Open a ChatGPT reply in Berlin next month, and the words on your screen may carry a secret. Not a visible stamp or a disclaimer, but a statistical pattern woven into the way sentences are built, something no human eye could ever catch but a machine, with the right key, could read instantly. OpenAI calls the system textGrain, and it is about to become a quiet feature of life for ChatGPT and Codex users across the EU, before anyone else in the world gets it. The timing is not an accident.

The European Union's AI Act is pushing companies toward more transparency about machine-generated content, and OpenAI is far from alone in responding. Anthropic announced a similar watermark in August, and both companies are building on an approach pioneered by Google DeepMind, called SynthID. OpenAI says its own version "matched or exceeded" SynthID's performance in testing, while producing text that scores just as well on standard AI benchmarks as unwatermarked output. In other words, users in Paris or Warsaw are unlikely to notice any difference in quality, only an invisible fingerprint tagging along underneath.

Yet OpenAI is notably cautious about what this fingerprint can actually prove. The company admits textGrain "does not guarantee reliable detection," and insists it cannot confirm who owns a piece of text, how much of it a person actually wrote, or whether any claim within it is true. For now, detection tools will be shared only with vetted researchers and institutions, not the public, partly because false positives remain a real risk. Beyond Europe, the rollout looks different: API customers worldwide can opt in voluntarily, letting businesses decide for themselves how to handle transparency.