AI Text Watermarking May Not Be Able to Meet the EU’s New Rules
This highlights the fundamental technical limitations of AI text watermarking and the compliance risks AI providers face when regulations mandate capabilities that current technology cannot reliably deliver.
Reported by IEEE Spectrum ·
What happened
Tests on OpenAI's new textGrain watermarking system, developed to comply with the EU AI Act, show that detection rates drop sharply when AI-generated text is under 150 words or undergoes minor editing. Because the EU regulation lacks clear technical benchmarks for robustness, it remains uncertain whether the tool will satisfy the new legal transparency obligations.
Who it’s about
- OpenAI
- textGrain
- Google DeepMind
- SynthID
Source
More on policy
Get a brief picked for you
Sign in to follow topics and companies, save stories and search the archive.
Sign in