Claude Watermarks Under the EU AI Act: Article 50 Explained

The short answer: The technical, machine-readable marking required by Article 50(2) is primarily the responsibility of the provider of the generative AI system. A company that merely uses Claude or another ready-made AI system does not automatically have to develop its own watermarks.

The position may be different where a company develops an AI system itself, integrates Claude or another model into its own product, and offers that product under its own name. Article 50(2) must also be distinguished from visible AI disclosures under Article 50(4).

What is Anthropic changing in Claude?

Anthropic is introducing machine-readable marking for AI-generated content produced by supported Claude models. It distinguishes between ordinary text and supported files.

Claude text receives an imperceptible watermark

When a supported Claude model generates text, Anthropic says an imperceptible watermark is woven directly into the text. Readers see a normal response; its meaning, quality and readability are intended to remain unchanged.

Because the signal is part of the text itself, it is intended to travel with the text when copied and pasted and to survive some later edits. This differs from conventional file metadata stored in a separate information field.

Marking is implemented at model level

For businesses, the model-level implementation matters. Marking is not limited to the standard Claude web interface. Anthropic lists delivery channels including:

• Claude and Claude Platform, including the Claude API
• Claude Code, Claude Cowork and Claude Tag
• AWS, Google Cloud and Microsoft Foundry

Supported models are intended to use the marking wherever they are offered worldwide. The change is driven by EU rules, but the technical implementation is not limited to users in the European Union.

Claude marks supported files differently

Anthropic does not use the same approach for supported files as it does for ordinary text. Claude can add digitally signed provenance metadata to generated or edited SVG, PNG and JPG files.

Anthropic uses the open C2PA – Coalition for Content Provenance and Authenticity standard. In simple terms:

• Generated text: embedded, imperceptible watermark
• Supported files: signed C2PA provenance metadata
• Shared purpose: machine-readable information about artificial generation or manipulation

Watermarks and metadata are often grouped together in public reporting. Technically, however, they are different methods with different limitations.

Why is Claude introducing watermarks now?

The change is linked to Article 50(2) of the EU AI Act. It applies to providers of AI systems, including general-purpose AI systems, that generate synthetic audio, image, video or text content.

Those providers must ensure that system outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. As far as technically feasible, the solution should be:

• effective
• interoperable
• robust
• reliable

The specific characteristics of different content types, implementation costs and the generally acknowledged state of the art may be taken into account.

Does the EU AI Act explicitly require watermarks?

Not exactly. “AI watermark” accurately describes Claude's current approach, but the legal requirement is technology-neutral. Article 50(2) does not prescribe one technical method.

Instead, AI-generated or manipulated content must be marked in a machine-readable format and detectable as artificially generated or manipulated. Recital 133 mentions methods including:

• watermarks and metadata
• cryptographic methods for proving provenance and authenticity
• logging and fingerprints
• combinations of different technologies

What does “machine-readable marking” mean?

Machine-readable means that information about artificial generation or manipulation does not necessarily have to be visible to people. Claude text may look entirely normal while a suitable technical system detects an additional signal.

This distinguishes Article 50(2) from visible notices such as “AI-generated” or the EU's labelling icons. Paragraph 2 creates a technical transparency layer beneath the visible content.

For a concise overview of the regulation, see our EU AI Act glossary entry.

When did Article 50(2) start to apply?

The transparency obligations in Article 50 generally apply from 2 August 2026. Shortly before that date, Regulation (EU) 2026/1744, the Digital Omnibus on AI, added a transition rule.

For relevant generative AI systems placed on the market before 2 August 2026, the measures needed to comply with Article 50(2) must be taken by 2 December 2026.

• 2 August 2026: Article 50 generally becomes applicable.
• 2 December 2026: transition period ends for relevant systems already placed on the market.

Anthropic accordingly states that Claude models launched in the EU on or after 2 August 2026 are intended to support machine-readable marking from launch. The company is working to add support to older models.

What does this mean for SMEs?

The decisive question for a business is its role: Are we using an AI system, or are we providing one ourselves? The EU AI Act treats providers and deployers differently.

Scenario 1: an SME uses Claude internally

A company may use Claude for marketing copy, summaries, research, internal documents, customer communications, translation or software development. That does not automatically make the company a provider of Claude.

The AI Act Service Desk gives a comparable example: an employer that provides employees with licences for a model already available on the EU market does not thereby become the provider of that model or have to verify the model's compliance.

Scenario 2: an SME builds its own AI product with Claude

The situation may be different where a company integrates Claude or another model into its own product and offers it under its own name. Typical examples include:

• a software vendor's own customer-service chatbot
• a Claude integration in a SaaS platform
• an AI assistant for machinery or industrial equipment
• a consultancy's AI application for clients
• a white-label AI system offered under the company's own brand

The company should then assess whether it qualifies as a provider of an AI system or a downstream provider. Using the Claude API does not automatically remove its own regulatory obligations.

Article 50(2) is therefore often more relevant to software vendors, SaaS companies and technology-focused SMEs than to organisations that merely use a finished AI tool internally.

Provider or deployer: the difference for businesses

In simplified terms, a provider develops or has developed an AI system or model and places it on the market or puts it into service under its own name or trademark. A deployer uses an AI system under its own authority.

• Provider: Anthropic supplies Claude – relevant to machine-readable marking under Article 50(2).
• Deployer: a company uses Claude – other obligations may arise depending on the use case.
• Downstream provider: a company integrates Claude into its own AI product – provider obligations may arise.

This role assessment belongs in every AI Act inventory and should not wait until a product has already been delivered to customers.

Which exceptions may apply?

Article 50(2) contains an exception where an AI system merely performs an assistive function for standard editing or does not substantially alter the input data or its semantics. European Commission guidelines provide further detail.

• short sequences of numbers, symbols or letters
• source code
• machine-to-machine outputs processed without exposure to people
• certain intermediate outputs in closed industrial or product-development environments where they are not the final output
• narrow business-to-business and industrial scenarios where the guideline conditions are met

The exceptions depend on context and should be interpreted narrowly. Industrial SMEs and companies with automated internal AI processes should document why a particular output is considered outside the marking obligation.

Does proofreading by Claude remain unmarked?

There is no universal answer. The legal minimum and a provider's technical implementation are separate layers. The AI Act provides exceptions for some assistive standard editing, while Anthropic describes marking generated text from supported models.

• proofreading
• translation
• summarisation
• rewriting
• editing an existing text

Text may therefore carry a Claude signal even though a person wrote the original and Claude only edited it. This is crucial when interpreting a detected mark.

Does a Claude watermark prove AI authorship?

No. Anthropic cautions against treating a mark as complete proof of provenance. A detected signal may indicate only that Claude processed the content.

• It does not prove that Claude wrote the entire text.
• It does not prove that Claude generated the underlying ideas.
• It does not prove that Claude supplied all the data.
• It does not prove that Claude was the original author.

Human-written text may merely have been proofread, translated or summarised before receiving a Claude mark. This limitation is particularly important for schools, universities and employers.

Can the Claude watermark be lost?

Yes. Text that was originally marked may later contain no detectable signal, particularly after substantial changes. Anthropic identifies causes including:

• heavy editing or extensive rewriting
• translation or mixing with other text
• substantial shortening or passages that are too short

File provenance metadata may also be lost through format conversion, resaving, screenshots or other processing steps.

Can users already detect the watermark themselves?

Anthropic says it is working on ways for users and third parties to detect embedded watermarks and provenance metadata. More detailed technical information is expected in future documentation.

This creates a transition period: supported content may already be marked while publicly documented detection tools and technical details are still being developed.

Is this the same as visible EU AI labels?

No. Both topics sit within Article 50, but they concern different transparency obligations.

Article 50(2): technical AI marking

Paragraph 2 requires machine-readable marking and detectability. Responsibility generally rests with the provider of the generative AI system – Anthropic in Claude's case.

Article 50(4): perceptible disclosure

Paragraph 4 applies to deployers and covers deepfakes as well as certain AI-generated or manipulated text published on matters of public interest. Any required disclosure must be perceptible to people.

A deployer therefore cannot rely solely on an invisible technical mark from the AI provider. Machine-readable provider marking and visible publication disclosures must be assessed separately.

• Article 50(2): technical, machine-readable, provider obligation, covering text, image, audio and video among other outputs
• Article 50(4): perceptible, deployer obligation, particularly for deepfakes and certain text publications

We explain the visible layer in detail in EU AI labels and content labelling under the AI Act.

What role does the EU Code of Practice play?

Anthropic has signed the European Code of Practice on Transparency of AI-generated Content. The distinction matters: Article 50 is binding law, while the Code is a voluntary tool for implementation and demonstrating compliance.

• Section 1: providers – machine-readable marking and detection under Article 50(2)
• Section 2: deployers – disclosure under Article 50(4)

Anthropic refers to the Section 1 commitments in its documentation. The Code also shows that machine-readable AI marking is not an isolated Claude project, but part of a broader industry response to the AI Act.

What penalties may apply?

Infringements of the Article 50 transparency obligations may be penalised under Article 99. The Regulation sets maximum levels of:

• up to EUR 15 million
• for undertakings, up to 3 per cent of total worldwide annual turnover for the preceding financial year

For undertakings, the higher amount generally applies. For SMEs, including start-ups, the lower of those maximums applies. The actual penalty depends on the individual case and national enforcement rules.

Conclusion: Claude makes an invisible part of the AI Act tangible

Claude watermarks show how an abstract legal requirement can be implemented in practice. Alongside visible AI notices, Article 50(2) creates a second transparency layer: AI-generated content can carry machine-readable provenance signals even when people cannot see them.

That does not make watermarks perfect AI detectors. Signals can be lost through editing, and an existing mark does not prove that Claude was the original author.

For businesses, the more important question remains: Are we merely using AI, or are we placing an AI system on the market ourselves? This role allocation now belongs in every AI Act and AI governance review.

If you would like to discuss your specific situation with us, contact Chifty.