Anthropic's watermark rollout leaves a loophole on two older Claude models

The watermark on older Anthropic models can be bypassed by tweaking sampling parameters, leaving older model outputs potentially unmarked.

Anthropic is rolling out watermarking on some of its Claude AI models to comply with the EU AI Act's December deadline. The watermarking, an imperceptible pattern in word choice, will be applied to Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6. Newer models like Claude Opus 5.5 already have this feature. Anthropic states the watermark does not alter the meaning, quality, readability, pricing, or request/response formats, and minimally impacts output speed. It also contains no user-specific information. Notably, on Claude Opus 4.6 and Claude Sonnet 4.6, the watermark is only active when sampling parameters like temperature, top_p, and top_k are left at their default settings. A slight adjustment to any of these parameters can bypass the watermark. The company offers a detection API for verification and is embedding the mark at the model level, making it platform-agnostic across AWS, Google Cloud, and Microsoft Azure, and it cannot be turned off by users.

Anthropic's decision to retrofit older models like Claude Opus 4.6, which remains a high-performing option, with watermarking indicates a broader industry push to meet regulatory demands, even if imperfectly. The ability to bypass the watermark on these specific models by altering sampling parameters highlights a potential loophole for those seeking to obscure AI-generated content, despite Anthropic's compliance efforts. This selective application and bypassability may lead to inconsistent detection rates and create challenges for downstream applications that rely on watermarks for content provenance.

The rollout strategy, especially the loophole on older models, suggests that achieving robust and universal AI content attribution remains a significant technical and policy challenge. As regulations like the EU AI Act come into effect, companies are balancing compliance with user flexibility and technical feasibility. The effectiveness of watermarking will likely be tested by users actively seeking to circumvent it, potentially driving further innovation in both watermarking and detection technologies, or leading to alternative methods for AI content identification.