Is Claude watermarking your code yet?
Anthropic gave its content-marking rollout a precise cutoff: 2 August 2026. It is a model-launch cutoff, not proof that every current Claude response is already marked.
1What the date actually says
Anthropic says Claude models launched in the EU on or after 2 August 2026 will support machine-readable marking at launch. The mark is meant to be embedded at the model level in generated text, across surfaces that include Claude Code, and used worldwide on supported models.
Models launched before that date are different. Anthropic describes a transition period and says it is working to add marking support to those models too. It has not published a model-by-model rollout table.
The careful reading is therefore: future models mark from launch; earlier models may be retrofitted. The announcement alone neither confirms nor rules out marking on a current model.
2Why that distinction matters now
As of this writing, Anthropic's current model dates all sit before the cutoff. Fable 5 and Mythos 5 became available on 9 June, Sonnet 5 launched on 30 June, and Opus 5 launched on 24 July. Even the newest public model therefore predates 2 August.
That makes headlines claiming that every Claude response is already watermarked too strong. But pinning an older Claude model is not a durable escape hatch either: Anthropic has said the earlier models are part of the transition, and it has not said which have already changed.
3Code is the unresolved part
There is another wrinkle for developer tools. The European Commission's Article 50 guidance expressly places source code outside the provider-side marking obligation. That exclusion is broad: it includes programming, scripting, markup, query and configuration languages, plus formats such as SQL, infrastructure-as-code, YAML and JSON configuration.
Anthropic's own description appears broader. It names Claude Code and says supported models mark all generated text. Until Anthropic publishes a code-specific explanation, we should not assume either that code is exempt from its product behaviour or that code produced today is marked.
We also do not know the algorithm. Nothing Anthropic has published says that the mark is hidden Unicode, unusual whitespace or file metadata. Statistical token selection is plausible—public systems such as SynthID-Text work that way—but it remains an inference, not a description of Claude's implementation. Public code-watermark research also shows why this is difficult: code is often short, low-entropy and intolerant of token substitutions.
4What OVM will watch
Anthropic says detection details and technical documentation are still forthcoming. Until those exist, OVM cannot distinguish a real provider watermark from stylistic guesswork, and a formatter or Unicode scrubber cannot prove that it removed one.
We will be watching for:
- a model support table and retrofit dates;
- an official detector or public technical specification;
- an explicit answer for source code, tool calls and file writes;
- code-specific minimum lengths and false-positive rates;
- local detection, or clear privacy and retention terms for remote detection.
If those pieces arrive, the responsible first step is a read-only provenance experiment—not a confident "AI versus human" label, and not an unverified cleanup button. For now, the honest status is announced, transitional and not yet independently detectable.
5References
- Anthropic — How Claude marks AI-generated content · rollout scope, cutoff date, transition and detection limitations
- Anthropic — Models overview · current model availability
- Anthropic — Claude Platform release notes · model launch dates
- European Union — AI Act, Article 50 · provider transparency obligation
- European Commission — Article 50 implementation guidelines · source-code exclusion in paragraph 70
- European Commission — Code of Practice on AI-generated content · implementation and detector expectations
- Nature — Scalable watermarking for identifying large language model outputs · SynthID-Text design and detection trade-offs
- ACL Anthology — SWEET: Who Wrote this Code? · source-code-specific watermarking constraints
- STONE — Syntax-aware source-code watermarking · a contrasting code-specific approach