OpenAI is quietly switching on invisible text watermarks in ChatGPT and Codex, starting only with users in the European Union.
The new system, called textGrain, embeds a machine-readable signal into generated text that humans can't see. OpenAI says its internal benchmarks show textGrain matched or beat other watermarking approaches, including Google DeepMind's SynthID for text, while keeping output quality close to unwatermarked text. The company published benchmark scores comparing watermarked and unwatermarked outputs to back up that claim. Even so, OpenAI concedes the system does not guarantee reliable detection in every case.
The EU-first rollout isn't a coincidence: both OpenAI and Anthropic are watermarking text to satisfy transparency requirements in the EU's AI Act, not because either company volunteered the feature. Anthropic got there first in August, building its tool on top of Google DeepMind's existing SynthID rather than developing its own detector from scratch. OpenAI went the other direction, engineering textGrain in-house and using SynthID mainly as the yardstick to prove its homegrown approach measures up.
That's a meaningful difference in approach even though the regulatory motive is identical: one lab built on an established system, the other built and benchmarked its own. Whether textGrain's self-reported numbers hold up outside OpenAI's own tests remains to be seen - and so does whether anyone outside Europe gets the feature at all.