That distinction changes the immediate takeaway for Claude users. The policy is global for supported models: Anthropic says marks will apply across the Claude API, the consumer chat service, Claude Code, Claude Cowork, Claude Tag, and other places Claude is offered. But there is no public confirmation that current pre-August 2 models have been retrofitted, no named list of marked models, and no public detector that customers or third parties can use today.
The story, first circulated by Glitchwire, correctly identifies Article 50 of the EU AI Act as the driver of a new generation of content-origin controls. It gets ahead of the available evidence, however, when it frames the mechanism as a customer-specific surveillance or tracing system. Anthropic says a detected mark indicates that content may have been processed by Claude; it does not establish a complete chain of custody, identify the person who used Claude, identify an account, or prove that a final document was wholly generated by the model.
Anthropic’s rollout is global, but it is not yet universal
Anthropic’s support documentation describes two different technical paths. For text, a supported Claude model will weave an imperceptible watermark into its generated wording. The company says the mark remains with copied text and can survive some edits because it is embedded in the statistical choices made during generation rather than attached as a visible label or ordinary document property.
For files such as PNG, JPG, and SVG, Anthropic says it will add cryptographically signed provenance metadata built around the C2PA standard. That is a different mechanism with different limits. Metadata can show a signed origin history when a compatible file remains intact; a statistical text watermark is an inference drawn from the words themselves. Users should not assume that a C2PA record exists in a pasted code block, a Slack message, a screenshot, or a re-created document.
Anthropic also says marking is imposed at the model level rather than selected by product surface. For an organization using Claude through the API, that means the practical behavior should match Claude’s chat interface once the same supported model is in use. There is no announced API parameter, enterprise control, or Claude Code setting to disable it.
Still, “model level” does not mean all existing Claude outputs have suddenly changed. The company’s wording matters: it commits to marking new models launched in the EU from August 2 onward and says it is working to bring older models into the program. Anthropic has not said which previously released Claude models are marked now, when retrofits will arrive, or whether every model family and region will transition at the same time.
That missing model-by-model rollout record is the central operational gap. A company cannot responsibly treat a Claude output generated on August 11, 2026, as marked or unmarked solely from the date. It needs the exact model identifier and a statement from Anthropic that the model is supported by the marking system.
Article 50 requires detection, not a global customer-tracking system
The European Commission’s Article 50 guidance says providers of generative AI systems must apply machine-readable marks to synthetic content and make it detectable as artificially generated or manipulated, subject to exceptions such as standard assistive editing that does not substantially alter the user’s input or its meaning. The legal obligation began applying on August 2, 2026.
The Commission also draws a line that much of the early coverage has blurred. Providers must make output machine-detectable. Deployers—the businesses or individuals publishing or operating AI-enabled systems—have separate duties to visibly disclose deepfakes and certain AI-generated text aimed at informing the public on matters of public interest, particularly where there has been no human review or editorial control.
In short, a hidden mark is not automatically a public label, and a machine-readable mark does not decide whether an organization has met every disclosure obligation. A company publishing AI-assisted material still needs to assess its own role, its audience, its review process, and the rules that apply in the jurisdictions where it operates.
The regulation also does not require Anthropic to deploy one global implementation. The European Commission’s transparency rules govern systems placed on the EU market. Anthropic’s decision to apply marking worldwide for supported models is a product and operations decision that avoids maintaining geographically separate model behaviors. It may also reduce the chance that unmarked output generated elsewhere is later distributed into the EU.
That is broader than the narrowest possible compliance strategy, but it is not evidence that the EU mandated universal tracking of every Claude user. Nor is it a surprise that a provider operating a single hosted model service would prefer one model configuration over regional forks.
Existing systems have a transition period. The EU’s 2026 AI Act amendment gives generative systems placed on the market before August 2 until December 2, 2026, to meet the Article 50 marking obligation. That four-month window explains Anthropic’s split between future launches and older models; it also means the assertion that Anthropic had to watermark every existing Claude model on August 2 is incorrect.
The maximum penalty cited in the underlying report—€15 million or 3% of worldwide annual turnover—is consistent with the Commission’s published enforcement summary. But those numbers do not establish that any particular Claude model is presently out of compliance, or that Anthropic had no other implementation options.
A positive result will be a signal, not proof
Anthropic says it is working to let users and third parties detect its watermarks and provenance metadata. That statement should not be read as confirmation that the public can already submit any text to a Claude detector, nor that Anthropic will provide an unrestricted lookup service. The company has not published the detector, the minimum text length needed for a meaningful result, error rates, supported languages, or the terms under which outside parties will receive access.
Its own described limitation is substantial: a detected Claude mark indicates content may have been processed by Claude, but does not establish the full provenance of the final work. Conversely, no detected mark does not prove the text was human-written or never touched by Claude. A user could edit, paraphrase, translate, merge, or selectively quote model output; a system could also encounter too little text to produce a reliable result.
This is where the “snitching” framing breaks down. The available description points to origin signaling, not a visible serial number tied to the account holder. That does not eliminate policy concerns—an employer, school, publisher, or platform could overinterpret a positive mark—but it is a materially different capability from a customer-identification database.
The European Commission’s own technical study on Article 50 flags why vendors must be careful. It identifies five broad approaches to generated-text detection—watermarking, structural marking, metadata, logging, and AI-text detectors—and evaluates them against effectiveness, robustness, reliability, accessibility, and interoperability. The study does not declare any one technique a universal solution.
A recent, not-yet-peer-reviewed evaluation of three representative watermark schemes reached an even harsher conclusion: meaning-preserving paraphrasing removed nearly all detectable marks in the tested configurations, while false negatives were already high in some baseline tests. Those findings do not prove Anthropic’s unpublished implementation will perform the same way. They do show why neither Anthropic’s assurance that quality is unchanged nor critics’ claim that the watermark is impossible to remove can be accepted without model-specific testing.
Claude Code users should plan for provenance, not panic
For Windows developers and IT teams, the immediate concern is less whether a future watermark exists than how it will be interpreted by downstream systems. Claude Code and API users often generate fragments: PowerShell commands, YAML files, code comments, test cases, release notes, documentation, and patches that are then extensively reviewed and revised. A mark on output is not a reliable measure of authorship, quality, licensing status, or security review.
Organizations should avoid turning detection into an automated enforcement trigger. A positive Claude mark should prompt a review of the work and the organization’s AI-use policy; it should not by itself establish misconduct, invalidate a contribution, or substitute for code review. A negative result should not be treated as a clean bill of health either.
The new marking system also does not resolve the harder enterprise questions around Claude Code: whether source code was sent to Anthropic, which model handled it, what retention and training terms apply, whether generated code passes license and security checks, and who is accountable for merging it. Watermarking describes a possible origin signal. It does not create a software supply-chain record.
The near-term facts are therefore narrower than the rhetoric: Anthropic has committed to worldwide marking for supported Claude models, with text watermarks and C2PA-style metadata for supported files. It has not publicly demonstrated that existing Claude models are all marked, published its detection tooling, disclosed independent robustness results, or provided a customer opt-out. The key date for older Claude models is now December 2, 2026, when the EU transition period expires—and Anthropic will need to replace broad promises with a precise supported-model list and a detector that can withstand real-world editing.