Claude now watermarks text from older models, with no opt-out

Text written with Claude now contains a machine-readable mark that identifies Claude as the source. Every Claude model released since August 2 has shipped with an invisible watermark, and Anthropic is now adding it to older models. Since November 3, Claude Mythos 5, Opus 4.7 and Sonnet 4.6 also mark their output, according to an email to API customers reported by The Decoder. The model table on Anthropic's help page already lists all three as supported. Opus 4.6 was also due on November 3, but Anthropic moved it to November 16 without giving a reason. Opus 4.5, Sonnet 4.5 and Haiku 4.5 follow on the same day, and on Amazon Bedrock the rollout takes a few extra days.

The EU AI Act is the reason

Anthropic has signed the code of practice for Article 50 of the EU AI Act, which calls for machine-readable marking of AI-generated content. The Decoder says the company cites a December deadline for the retrofit. We explained what the EU labeling rules mean for users in August. The marking applies worldwide and across products, including the Claude apps, Claude Code, Cowork and the API. Because it is applied at the model level, it also appears in text generated through AWS, Google Cloud or Microsoft Foundry. According to The Decoder, customers cannot turn it off.

What the watermark contains

Claude subtly adjusts its word choice to create a statistical pattern that readers do not notice. The Decoder reports that Anthropic uses a variant of Google's SynthID, the same technology used for Apple's AI images in iOS 27.2. In its email, Anthropic says the watermark "contains no information about the user, their organization, or their conversations with Claude." The help page says the mark will "travel with the text when it's copied and pasted elsewhere" and may persist through some editing. Heavily rewritten, translated or very short passages may not carry a detectable mark.

Hardly anyone can check the watermark yet. Detection is in private preview for regulators, law enforcement, media, fact-checkers, researchers, educational organizations and companies with their own obligations under the law, and access requires an application. For PNG and JPEG files, Claude also adds Content Credentials based on the C2PA standard, which anyone can read with the free Claude Content Checker. Anthropic notes that a detected mark does not prove authorship. Text that Claude only proofread or translated carries the mark as well.

A loophole mostly affects developers

On Sonnet 4.6 and Opus 4.6, the watermark is only reliably applied when temperature, top_p and top_k stay at their default values. The Decoder says the same limit applies to all four models in the second wave. Developers who change one of these settings through the API get text that may not be detected. The Claude apps do not offer these settings. OpenAI handles this differently. According to The Decoder, its watermark has been on by default since October only for ChatGPT and Codex users in the EU, while API customers have to turn it on themselves.