Claude Is Putting Invisible Watermarks In AI-Generated Text. Even Your Human-Written Copy Could Get One

Anthropic is adding invisible, machine-readable watermarks to text generated by its Claude AI models, opening a new front in the effort to identify content created or processed by artificial intelligence.

The change could affect more than people who ask Claude to write something from scratch. A human-written press release, article or other document that is simply proofread, translated or reformatted with Claude could also end up carrying a detectable AI signature.

The move comes as AI companies adapt their products to meet transparency requirements under the European Union's AI Act.

Anthropic is using two methods to mark content associated with newer Claude models. For text, the company embeds patterns into the output that are designed to be readable by detection systems while remaining imperceptible to people.

Files generated through Claude can also contain metadata with digital signatures showing that the content was processed by the AI system.

Axios reported that the measures apply to models launched in the European Union after Aug. 2, with Anthropic implementing the marking technology wherever those models are offered worldwide.

That means users outside Europe could encounter the same system even though the regulatory requirement originates in the EU.

The technology also comes with an important limitation. A watermark does not necessarily mean Claude wrote the underlying material.

Someone could draft a document themselves and then use Claude to correct grammar, translate the text or change its formatting. The resulting material could still contain a detectable Claude mark.

That distinction could matter for communications teams, journalists, marketers and other professionals who increasingly use AI as an editing or production tool rather than as the original author.

Anthropic also says its system can lose effectiveness when text is heavily rewritten, combined with other material or is too short. In those situations, the watermark may no longer be detectable.

The result is that a watermark can provide evidence of Claude involvement without necessarily establishing who wrote the original material or how much AI contributed to it.

The rollout is linked to Article 50 of the EU AI Act, which introduces transparency obligations for certain AI-generated and manipulated content.

The European Commission said providers covered by the rules must use machine-readable markings that allow AI-generated or manipulated content to be detected. The requirements are intended to make synthetic material easier to identify and address risks such as deception and manipulation.

The rules do not mean every piece of AI-assisted text will have to display an obvious warning to readers. The EU framework distinguishes between technical marking by AI providers and disclosure obligations for specific types of content, including certain deepfakes and AI-generated material concerning matters of public interest.

The European Commission has also published a voluntary Code of Practice intended to help AI providers and users implement the transparency requirements.

Anthropic is not the only major AI company working on content provenance.

OpenAI has also outlined its approach to the EU AI Act, although its current watermarking and provenance efforts have focused primarily on media such as images and audio rather than introducing a comparable watermarking system for text.

The push is also spreading to platforms that host user-generated content.

Axios reported in July that Substack partnered with AI-detection company Pangram to give users tools for assessing whether posts, notes, comments and replies appear to have been generated by AI.

Snap has taken a different approach by changing how AI-generated material is promoted on its platform.

Snap said it is prioritizing original content on Spotlight and reducing the visibility of synthetic AI videos and widely syndicated material as it seeks to emphasize content created by users.

These moves point to a wider change in how the technology industry is approaching AI content. As generative AI becomes part of everyday writing, editing, translation and media production, companies are increasingly looking for ways to establish whether and how AI was involved.

The EU's rules could accelerate that process.

AI systems that were already on the market before Aug. 2 have a transitional period to comply with the marking requirements. According to European Commission guidance, the deadline for those legacy systems is Dec. 2, 2026.

That gives other major AI providers several months to reveal how they intend to identify content produced by their older models.