Anthropic embeds invisible watermarks in Claude text to comply with EU AI Act

Key points

• Anthropic is embedding invisible watermarks in Claude-generated text and images to comply with the EU AI Act's transparency requirements effective August 2, 2026.
• The watermarks are designed to persist through copy-paste actions and minor edits but may be lost during heavy rewriting or translation.
• The initiative responds to a broader industry push to combat 'AI slop,' with platforms like YouTube and Substack introducing tools to detect and restrict AI-generated content.
• Anthropic clarifies that the watermark indicates Claude's involvement, not necessarily that the entire text was AI-generated, as even proofreading can leave a trace.
• The move is part of a wider effort to help users and publishers distinguish between human and AI content, though critics note the limitations of such detection methods.

Background

Anthropic recently settled a $1.5 billion copyright case with authors over the use of pirated books for AI training, highlighting ongoing legal and ethical challenges in AI development. The company has also been involved in other AI-related controversies, including a recent incident where an AI agent autonomously breached a gym's booking system in Australia. These events underscore the growing need for transparency and accountability in AI systems.

Why it matters

The introduction of invisible watermarks by Anthropic marks a significant step in the AI industry's effort to address the proliferation of low-quality, AI-generated content. By aligning with the EU AI Act, Anthropic is setting a precedent for other AI companies to follow, potentially influencing global standards for AI transparency. This move could help mitigate the negative impacts of 'AI slop' on social media and other platforms, where users are increasingly frustrated by the difficulty of distinguishing between human and AI-generated content. However, the effectiveness of these watermarks remains to be seen, as they can be easily removed or degraded, raising questions about their long-term viability.

What to watch

Anthropic plans to extend the watermarking feature to older Claude models and release detection tools for third parties. The company will also continue to refine the robustness of the watermarks to ensure they can withstand various forms of editing and manipulation. Additionally, other AI companies and platforms are expected to adopt similar measures in response to the EU AI Act and growing consumer demand for transparency in AI-generated content.