Anthropic's new text watermarking leaves a loophole on older Claude models

The new watermarking rollout

Anthropic is expanding its text watermarking system to include four additional Claude models. This expansion covers the Mythos 5, Opus 4.7, Opus 4.6, and Sonnet 4.6 versions. The company is moving to include these older models in its effort to provide detectable patterns in AI-generated text.

The watermark functions as an imperceptible pattern in word choice. Anthropic claims that this method does not change the meaning, quality, or readability of the output. It also does not add extra characters or tokens, which means it should not affect pricing or the format of the requests and responses.

The company states that the watermark does not include any data regarding the person using the tool, their company, or their specific chats with Claude. It is designed to be a statistical signature rather than a data tag. This rollout is part of a broader strategy to ensure models meet upcoming regulatory requirements.

A technical loophole for users

While the rollout is happening, a specific loophole exists for API customers using two older models. On the Opus 4.6 and Sonnet 4.6 versions, the watermark only applies when developers leave the three sampling parameters at their default values. These parameters include temperature, top_p, and top_k.

Sampling parameters are settings that control how much randomness a model uses when picking words. By making even a minor tweak to any one of these three settings, a user can effectively dodge the watermark. This means the statistical pattern required for detection is no longer consistently applied by the model.

Anthropic warned in an email to customers that they should keep these settings at their defaults if they rely on the watermark for compliance. Interestingly, the company did not include this specific warning for Claude Mythos 5 or Claude Opus 4.7. This suggests the watermark on those two models applies regardless of the sampling settings used.

Compliance and the EU AI Act

Anthropic cited the EU AI Act as a primary reason for retrofitting these older models with watermarks. The regulation sets a deadline for content labeling, and the company is working to ensure its tools are compliant. Every model released after August 2, 2026, will ship with the watermark enabled on day one.

Because the watermark is embedded at the model level, it works across every platform that hosts these models. This includes major providers like Amazon Web Services, Google Cloud, and Microsoft Foundry. Customers using these platforms cannot turn the watermark off, except where the technical loophole applies to the older models.

The technology behind the watermark is a variant of Google DeepMind's SynthID text method. This method tweaks the randomness behind word selection to produce a pattern that can be verified. Anthropic currently offers a detection API in private preview for users who need to verify these marks.

Debate over quality and utility

There is ongoing debate regarding whether watermarking truly leaves text quality untouched. Some critics argue that because no two synonyms mean exactly the same thing, picking words based on a watermark key instead of semantic fit will eventually degrade precision. They suggest that the nuance of language might be lost in the process.

Legal experts have also raised concerns about how these marks might appear in professional settings. For example, law firms might face issues if clients who have banned AI use discover the mark in a document. There is also a concern that multiple watermarks layered inside a single document could create technical complications for those reviewing the files.

In comparison, other companies are taking different approaches to watermarking. OpenAI launched a system called textGrain in October 2026, but it only activates by default for specific users in the EU. For their API customers worldwide, the system ships disabled and can be flipped on or off at the user's discretion.