OpenAI enables EU ChatGPT text watermarking by default

Regulators have been searching for concrete levers to make AI outputs traceable, and OpenAI’s decision to turn on textGrain watermarking by default in the EU is the clearest sign yet that provenance tools are moving from research to compliance infrastructure. The company is not claiming this watermark is foolproof, but it is accepting product friction and some model capacity for the sake of meeting the AI Act’s detectable content requirements. That is a notable strategic pivot from the days when watermarking was framed as an experiment rather than a default behavior.

For the broader race to AGI, this move shows how frontier labs will increasingly ship features dictated by law rather than by user demand. It also indirectly benefits European regulators, who gain a real-world testbed for provenance mandates while avoiding the hardest technical work themselves. The fact that watermarking remains off by default outside the EU highlights a fragmented future, where global models exhibit region-specific safety and transparency properties. As capabilities climb, such regional forks in behavior and tooling could matter as much as raw benchmark scores when enterprises choose which AGI-adjacent platforms to trust.