI output. Systems already on the market before that date must comply by December 2. The technical report was co-written by five OpenAI researchers and four academics from the University of Pennsylvania and Yale, and OpenAI says it plans to open-source the method.
WangDou's Take
Ninety-five percent sounds great until you read the next line: swap a quarter of the words and you are down to 17%. In plain language, this lock stops honest people and nobody else. A high-schooler who wants to hand in AI homework runs the essay through a translator and back, and the stamp is gone. The person who used ChatGPT to fix a few typos gets stamped good and hard. Then look at the map: on by default in the EU, off by default on the API everywhere else. The watermark goes exactly as far as the law pushes it and not an inch further. Romanian at 42% is the awkward part. Half of the EU's own languages slip through, so this is a lock that only closes halfway, and the half it closes is the rule-following half. The detector goes to institutions only. Teachers can't catch homework, but regulators can check journalists. Ask yourself who this stamp was really made for.
Source: OpenAI · Our approach to EU text provenance rules · ActuIA · Interesting Engineering