Anthropic text watermark can be bypassed on older Claude models

What happened

Anthropic is rolling out text watermarks to four more Claude models ahead of the EU AI Act's December deadline. On older models, including Claude Opus 4.6 and Sonnet 4.6, API users can disable the watermark by changing a single sampling parameter, according to The Decoder, which first reported the loophole.

The report describes Anthropic's response so far as a compliance warning. The watermarks are designed to signal that text was machine-generated, so a setting that removes them undermines the disclosure they are meant to provide.

The timing is tied to the regulatory calendar. The rollout is positioned around the EU AI Act's December deadline, which puts pressure on providers to make AI-generated content identifiable before it reaches end users.

Why it matters

For organisations building on Claude through the API, the story is less about one parameter and more about where a compliance control actually sits. If disclosure depends on a setting that developers can change, transparency becomes a design choice made downstream rather than a guarantee built into the model.

The lesson extends beyond one vendor. As regulators move to require disclosure of AI-generated content, the operational question shifts from "does our model watermark its output?" to "does the product still behave that way once developers start configuring it?" That is a service-design question as much as a legal one, and it will shape how AI features are documented, tested and governed.

By the numbers

• Four additional Claude models receiving text watermarks
• Two named examples of older models where the watermark can be bypassed: Claude Opus 4.6 and Sonnet 4.6
• One sampling parameter that users of the API can change to disable the watermark
• December the EU AI Act deadline the rollout is timed against

The Renascence take

Most coverage will treat this as a vendor loophole. We read it as a configuration problem: a safeguard that lives in a parameter is only as reliable as the defaults, documentation and governance wrapped around it.

A watermark that the end user can switch off is a disclosure preference, not a control. The behavioural principle underneath is defaults: most people never change them, but a motivated developer will, and that is exactly the population a compliance regime needs to account for. Customer-obsessed operators should treat AI disclosure as a product feature to be tested end to end, including the paths a developer can take to remove it. Audit your own integrations this quarter, and make disclosure the default rather than an option.