Anthropic says future Claude models will add invisible watermarks to AI-generated text as part of its compliance with new EU AI Act transparency rules.
The company explained that upcoming Claude models will generate text with a hidden signal that can help determine whether Claude was likely involved in writing it. The watermark will not be visible to readers and will not add secret text, hidden characters, unusual formatting, or metadata to ordinary written output.
Instead, Anthropic said the watermark is built into the way Claude chooses words.
When Claude generates a sentence, it often has to choose between words that are both acceptable. Anthropic said its watermarking method slightly changes the source of randomness used for those word choices. Over a longer passage, those small choices can form a pattern that a detector with the right key can check.
For readers, the text should look the same.
Anthropic said the watermark should not affect the meaning, quality, readability, creativity, speed, or cost of Claude’s responses. It also said the watermark does not identify a specific user, organization, or chat.
The move is tied to the EU AI Act, which requires AI providers serving the European market to mark AI-generated content. Anthropic said other major AI providers that signed the same Code of Practice are expected to use their own marking systems as well.
The company’s approach is based on a version of Google DeepMind’s SynthID-Text method, which was published in 2024.
Still, Anthropic made clear that the watermark is not a perfect AI detector.
A detected watermark may show that Claude likely processed part of a text, but it does not prove that a human did not write it. It also cannot detect text created by another AI model that uses a different marking method.
The system also works better on longer passages. Short text, factual writing, proofreading, and code may carry weaker signals because Claude has less room to choose between different words without changing accuracy.
That distinction could matter for students, writers, developers, companies, and communications teams that use Claude for editing, translation, formatting, or drafting.
If Claude only fixes grammar or punctuation in a human-written passage, most of the words may still come from the user. In that case, the final text may carry little watermark signal. But if Claude rewrites or translates a full passage, more of the output may carry Claude’s mark.
Claude-generated files will be handled differently.
Anthropic said supported files such as PNG, JPG, and SVG will use C2PA content credentials. That means a file can include signed metadata showing that Claude was involved in creating or processing it. This is separate from the invisible text watermark.
The rollout will apply globally to supported Claude models wherever Claude is offered. That includes Claude, Claude Platform, Claude Code, Claude Cowork, Claude Tag, and supported cloud partner access through AWS, Google Cloud, and Microsoft Foundry.
New Claude models launched in the EU on or after August 2, 2026 will support marking at launch. Anthropic said older models are also being updated.
The Verge noted that file metadata can sometimes be stripped when content is edited or uploaded to other platforms, which means provenance labels may not always survive.
Axios raised another practical concern for communications teams: content that was only proofread, translated, or formatted by Claude could still carry a signal that Claude processed it.
Business Insider reported that some Claude users have questioned whether watermarked AI-assisted work could be judged differently in academic, professional, or client settings. Anthropic said a detected watermark only indicates that Claude may have processed the content and does not change a user’s ownership or authorship rights under its terms.
Anthropic said it is also working on a watermark detection API so users and third parties can check whether text likely contains Claude’s watermark.
For now, the takeaway is simple: future Claude text may carry an invisible signal, but readers will not see it.
A light edit may not remove the watermark. A complete rewrite likely will. And a missing watermark does not prove that a piece of content was written by a human.
As AI watermarking becomes part of EU AI Act compliance, more major AI companies are expected to move toward machine-readable labels for AI-generated content.
via: Anthropic | Business Insider
