Research Summary: Watermarks Without Verification: AI Text Watermarking After the EU AI Act

The implementation of Article 50 of the EU AI Act on August 2, 2026, mandated generative AI providers to mark AI-generated content, ensuring its detectability. Following this, Anthropic embedded watermarks, based on SynthID-Text, into its Claude models, enabled by default. Google had already deployed similar technology in Gemini since 2024. This move has been met with user objections concerning quality degradation, particularly for code, potential encoding of identifying information, and conflicting views on watermark removability and inescapability, prompting vendor assurances regarding quality and robustness.

Why it matters

This development underscores a significant regulatory shift towards transparency and accountability in AI content generation, impacting compliance strategies for AI providers globally. The tension between regulatory mandates, technological implementation, and user experience highlights critical challenges in deploying AI at scale, influencing future product development and policy frameworks for responsible AI.

Key insights

• The EU AI Act's Article 50, effective August 2, 2026, requires generative AI providers to mark AI-generated content for detectability.
• Anthropic implemented watermarking in all Claude models released post-August 2, 2026, using SynthID-Text, with no user opt-out.
• Google had previously deployed SynthID-Text in its Gemini models since 2024.
• Users expressed concerns about watermark-induced quality degradation, especially for code.
• User objections also included the possibility of watermarks secretly encoding identifying information.
• Conflicting user feedback suggested watermarks were both easily removable and inescapable.
• Vendors, such as Anthropic, responded by assuring unchanged quality, absence of identifying information, and robustness against minor alterations.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2609.09604