AI Content Watermarks Become Mandatory Under EU Rules

AI content watermarks are moving into practice as technology companies face new transparency duties under the EU AI Act.

Anthropic has begun adding machine-readable marks to content generated by Claude. The company’s move follows Article 50 requirements that aim to help people identify synthetic or manipulated material.

The rules apply to providers of general-purpose AI models. Companies that fail to comply could face fines of up to €15 million or 3% of their global annual turnover.

Anthropic begins watermarking Claude output

Anthropic said Claude-generated content will carry embedded markers that machines can detect. The system applies to text, images and files where the relevant technology supports it.

For text, the company uses invisible watermarking. These marks should not affect the wording, quality or readability of an output. They are designed to remain associated with text when a user copies and pastes it elsewhere.

Files generated by Claude can also carry digitally signed provenance metadata. This metadata can apply to formats such as SVG, PNG and JPG files.

A valid signature can indicate that Claude processed a file. It can also help reveal whether someone altered the marked file after generation.

Anthropic has started the rollout for Claude models launched in the European Union on or after August 2, 2026. It plans to add AI content watermarks to older models over time.

EU rules put pressure on AI companies

Article 50 requires organisations to make AI-generated or manipulated material machine-readable in relevant cases. The goal is to give platforms, regulators and users better ways to assess where content came from.

The requirements reach beyond one company. General-purpose AI providers operating under the EU framework must meet the same standards.

Fines can reach €15 million, which is roughly $17 million, or 3% of total worldwide annual turnover for the previous financial year. The potential penalty gives AI developers a strong reason to build provenance systems into their products.

Anthropic plans to apply its marking system globally, even though the legal requirement comes from the EU. This could make its approach visible to users across markets.

Watermarks cannot prove everything

AI content watermarks have important limits. A detected mark does not prove that an entire document, image or file was created by AI.

For example, a person may submit an original text to Claude for proofreading, summarising or translation. The resulting version may receive the same mark as content written entirely by the model.

The absence of a watermark also does not prove that material came from a human. Older files, unsupported formats and outputs created before marking systems were available may not include any provenance signal.

Heavy editing can also weaken these systems. Paraphrasing, translating or combining AI-generated text with other material may remove a watermark or make it harder to detect. Very short outputs can create similar challenges.

Conclusion

AI content watermarks will become an important part of online transparency under the EU AI Act. Anthropic has already started the process with Claude, but no marking method can offer perfect proof of a file’s origin. Effective identification will require watermarks alongside metadata, standards and reliable verification tools.