AI-generated text, files will soon become machine-identifiable: What Anthropic's Claude watermarks mean for you

AI giant Anthropic has said it will add watermarks that are undetectable to the human eye to its new Claude models.

The company announced this move in response to the transparency requirements under the European Union's AI Act, Article 50.

"Generated text will carry embedded watermarks, and generated files will include digitally signed provenance metadata where supported," the company said in a help document published on its website on Monday.

Claude models launching from 2 August will have this feature, in which content generated at the model level will carry the said mark. The company said these markings would be applied worldwide, and not to European users alone.

How will watermarks work?

As per Anthropic, the plan is to embed the feature directly into the text it generates for users. While the watermark will not be visible when the user reads the text generated, it will be detected by machines if the text is copied and pasted from Claude. Some of these watermarks may also survive editing, Anthropic said.

Since the mark will be applied at the model level, it will be supported across other outputs like API, Claude, Claude Code, Claude Cowork and Claude Tag. Cloud customers will thus receive marked outputs if they access supported models through major platforms.

Anthropic has specifically listed AWS, Google Cloud and Microsoft Foundry, according to a report by Interesting Engineering.

Not only text, but also files generated using Claude models will carry signed provenance metadata in accordance with the Coalition for Content Provenance and Authenticity (C2PA) standards.

How will you be affected?

The new approach by Anthropic not only means that there will be a Claude watermark if a piece of text is directly copied and pasted or lightly edited, but also that, from now on, AI-generated text may become machine-identifiable despite looking normal to the naked eye.

The biggest implication of this is that AI-generated text can now be traced at scale, with a degree of certainty that was previously impossible.

This could eventually apply to school essays, corporate reports, news articles, research papers, software documentation, social media posts, and legal documents.

The significance lies not merely in AI detection but in provenance, reported Interesting Engineering.

If this eventually becomes standard industry practice, it will open a new frontier in AI detection. However, one has to keep in mind that heavy paraphrasing can remove existing watermarks from texts.

The biggest impact this will have is likely on education and social media. While it will be easier to verify whether social media content is AI-generated, when it comes to education, schools and universities will be able to say with a degree of certainty whether a student's submission has been generated by AI.