OpenAI announced it will start embedding an invisible watermark into text generated by ChatGPT and Codex for users in the European Union, as part of efforts to comply with the EU AI Act. The company disclosed the move in a blog post published on Monday.
Why this change?
The EU AI Act’s transparency provisions, which took effect on August 2, require that content produced by AI systems be marked in a way that other systems can detect. OpenAI says adding a watermark is intended to satisfy that requirement.
Rollout details
- OpenAI said the watermark will be rolled out over the coming weeks to eligible ChatGPT and Codex users in the EU across all plans.
- Developers using OpenAI’s API anywhere in the world can enable the watermark for selected models starting today; it is off by default.
- The company is not making text watermarking a global default at launch.
How the invisible watermark works
The watermark is not a visible mark or separate metadata tag. Instead, it subtly shifts the model’s word choices to create a pattern that is imperceptible to human readers but detectable by a dedicated detector. Because the signal is embedded in the words themselves, it persists when text is copied and pasted. OpenAI says the watermark does not identify individual users, and that it observed no meaningful degradation in model performance when the watermark was active.
The technical approach: textGrain
OpenAI published a technical report describing the method, called textGrain. The paper was co-authored with researchers from the University of Pennsylvania and Yale. The approach uses a secret key to reorder next-word predictions; by applying hundreds of small nudges to token selection, a detector that has the key can distinguish AI-generated content using only the text and the key.
Resilience to editing and known limitations
OpenAI’s tests indicate the watermark can be weakened or obscured by editing. In one experiment, replacing 10% of words with synonyms reduced detection rates from about 92% to 66%. The company also noted that short passages, mathematical answers, and translated text are more difficult to detect.
Access to detectors and caveats
Because of these limitations, OpenAI is initially providing detector access only to approved researchers and expert organizations to help evaluate reliability and responsible uses. The company cautioned that an absent watermark does not prove human authorship—texts can be too short or heavily edited, or may originate from another provider’s AI system.
Market context
The announcement follows Anthropic’s decision two months earlier to watermark text generated by its Claude model, a policy Anthropic is applying globally. Major industry players including Google, Meta, Microsoft and OpenAI have committed to the EU’s code of practice on AI-generated content.
Summary
OpenAI’s invisible watermark aims to meet EU regulatory requirements, but the company acknowledges practical limitations and is restricting detector access initially. The technical details are laid out in the textGrain report, while real-world effectiveness will depend on edits, text length and the nature of the content.
