Anthropic to Watermark Claude AI Text for EU Users Under New Act

Image credit: Search Engine Journal

Anthropic announced plans to watermark text generated by its Claude artificial intelligence models for European Union users starting August 2, aligning with new EU AI Act regulations.

The move aims to comply with Article 50 of the EU AI Act, which mandates machine-readable marking for generative AI outputs, but it has prompted concerns among writers and developers regarding implementation and potential for circumvention.

Article 50(2) of the EU AI Act requires providers of generative AI systems to ensure that text, audio, images, and video generated by their models are clearly identifiable as artificially produced. These markings must be effective, interoperable, robust, and reliable, according to the regulation.

Writers have voiced objections to their edited work being marked as AI-assisted, particularly if an AI model only provided initial drafts or minor contributions, Forbes reported.

Developers also raised questions about the practicality and efficacy of watermarking technologies, citing potential difficulties in their application and detection across diverse content forms, TechCrunch reported.

The marking obligation does not extend to standard editing assistance or instances where the AI system does not substantially alter input data, according to the EU AI Act. Specific exclusions also apply to source code and short text sequences.

Anthropic, alongside major tech companies including Google, Meta, Microsoft, and OpenAI, signed a voluntary Code of Practice on Transparency of AI-generated Content. This framework provides guidelines for compliance ahead of formal regulatory enforcement.

However, the presence of a detected Claude watermark does not definitively establish AI authorship or that Article 50(2) necessitated the marking, according to Anthropic. User editing, translation, and other content changes post-processing can alter or remove such marks.

Researchers have already demonstrated methods to remove or spoof AI watermarks with high success rates and at low cost, raising questions about their long-term reliability. Roger Montti, writing for Search Engine Journal, highlighted these vulnerabilities.

ETH Zurich’s SRI Lab, for instance, has published findings detailing techniques that can effectively bypass current watermarking schemes, underscoring the technical challenges in creating truly indelible marks.

The ongoing debate highlights the complexities of balancing regulatory demands for transparency with the practical realities of AI development and content creation.

Source: Search Engine Journal