Claude can now leave an invisible mark on your writing, even if you wrote it yourself

Anthropic has introduced a new system that embeds invisible watermarks directly into text generated by Claude. Unlike metadata attached to a document or image, the new marking system is designed to follow generated text when it is copied and pasted elsewhere and can even persist through some subsequent editing.

The European Union can take at least some credit, or blame, for the development. Anthropic says it has signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content as a provider of both generative AI models and systems. Its new marking scheme is how the company currently intends to put those commitments into practice, although the company is deploying it worldwide rather than confining it to the bloc responsible for the regulatory impetus.

The company says all Claude models launched on or after August 2, 2026, support the technology from launch. Anthropic also intends to extend it to older models over time. Nor is the system confined to markets where AI-content labelling requirements might make such measures necessary, as the company is deploying it worldwide.

Anthropic isn't saying how the watermark works

According to Anthropic, the watermark is woven directly into Claude’s text output without visibly altering the resulting response. This makes it fundamentally different from conventional provenance metadata, which can simply disappear the moment content is liberated from its original file.

Precisely how Anthropic accomplishes this remains something of a mystery. The company has heretofore declined to publish detailed technical documentation explaining how the text watermark is embedded or detected. It nevertheless maintains that the marking does not affect the “meaning, quality, or style” of Claude’s responses.

The system also extends considerably beyond the consumer-facing Claude chatbot. Anthropic says supported models produce marked text when accessed through Claude, Claude Code, Cowork, its API and third-party cloud platforms. Developers employing Claude models via non-native means are therefore not exempt from the new provenance regime.

Anthropic eventually plans to make detection mechanisms available to third parties, although it has yet to divulge how these tools will work or, indeed, who will be permitted to use them.

Ironically, a Claude watermark doesn't prove Claude wrote it

There is, however, a rather substantial caveat. Detecting Claude’s watermark does not necessarily mean Claude actually wrote the underlying material. Anything that passes through Claude can potentially acquire the mark, making its presence evidence of Claude's involvement rather than proof of AI authorship.

Anthropic itself acknowledges that entirely human-written text can acquire a watermark after being processed by Claude. An author could, for instance, write an entire article themselves and use Claude merely to proofread, translate or edit it. The resulting copy may nevertheless emerge bearing Claude’s invisible mark.

Anyone determined to rid themselves of said watermark will have to resort to extensive editing, paraphrasing, translation or combining Claude output with other material, effectively undermining its purpose to begin with. It also seems inevitable that the technology will spawn a plethora of cottage industries dedicated to excising the dreaded watermark from Claude-generated text.

The resulting situation is therefore somewhat peculiar. A positive result may indicate that Claude processed a piece of text at some point, but it cannot conclusively establish that Claude wrote it. Conversely, the absence of a detectable watermark does not necessarily prove that Claude was never involved. 

Will this actually reduce AI slop? Probably not

That essentially brings us back to square one because the presence, or lack thereof, of the Claude watermark proves nothing. The AI-slop-ification, it appears, will go on unhindered, at least for the foreseeable future. 

It does, however, set a worrying precedent for other AI companies, who will undoubtedly follow Anthropic and try to bake in provenance mechanisms in their AI outputs. "We don't taint our output with invisible markings", therefore, will become a legitimate marketing slogan for AI companies in the coming future. 

Images and certain other files generated using Claude are handled differently. Anthropic uses C2PA Content Credentials for supported non-text output, attaching cryptographically signed provenance information to generated files rather than subjecting them to the same text-watermarking technique.

For now, perhaps the biggest unanswered question is what Claude’s invisible text marking actually consists of. Should the system influence token selection during generation, its implementation and potential effect on model output could prove considerably more interesting than conventional metadata-based provenance.

Anthropic says additional technical documentation concerning its watermarking approach will be published in the future. Until then, precisely what Claude is doing behind the curtain remains unknown.