Anthropic's AI Watermark, Loopholes in Legacy Models

Anthropic has expanded its digital watermarking feature, which embeds into text, to four new Claude models. This measure was taken to meet the December deadline set by the European Union Artificial Intelligence Act, but it has been revealed that in legacy models Claude Opus 4.6 and Sonnet 4.6, developers can bypass the watermark by simply adjusting a single parameter.

Digital watermarking in artificially intelligent generated text is fundamentally a technology designed to make it easier to identify text produced by artificial intelligence. The Artificial Intelligence Act that the European Union is implementing requires that providers implement such identification measures, and Anthropic's current expansion is positioned as a response to that requirement. This is part of a broader trend in which major artificial intelligence companies are progressively implementing compliance measures as regulatory deadlines approach the end of the year.

The problem lies in a technical loophole in two legacy generation models. Simply changing the sampling settings of generated text, the parameters that control how the next word is selected, causes the watermark to stop functioning. In other words, users can exercise a certain degree of control over the presence or absence of the watermark.

Anthropic has responded to this issue only by issuing compliance warnings and has not yet implemented any technical fixes or restrictions. While the company takes the form of fulfilling regulatory accountability through warnings, the loophole itself has not been closed.

What this situation demonstrates is the gap between formal compliance with regulations and substantive compliance. While it is true that the watermarking feature is implemented, if it can be easily bypassed, the original purpose of identifying artificially intelligent generated content cannot be adequately achieved. The extent to which regulatory authorities will require technical robustness remains to be seen and will be shaped by future implementation guidelines and enforcement posture.

Additionally, the difference in compliance levels between legacy and new models cannot be overlooked. Companies often continue using legacy models for cost and compatibility reasons, and the effectiveness of regulatory compliance is meaningful only when uniformly ensured across all models. Whether Anthropic will technically fix the loophole in legacy models going forward can serve as an indicator of substantive compliance posture.

As the European Union Artificial Intelligence Act's December deadline approaches, the robustness and verifiability of watermarking technology implementation, not merely its existence, will continue to be questioned. This case can serve as a specific example of the difficulties in regulatory compliance and may become a reference case for the entire industry.