OpenAI will begin adding an invisible watermark to text generated by ChatGPT and Codex for users in the European Union to comply with the EU AI Act, the company said Monday.
The watermarking system will be rolled out over the coming weeks to eligible ChatGPT and Codex users across all plans in the EU.
OpenAI said developers using its API anywhere in the world can also enable the feature for select models from Monday, although it will remain switched off by default.
The company said it is not making text watermarking a global default at launch.
How will the watermark work?
The watermark will not appear as a visible symbol in AI-generated text. Instead, OpenAI’s system will subtly influence the model’s word choices to create a hidden pattern that can be detected by specialised systems.
Because the pattern is embedded within the text itself, OpenAI said it should remain detectable when the content is copied and pasted.
The company said the watermark does not identify the person who generated the text and that testing showed no meaningful impact on model performance when the feature was enabled.
OpenAI has named the technology textGrain and published a technical report explaining how it works.
The system uses a secret key to influence how the model selects among possible next words. When these small changes are repeated across a passage, a detector can identify the resulting statistical pattern as AI-generated content.
Watermark can be weakened by editing
OpenAI acknowledged that the watermark is not impossible to remove or weaken.
In one company test, replacing around 10% of the words in a passage with synonyms reduced detection accuracy from about 92% to 66%.
OpenAI also said short passages, mathematical answers and translated text can be more difficult for its system to identify.
“These limitations” were among the reasons the company said it would initially provide access to its detector only to approved researchers and expert organisations.
The aim, OpenAI said, is to help assess the technology’s reliability and responsible uses.
A watermark does not prove AI authorship
OpenAI also cautioned that the absence of a watermark should not be interpreted as proof that a person wrote the content.
A passage may be too short, heavily edited or generated by another AI system, the company said.
The watermark can indicate that an OpenAI system generated or processed part of a passage, but cannot establish how much human judgement, editing or creativity was involved.
EU rules drive move
The move follows transparency requirements under the EU AI Act, which require providers of certain AI systems to make AI-generated or manipulated content identifiable in a way that can be detected by systems.
OpenAI is among several major technology companies, including Anthropic, Google, Meta and Microsoft, that have committed to following the EU’s code of practice on transparency for AI-generated content.
Anthropic announced in August that it would introduce watermarking for text generated by its Claude models globally.
