AI Text Watermarking May Struggle to Meet New EU Transparency Regulations

Tests show that textGrain, a text-watermarking technique developed by OpenAI to comply with EU AI Act transparency rules, suffers a sharp drop in detection rates for short texts and post-edited content, potentially falling short of compliance requirements. In tests, the detection rate reached 80% for passages of around 150 words, but plunged to 17% when 25% of the words were replaced with synonyms. The EU AI Act mandates robust, machine-readable markings for synthetic content; researchers note the method still lacks independent validation and could expose companies to heavy fines if found non-compliant.