Claude now marks its text worldwide. Only a few can check the mark

Facts are attributed and linked to their sources. The analysis and opinions are the author's own.

What it is

Anthropic has signed the EU AI Act's Article 50(2) code of practice on marking AI-generated content. Its help centre says models launched on or after 2 August 2026 are marked from launch, and that the marks apply worldwide across the API, the Claude apps, Claude Code and Cowork, as well as through AWS, Google Cloud and Microsoft Foundry. Generated files also carry C2PA content credentials. Detection is in private preview, open only to eligible organisations.

On Friday Anthropic told customers it would extend the watermark to Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6, starting on 3 November. The Decoder reported a catch in that email. On Opus 4.6 and Sonnet 4.6 the mark applies only if temperature, top_p and top_k are left at their defaults. Anthropic's advice: "If you rely on the watermark for your own compliance, keep these settings at their defaults." The Decoder says the method is a variant of Google's SynthID text approach, that users cannot opt out, and that the trigger is the EU AI Act's December deadline.

OpenAI chose differently on 5 October. Its textGrain watermark is on by default only for eligible ChatGPT and Codex users in the EU. For API customers anywhere it is opt-in and off by default, and detector access is limited to approved researchers and expert organisations. OpenAI published its own test results. At a 1% false-positive target, it detected the mark in about 80% of 200-token passages and about 95% of 400-token passages on psychology content. Replacing 10% of words with synonyms cut detection from about 92% to 66%. Replacing 25% cut it to 17%.

Why it matters

A watermark that the public cannot check creates an imbalance. A regulator, a newspaper or a fact-checker may be able to test text you publish. You, the publisher, may not. Both companies are also clear about the limits. Anthropic's help page says proofreading by Claude can leave a mark on human writing, and that short or heavily edited text may carry none. OpenAI says the absence of a mark does not prove a human wrote the text.

A watermark can tell you a model touched the text. It cannot tell you who is responsible for what it says.

In my view, the practical issue for a regulated firm is disclosure policy, not detection. Customer letters, marketing copy and research notes drafted with Claude may now carry a mark whether or not anyone intended it. The sampling-parameter catch also creates a control question: an engineer who changes temperature on an older model may switch the mark off without knowing. If a firm relies on the mark for compliance, that setting belongs in change control.

What I'd watch

• Whether either company opens detection beyond its preview or approved-researcher groups.
• Whether compliance teams start treating sampling-parameter changes on older models as control changes.
• How EU supervisors treat marks that can be weakened by light editing or translation.