OpenAI to auto-watermark ChatGPT in EU with textGrain

OpenAI's decision to turn on automatic watermarking by default only in the European Union reflects the bloc's new rulebook. The EU AI Act took effect in August and requires marking AI-generated content so another tool can detect it. Outside Europe, the same feature will be available but switched off. That split suggests the company is treating the requirement as a regional compliance matter rather than a global default.

The method itself, called textGrain, belongs to a family of watermarking approaches that hide patterns in word choices. Those patterns are not obvious to a human reader and do not meaningfully change output quality, but someone holding a key can use a specialized detector to find them. Existing standards like SynthID and the C2PA project already work on similar principles, and the article notes they are relatively easy to circumvent for anyone with basic know-how. OpenAI says it will give detector access to a limited number of researchers and organizations, with an approval process for others over time.

How much this actually changes things may hinge on two things: whether the watermark holds up against attempts to remove it, and how widely the detector is shared. For EU-based compliance and trust-and-safety teams, an automatic marker offers a starting point, but one whose reliability the article does not vouch for. The test is likely to come as soon as researchers and organizations begin probing textGrain in practice.