Watermarks Without Verification: AI Text Watermarking After the EU AI Act (arxiv.org)

On August 2, 2026, the European Union's AI Act mandates that generative AI providers must mark AI-generated content for easy detection. Following this, Anthropic announced that their Claude models would incorporate a default watermark based on SynthID-Text, further emphasizing transparency in AI-generated outputs. However, user feedback includes concerns about the watermark affecting the quality of generated content, especially code, while also raising issues about the potential encoding of identifying information and the watermark's detectability. Anthropic has countered these claims, asserting that both quality and security remain intact. This situation highlights a critical governance challenge within the AI/ML community: the inability to verify the effectiveness and claims surrounding AI watermarks. Research conducted on open-weight models has indicated that while the watermark does not significantly degrade text quality, it can impact code correctness. The team's analysis suggests that current tools for testing watermarks are inadequate, underscoring a broader need for transparency measures. The study calls for crucial steps like matched output releases, configuration disclosures, and accredited audits to enhance verification processes and ensure accountability within generative AI systems.