OpenAI released the textGrain watermark in October to comply with the new EU rule on labeling AI generated text. Tests show the mark often disappears in passages under 150 words or after editing. The false positive rate stands at one percent.
The European Union now requires AI generated text to carry a detectable watermark. OpenAI built textGrain for this purpose, released in October. The system treats entropy, a measure of how unpredictable a models word choices are, as a limited resource that developers can allocate to embed a statistical pattern. Tests show that pattern often cannot be detected once a passage drops below 150 words or gets edited. The detectors false positive rate is one percent.
For companies that deploy or offer text generators, this creates uncertainty about legal compliance. There is still no binding technical benchmark for how reliable a watermark must be to satisfy the EU rule. Relying solely on a vendors assurance carries risk, since a later review could rule the labeling insufficient, with possible fines and costly rework of systems.
A positive match also does not show how much a human contributed to a passage, and a negative result does not prove a text was written without AI help. It remains open what minimum standard the European Commission will eventually set, and whether OpenAI or rivals will need to refine their methods before penalties apply.
What this means for decision-makers
- Check whether your text generation vendor still marks short or edited passages reliably.
- Ask legal counsel what proof the EU rule actually requires for compliant labeling.
- Monitor whether the European Commission sets a minimum technical standard before funding any system changes.
This story was produced automatically from the source named above and checked by software before publication. The image is symbolic and shows neither the event nor a real person. How this paper is made