OpenAI to Begin Watermarking ChatGPT Text in the European Union
OpenAI is about to start embedding an invisible watermark in text produced by ChatGPT and Codex for users in the European Union. The company announced the change on Monday in a blog post, saying it is needed to comply with the EU AI Act.
The reason is regulatory. The Act’s transparency rules took effect on 2 August and require AI companies to mark AI-generated content in a way that other systems can recognise.
Who gets it, and when?
The rollout will happen over the coming weeks. It covers eligible ChatGPT and Codex users on every plan, but only inside the EU.
Developers using OpenAI’s API, in any country, can switch the watermark on for selected models starting now. It is off by default. OpenAI also made clear that it is not making text watermarking a global default at launch.
How an invisible watermark works?
There is no visible mark or symbol. Instead, the system quietly nudges which words the model picks as it writes. The result is a statistical pattern that a reader cannot see but a detector can find.
Because the pattern sits in the wording itself, it stays with the text when someone copies and pastes it. OpenAI says the watermark does not identify the user, and that it saw no meaningful drop in model performance with it turned on.
The method behind it
Alongside the announcement, OpenAI released a technical report on its method, called textGrain. It was co-written with researchers from the University of Pennsylvania and Yale. The report describes using a secret key to sort the model’s next-word predictions. One such nudge means little, but across hundreds of them a detector holding the key can flag AI-generated text from the text alone.
The limits OpenAI admits
The watermark can be weakened by editing. In one test, swapping 10% of the words for synonyms cut detection from about 92% to 66%. Short passages, maths answers and translated text are also harder to detect.
For that reason, OpenAI says it will give detector access at first only to approved researchers and expert organisations. They will help test how reliable it is and how it should be used responsibly.
The company also warns against reading too much into a missing watermark. It “does not prove human authorship,” OpenAI said. The text may be too short, heavily edited, or produced by another company’s AI. Even when a watermark is present, it only shows that an OpenAI system produced or handled part of a passage. It says nothing about how much human judgement or creativity went into the final text.
Where this sits in the wider industry?
The move comes two months after Anthropic said it would watermark text generated by Claude, and Anthropic is applying that worldwide. That decision drew criticism from some Claude users. They argued that they supplied the instructions, context and decisions, and that Claude was simply a tool.
OpenAI had built a text watermark before but held it back. The Wall Street Journal reported in 2024 that one concern was users switching to rivals that did not watermark.
Anthropic, Google, Meta, Microsoft and OpenAI are among the companies that have committed to the EU’s code of practice on AI-generated content.
What to watch next?
Three things will matter. The first is how reliable detection proves to be once approved researchers start testing it. The second is whether OpenAI extends watermarking beyond the EU. The third is whether students, workers and publishers push back, as some Claude users did.
