Copy a response from Claude into another app and traces of the AI tool’s involvement may follow the text.
Anthropic’s new marking system took effect August 2 as EU AI Act transparency requirements for AI-generated content came into force. European regulation prompted the change, but the company says supported models use the markings worldwide.
Text and file outputs are marked in different ways, affecting what other systems can detect later.
Text watermarks and file provenance use separate methods
According to Anthropic, supported models add an imperceptible watermark to generated text. Readers cannot see it, but compatible systems can potentially detect the hidden pattern.
Supported file outputs can instead carry digitally signed C2PA metadata, creating a provenance record without requiring a visible label.
Across the company’s API and services such as Claude Tag in Slack, supported models can carry the markings outside the main chatbot. Easier ways to check them are still coming, with the AI company saying detection tools and additional technical guidance remain in development.
AI markings can weaken outside the chatbot
Copying marked text into another application may leave the watermark intact, but durability drops as the wording changes. Anthropic says substantial rewriting or translation can make detection harder or impossible. File-based provenance depends on attached metadata, which may disappear during editing or conversion, or when software strips it.
Even a successful detection does not settle authorship. Human-written material sent through the AI tool for editing or translation can come back marked, while heavily revised AI-generated text may no longer be detectable. A positive result can indicate that Claude processed the content without proving it wrote every word, while an unmarked document cannot be treated as proof of human authorship.
AI provenance complicates review and disclosure
People submitting professional or academic work should keep a record of where AI entered their workflow and follow any disclosure requirements attached to the work. Keeping documentation gives them something to point to if an organization later questions how the material was produced.
Organizations reviewing submitted or published content need policies that account for partial AI use and inconclusive detection. A detected marker should prompt a closer look at how the tool was used, while AI-use policies should spell out what requires disclosure and how disputed findings will be handled.
Businesses moving AI output between applications should test whether model integrations preserve provenance and document processing steps that can remove it. Multinational teams should also watch EU transparency requirements because vendor changes made for Europe can reach workflows elsewhere.
As checking tools improve, organizations will have to decide how much weight a marker deserves when authorship or disclosure is disputed.
Also read: US grid data suggests AI data centers are creating concentrated power pressure in certain regions instead of overwhelming the entire system.
