OpenAI plans to make text watermarking a default feature for ChatGPT users in the European Union, creating another example of regulation directly shaping AI product design.
The supplied source says the move is tied to the EU AI Act, which requires AI-generated content to be marked in a machine-detectable way.
OpenAI’s system, called textGrain, uses subtle patterns in word choice that are intended to be invisible to readers but detectable with a specialized tool.
The company says its best-case tests produced a 92% detection rate.
The weakness is robustness. According to the source, changing 10% of the generated text reduces detection success by almost 30%, while changing 20% can lower it by nearly 75%. Shorter and translated text also performs worse.
That means watermarking may help with compliant, unchanged AI output without solving the broader problem of attribution.
OpenAI plans to provide detector access to a limited number of researchers and organizations, with an approval process for additional access.
Outside the EU, watermarking will be available but not enabled by default. The source says the feature will remain optional in the API.
BTI’s bottom line: the EU AI Act is beginning to create measurable product differences by geography. For AI companies, compliance is becoming part of the product stack. The technical challenge is that watermarking systems can satisfy regulatory requirements without providing perfect real-world detection.