Anthropic has begun attaching machine-readable marks to the text and images that its Claude models generate, and will apply them to users worldwide rather than only in Europe, after signing a European Union code of practice tied to new transparency rules that took effect on 2 August 2026.

The company confirmed in a support document that it had signed the EU AI Act’s Article 50(2) Code of Practice on Transparency of AI-generated Content, as a provider of both generative AI models and generative AI systems.

The step makes Anthropic one of the first large AI developers to set out, in operational terms, how it intends to comply with a law that will eventually require most AI-generated material circulating in the EU to be identifiable as such.

The obligations reach further than one company because they divide responsibility between two groups.

Providers, the firms that build and supply AI models, must ensure their systems’ output can be recognised as artificial.

Deployers, the far larger set of organisations that use those systems to create and publish content, carry separate duties to disclose it. The distinction determines who is responsible, and for what.

How Anthropic is marking Claude’s output

According to Anthropic’s documentation, Claude uses two techniques. For text, it embeds an imperceptible watermark that the company says does not change the meaning, quality, or readability of a response, and that remains in place when the text is copied and pasted or lightly edited.

For files, it attaches signed provenance metadata to supported image formats, including .png, .jpg, and .svg, following the open C2PA standard maintained by the Coalition for Content Provenance and Authenticity.

That metadata records where a file came from and can reveal whether it has been tampered with. The same scheme has been adopted by Adobe, the BBC, and a number of camera manufacturers to label the origin of images.

Claude models launched in the EU on or after 2 August 2026 will carry the marking at launch, Anthropic said, while models released before that date are being updated to support it.

The marking spans the company’s products, including its developer API, the Claude apps, Claude Code, and its enterprise deployments, and is being applied globally rather than restricted to European users.

Anthropic attached an important qualification. A detected mark shows only that content “may have been processed by Claude,” the company said, and the absence of a mark does not rule out that content was AI-generated, because output from older models or heavily edited text may carry no mark at all.

The company said it would publish further technical documentation on how third parties can detect its marks.

A machine-readable mark is not the same as a visible label. The watermark and metadata are signals a detection tool can read, not a notice a reader sees on the page.

Turning one into the other, an on-screen disclosure that a human can understand, is a separate step, and under the AI Act it is largely a duty for the organisations that publish the content rather than for the model provider.

How it compares with rival AI companies?

Anthropic is not the first major developer to move on content marking. In May 2026, OpenAI joined the C2PA coalition and partnered with Google to embed the latter’s SynthID watermark into its image outputs, and previewed a tool that lets users check whether an image was generated by its models.

SynthID, developed by Google DeepMind, embeds invisible marks into images that are designed to persist through screenshots, resizing, and compression. C2PA had attracted more than 6,000 members and affiliates by early 2026, though adoption across the wider industry remains uneven.

What sets Anthropic’s commitment apart is its emphasis on watermarking text, the format Claude is most used to produce and the hardest to mark reliably, and its framing of the work as compliance with a specific legal obligation rather than as a voluntary provenance effort.

What Article 50 requires?

The obligations stem from Article 50 of the AI Act, the bloc’s risk-based law governing artificial intelligence, which entered into force in 2024 and is being phased in over several years.

The transparency provisions in Article 50 became applicable on 2 August 2026.

Paragraph 2 of the article addresses providers. It requires that the outputs of generative AI systems be “marked in a machine-readable format and detectable as artificially generated or manipulated,” using technical solutions that are “effective, interoperable, robust and reliable as far as this is technically feasible.”

The requirement carries narrow exceptions, including cases where an AI system performs only an assistive or minor editing function and does not substantially alter the input.

Paragraph 4 addresses deployers, and it is here that the obligations extend to organisations that simply use AI tools. Anyone deploying a system that produces a deep fake, meaning a synthetic or manipulated image, audio, or video resembling real people, objects, or events, must disclose that the content has been artificially generated.

Deployers that publish AI-generated text to inform the public on matters of public interest must also disclose it, unless the content has undergone human review and a person or organisation holds editorial responsibility for it.

Material that is evidently artistic, creative, satirical, or fictional is treated more lightly, needing only an indication that synthetic content is present in a way that does not spoil the work.

In practice, the duty to label often falls on the user rather than the model. A marketing team producing a synthetic product image, a newsroom drafting copy with an AI assistant, or an agency generating an AI voiceover would each count as a deployer under the Act.

A provider’s machine-readable mark can make that disclosure easier to automate further down the chain, but it does not shift the legal responsibility from the deployer onto the company that built the model.

The Code of Practice that Anthropic signed is a voluntary instrument intended to help companies meet the Article 50 obligations.

It was published by the European Commission on 31 July 2026, two days before the rules applied, and sets out what the Commission describes as a “streamlined, predictable and legally certain pathway” to compliance.

Signing the code and following it gives a company a presumption that it is meeting the relevant obligations.

Around 190 organisations have signed, according to the Commission, spanning sectors from IT and telecoms to education and retail, with about half of them small or recently founded companies.

The code is split into two sections: 82 organisations have signed the section covering providers and their marking duties, and 152 have signed the section covering deployers and disclosure.

Provider signatories include Google, Meta, Microsoft, OpenAI, and the AI video company Synthesia, while the deployer list includes Getty Images, Lenovo, and Lufthansa.

The Commission said two task forces would begin work in September 2026 to share best practice and develop the technical detail the law leaves open, an indication that key elements of implementation are still being settled.

Enforcement and open questions

Enforcement of the transparency provisions is only beginning, and questions remain over how robust the marking can be in practice. T

ext watermarks are particularly hard to make tamper-proof: paraphrasing a passage, passing it through a different model, or simply retyping it can strip out the signal.

That difficulty is part of the reason the Act qualifies its requirement with the phrase “as far as this is technically feasible,” and it is consistent with Anthropic’s own decision to present its marks as indicative rather than conclusive.

The financial stakes for non-compliance are set to be significant once oversight matures. Under the AI Act’s tiered penalty structure, breaches of the transparency obligations can draw fines of up to €15 million or 3% of a company’s total worldwide annual turnover, whichever is higher.

As we reported, regulators have been granted powers to inspect models and restrict market access, though the bodies responsible for enforcing the AI Act are still building up capacity.

The EU has positioned the AI Act as an attempt to set a global benchmark for AI governance, and Anthropic’s decision to apply Claude’s marking worldwide, rather than maintain a separate compliant product for Europe, is a further example of European rules shaping products well beyond the bloc’s borders.

It is a pattern that has drawn both praise and criticism, with some industry figures arguing that Europe risks over-regulating a field where it lags the United States and China on investment.

For the businesses that build on Claude and rival models, the more immediate task is a practical one: identifying which of their own AI-assisted outputs now require a label under Article 50, and how to attach one, before enforcement of Europe’s new transparency regime gathers pace.

Get the TNW newsletter

Get the most important tech news in your inbox each week.