Anthropic Explains Claude Watermarking Amid EU AI Act Compliance

Anthropic clarifies watermarking approach after user backlash

Anthropic published a detailed explanation Friday of its plans to watermark text generated by Claude, its AI assistant, following customer concerns that prompted some subscription cancellations earlier this week. The company is implementing the technology to comply with the European Union's AI Act, which requires AI providers to mark synthetic content.

According to Business Insider, which first reported the details, Anthropic will apply watermarking globally at launch because it lacks a reliable method to limit the feature by region. The watermarks will roll out first to new Claude models, then to older versions in coming months. Anthropic told Business Insider it has not observed a trend of increased cancellations since announcing the watermark.

How the watermark technology works

The watermarking system is based on a 2024 Google DeepMind research paper. Rather than embedding hidden characters or special formatting, the watermark operates through word selection patterns. When Claude generates text, it makes numerous decisions about which synonyms or phrases to use—choices often settled by random selection. Anthropic's watermark replaces the standard random number generator with a specialized one that creates detectable patterns.

Anyone with Anthropic's "key"—which the company plans to distribute through an API—can analyze text to determine the likelihood it was generated by Claude. The company emphasized that readers cannot distinguish watermarked text from unmarked content by appearance alone.

Why it matters

The EU AI Act's extraterritorial reach means watermarking requirements designed for European markets will affect AI tools worldwide. As Anthropic noted, other major AI providers including OpenAI will face the same compliance obligations. This represents a significant shift in how AI-generated content is tracked and identified, though the technology's limitations suggest watermarking alone won't solve attribution challenges.

Significant limitations remain

Anthropic's blog post repeatedly highlighted constraints of the watermarking approach. The system can only identify Claude-generated content, not AI text in general. Longer passages are easier to detect because they contain more word choices, while factual content and code receive fewer watermark signals to preserve accuracy.

The company acknowledged that light editing likely won't remove watermarks entirely, but comprehensive rewrites will. Anthropic also clarified that watermark detection indicates Claude processed the content but doesn't affect user ownership rights under its terms of service.

OpenAI updated a support page two weeks ago indicating it also plans to add text watermarks, though the company did not respond to Business Insider's request for comment.

Business Insider first reported these developments.