OpenAI’s textGrain watermark fails EU tests on short or edited text
The EU’s AI Act requires AI-generated text to be watermarked for transparency, but OpenAI’s new textGrain system—designed to comply—struggles in real-world use. Tests show detection rates drop sharply for texts under 150 words or after editing, with synonym replacement cutting accuracy from 92% to 17%. OpenAI claims textGrain outperforms Google’s SynthID Text in its own tests, but those results…
Key points
- OpenAI’s textGrain watermark detects AI-generated text at 95% accuracy for 400-token passages but drops to 80% for 200 tokens
- Editing text reduces detection rates: replacing 10% of words cuts accuracy from 92% to 66%
- EU AI Act lacks measurable thresholds for watermark effectiveness, leaving compliance uncertain
AI Text Watermarking May Not Be Able to Meet the EU’s New Rules
IEEE Spectrum AI · 5 November 2026
Loading the full article…
This text was published by IEEE Spectrum AI and written by Anna Desmarais. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Policy & Regulation
All →- Microsoft patents AI that could edit live speech and match lips · 2 src
- Republicans ask Nvidia CEO to brief Congress on alleged Chinese spy ties · 1 src
- Google says schools must accept AI features to use Search starting next year · 39 src
- SpaceX shares drop 7% amid AI and tech market sell-off after Democratic midterm wins · 8 src
- OpenAI pledges to avoid NDAs and cover electricity costs for data centers after election backlash · 5 src
Comments
via GitHub Discussions