To meet the EU AI Act, OpenAI is rolling out a phased, opt‑in watermarking feature that embeds unseen markers in the text generated by ChatGPT and Codex for European customers.
OpenAI is introducing an optional, invisible watermarking system for text generated by ChatGPT and Codex, aimed at complying with the European Union’s AI Act.
Why watermarking is needed
The EU AI Act requires high‑risk AI systems to be traceable, meaning developers must provide mechanisms that allow regulators and users to identify machine‑generated content. OpenAI’s new feature embeds a cryptographic marker in the output without altering its appearance.
How the feature works
When the opt‑in is enabled, the model adds a subtle pattern of token selections that can be detected by a verification tool. The watermark is invisible to readers but can be extracted by OpenAI’s internal scanner or third‑party utilities that understand the encoding scheme.
The system is phased: early rollout targets European enterprise accounts, followed by broader availability to individual users who choose to activate it in their settings.
User control and privacy
Customers can toggle the watermark on or off at any time. OpenAI assures that the marker does not store personal data and is not used for profiling. The company also states that the watermark does not affect the quality or speed of the generated text.
Potential impact on developers
Developers building on the OpenAI API will receive an additional header indicating whether a response includes a watermark. This allows downstream applications to respect compliance requirements or to flag content for review.
- Enhanced traceability for regulators
- Voluntary adoption for EU customers
- No visible changes to output
- Easy toggling via API or UI
OpenAI plans to monitor the effectiveness of the watermark and may adjust the algorithm as the EU AI Act evolves.