Anthropic Watermarks: Why Some Older Claude Models Remain Vulnerable
The company is updating its labeling system to meet EU regulatory requirements
Safety and transparency standards in the artificial intelligence sector are tightening. Anthropic has announced that it is implementing its text watermarking system across four additional models. This decision is linked to the December deadlines set by the EU AI Act.
Changes in Claude Models
The updated system applies to Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, and Claude Sonnet 4.6. According to the company, the watermark uses a sophisticated word-selection scheme that is imperceptible to humans and does not alter the quality, content, or readability of the text. This labeling operates at the model level, meaning it is automatically activated on Amazon Web Services, Google Cloud, and Microsoft Foundry platforms.
However, Anthropic has acknowledged a technical nuance that allows some users to bypass the watermark. Specifically, for the Claude Opus 4.6 and Claude Sonnet 4.6 models, the labeling only functions if API parameters—such as temperature, top_p, and top_k—are left at their default values. Even minor adjustments to these settings can prevent the system from applying the watermark. The company warns developers that if regulatory compliance is a priority, they should keep these parameters unchanged.
Technological Challenges and Criticism
The method used by Anthropic is based on Google DeepMind’s SynthID technology, which creates statistical patterns during word selection. While the company maintains that this process does not affect text quality, industry experts remain sceptical. Critics point out that selecting synonyms based on watermark encoding rather than semantic precision inevitably degrades the quality of the text.
There is also a legal dimension to consider. Law firms have expressed concern that such markers in documents could become problematic in environments where the use of AI is restricted. Furthermore, the presence of multiple different watermarks in a single document could lead to further technical complications.
What This Means for Users
Anthropic is currently working on a detection API that will be able to verify these watermarks. Despite the existing flaws, this step demonstrates that technology giants are striving to comply with international regulations, which will make the identification of AI-generated content more standard in the future. For users in Georgia, where international API services are actively utilized, these changes are significant for both software developers and businesses that use AI for content generation.