Anthropic to Add Invisible Watermarks to Claude-Generated Text

Anthropic says it will add invisible, machine-readable watermarks to text generated by Claude, Claude Code, Claude Cowork, Claude Tag, and its API, as well as to supported Claude options through AWS, Google Cloud, and Microsoft Foundry.

The company says supported models will embed the marks directly in their text where readers cannot see them, similar to Google’s SynthID. It also claims that these watermarks will “stick” to copied text and can survive some editing.

Anthropic ties the policy to Article 50(2) of the European Union’s AI Act and its Code of Practice on Transparency of AI-Generated Content. The EU requires that generative AI systems use tools to mark and detect synthetic content when it’s feasible. The obligations began on August 2, 2026, although older systems have until December 2026 to meet this requirement.

Credit: Bloomberg/Getty Images

Going forward, models launching in the EU will watermark content from their first day; meanwhile, Anthropic is working to retrofit watermarking technology for older models.

Interestingly, a detected watermark doesn’t necessarily indicate that Claude wrote that particular content, because Claude will even mark text after someone asks it to translate, summarize, or edit their work. On the other hand, Anthropic says that a watermark-free piece of writing does not prove human authorship or confirm that no AI assisted the work. Using heavy edits, paraphrasing, translations, or mixing material can avoid detection. Similarly, re-saving, conversion, and screenshots can also remove file metadata, as reported by TechCrunch.

Anthropic has not published the technical method for the text watermark, its detection rates, or a public detection tool. It makes me wonder how long it would take for someone to use Claude AI itself to develop a method to remove hidden watermarks, or even to maliciously add one to an original text.