OpenAI will begin adding an invisible watermark to text produced by ChatGPT and Codex for users in the European Union, a step it ties to the transparency obligations of the EU AI Act, TechCrunch reported Monday. The company's own test results, detailed by BleepingComputer, show the mark can be weakened substantially with modest edits.

What Happened

The rollout will reach eligible EU users on all ChatGPT and Codex plans over the coming weeks. It is not a worldwide default. Developers using OpenAI's API anywhere can switch the watermark on for select models starting Monday, but it is off unless they enable it.

The technique is called textGrain. Rather than inserting hidden characters, it nudges which words the model picks, using a secret key to sort the model's candidate next words. The result is a statistical pattern that a reader cannot see and that survives copying, but that a detector holding the key can find. TechCrunch reported that the accompanying technical report was written with researchers from the University of Pennsylvania and Yale.

RelatedEight States Sue OpenAI as Iowa Seeks Court Halt on Testing of Advanced AI Models

OpenAI said the watermark does not identify the user, according to 9to5Mac, and BleepingComputer reported it also does not reveal account details, the prompt, conversation history or how much a person edited the text. OpenAI said it found no meaningful difference in overall model performance with the watermark applied, 9to5Mac reported.

Why It Matters

The AI Act's transparency rules took effect Aug. 2 and require AI-generated content to be marked in a way machines can detect. Text is the hardest medium to label, because there is no file metadata to carry a tag once words are pasted elsewhere. OpenAI's answer is to hide the label in the wording itself, and to switch it on only where the law requires it.

The company was candid about the limits. In its tests, the detector flagged about 92% of unedited watermarked samples; replacing 10% of the words with synonyms lowered that to 66%, and replacing 25% brought it down to 17%, BleepingComputer reported. Length matters too: at a 1% false-positive rate, detection was around 80% for 200-token answers and about 95% for 400-token ones. Math answers performed worse because, as OpenAI put it, the model has "less freedom to choose different words." Translated, short or heavily edited text may not be detected at all.

OpenAI also warned that the absence of a watermark does not prove a human wrote something, a point aimed at schools and employers that might treat a clean result as an all-clear.

Context

OpenAI has been slow to ship text watermarking. The Wall Street Journal reported in 2024 that the company had built such a system but held it back, partly over concern that users would move to rivals that did not mark their output. Anthropic said in August it would watermark text from its Claude models worldwide, a move that drew complaints from some users who argued they supplied the instructions and decisions behind the output, according to TechCrunch.

What to Watch

Access to the detector will initially be limited to approved researchers and expert organizations, and OpenAI is taking applications. 9to5Mac reported the company plans to open-source the technology. Key questions are whether independent researchers confirm OpenAI's accuracy figures, whether EU regulators accept a mark that light editing can erode, and whether OpenAI extends the watermark beyond the EU or makes it a default for API customers.