Anthropic's AI Watermark Easily Disabled in Older Models

Anthropic announced that it will deploy text watermarking features to four Claude models ahead of the December deadline set by the EU AI Act. Text watermarking is a technology that embeds invisible identification information into text generated by artificial intelligence, enabling the detection and tracking of AI-generated content.

The EU AI Act mandates certain identification measures for AI-generated content, requiring each AI company to comply by the end of the year. Anthropic's deployment plan is driven by a clear motivation to meet this regulatory deadline. The introduction of text watermarking aligns with the regulatory trend demanding transparency in AI-generated content.

However, the deployment has revealed a critical issue. In older-generation models such as Claude Opus 4.6 and Claude Sonnet 4.6, the watermark can be disabled simply by modifying a single API sampling parameter when accessing the model through an API. An API is an interface that allows developers to directly control Claude from their programs, providing greater technical control than end-user interfaces.

Anthropic's response to this loophole was limited to issuing a compliance warning rather than a technical fix. In other words, they chose to address the issue through notices in terms of service and usage policies. This approach, which relies on policy-level deterrence rather than technical blocking, raises questions about its effectiveness.

This issue reveals the limitations of watermarking technology itself. If the watermark can be bypassed through output control settings such as sampling parameters, technical workarounds will continue to exist, raising questions about whether watermarking is sufficient as a regulatory compliance mechanism. Particularly, developer and business usage through APIs offers greater control flexibility than end-user interfaces, presenting a structural challenge to maintaining watermark effectiveness.

As the EU AI Act approaches implementation, how each AI provider will prepare technical and institutional measures has become a common industry challenge. The existence of such loopholes demonstrates that for watermarking technology to function as a regulatory compliance tool, improvements are needed in both technical specifications and operational rules. Going forward, attention will focus on whether technical fixes will be applied to older models or whether policy-level measures will be considered as de facto compliance.