Anthropic says future models it launches in the EU will mark generated text and files to meet AI Act transparency rules.
The Register reported that Anthropic plans to embed watermarks in text and add digitally signed provenance metadata to files generated by future models it launches in the EU. Anthropic said in a help document that the change is part of its effort to comply with content and transparency rules in the EU's AI Act.
The basic idea is straightforward: content produced by Claude would carry signals intended to show that AI was involved. For text, Anthropic says the mark will be built into the text itself. For files, it says it will use signed metadata where that is supported.
What Anthropic says it is changing
According to The Register's account of Anthropic's help document, generated text from supported Claude models will carry embedded watermarks. Generated files will include digitally signed provenance metadata where supported.
Anthropic also said it is working on models it has already released so that output marking can be added during the transition period allowed under EU law, as reported by The Register. That means this is not presented only as a feature for brand-new systems; the company says it is also trying to retrofit marking to existing models.
The scope is broader than Anthropic's own website or app. The report says the marking is intended for output from covered Claude models on the Claude Platform API, Claude, Claude Code, Claude Cowork and Claude Tag. It also says this includes third-party providers of Anthropic models such as AWS, Google Cloud and Microsoft Foundry.
Anthropic further said that marking will apply to output from supported models wherever Claude is offered, worldwide. In practice, that suggests a compliance change linked to an EU launch could affect organisations using Claude outside the EU as well, if they rely on the supported models and services named in the document.
How the marking is meant to work
The Register quoted Anthropic saying that when a supported Claude model generates text, it "weaves an imperceptible watermark directly into the text itself" and that the mark does not change the meaning, quality or readability of the response.
In plain English, Anthropic is describing a hidden marker that people are not supposed to notice when reading the text. The company also says the mark is applied at the model level, so it would remain present regardless of which Claude product or interface produced the text.
The same report says Anthropic believes the text watermark should travel with copied and pasted text and may persist through some editing. That matters because a watermark that disappears as soon as text is moved from one tool to another would have limited practical value for tracing where content came from.
For files, Anthropic says it will rely on signed provenance metadata that conforms to the C2PA standard, according to The Register. Here, provenance metadata means attached information intended to record where a file came from and whether it has been altered. The report says Anthropic expects to publish details on how people can detect Claude's marks in forthcoming documentation, which it describes as required under EU law.
What remains uncertain
The available detail is still thin. The Register noted that, without examples or technical documentation, it is unclear how Anthropic will make text watermarks difficult to remove.
The same report also points to limits on file-based marking. It says Anthropic plans to use C2PA-signed provenance metadata, but also notes that open source tools already exist to remove that metadata. The article adds that researchers have previously shown image-based watermarking can be undone, though it does not claim Anthropic is using the same method for text.
Anthropic itself appears to be setting expectations cautiously. According to The Register's summary of the company's wording, a detected mark is not conclusive proof that Claude created the content, and the absence of a mark does not guarantee that AI was not involved. That is an important limitation: these signals are being framed as indicators, not definitive forensic evidence.
Why it matters for customers and IT teams
The immediate significance is regulatory. Anthropic is explicitly tying the change to EU AI Act content and transparency requirements, as The Register reported. For organisations using Claude through Anthropic's own products or through cloud providers, that means AI-generated output may start carrying origin signals by default on supported models.
That could create practical compliance and workflow questions for businesses that publish, redistribute or store Claude-generated content. However, the source material does not show how enterprises are responding yet, and it does not provide evidence of specific customer policies or deployment changes.
The report does say the move may alienate some Claude customers who would prefer end users not to know that content was AI-generated. It also notes user scepticism about whether text watermarking will work in practice. Those points suggest two tensions at once: regulators may want clearer disclosure, while some users may worry about reliability, removability or the commercial effect of making AI involvement easier to detect.
A likely implication is that organisations using Claude across several channels will need to pay attention to Anthropic's forthcoming detection documentation. If the company applies marking at the model level across its own services and third-party platforms, the change may be less a per-product setting than a characteristic of the underlying model output.
What to watch next
What is known now is narrow but concrete. Anthropic says future models it launches in the EU will add hidden text watermarks and signed provenance metadata for generated files where supported, and it says it is also working to add output marking to already released models during the legal transition period.
What remains open is how robust the system will be, how detection will work in practice, and how much confidence customers or regulators will place in the marks. The strongest conclusion supported by the source is that this is a real compliance move, but not yet proven as a reliable way to establish authorship or prevent removal.