OpenAI will watermark ChatGPT outputs by default—but only in the EU

OpenAI said it will begin automatically embedding a watermark into text generated by ChatGPT for users in the European Union, a change the company ties directly to compliance with the EU AI Act that went into force in August. The firm will offer the same watermarking feature in other jurisdictions, but it will not be turned on by default outside the EU.

OpenAI describes its system as textGrain and has published a technical paper outlining its approach. Like other watermarking schemes for large language models, the technique alters word-choice patterns in ways that are intended to be imperceptible to human readers but detectable by a tool with the correct key or detection algorithm. OpenAI plans to grant detector access to a limited group of researchers and organizations and says it will accept requests for approval over time.

Experts and existing standards suggest such watermarking is not a foolproof identifier. Initiatives like SynthID and the C2PA project already provide ways to mark AI outputs, but they can be defeated by users with modest technical skills. OpenAI's watermark is proprietary, and the company acknowledges the broader limitation that no current method offers full reliability against deliberate circumvention.

The announcement illustrates how companies are adjusting product behavior to meet new regulatory demands while balancing transparency and access to detection tools. It also highlights an ongoing technical and policy debate about how—or whether—AI-generated content can be robustly distinguished from human-authored material when motivated actors seek to remove or obfuscate such markers.