Anthropic plans to add an invisible mark to AI text—as the industry scrambles to police AI slop
Fortune Beatrice Nolan ● Covered by 5 sources
Anthropic will start hiding a machine-readable mark in Claude text and images. It’s meant to help spot AI slop, and it may survive copy-paste and some edits.
Based on reporting by Fortune, Beatrice Nolan — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Anthropic is adding a subtle watermark to Claude’s output, starting with models released on or after August 2. The mark is invisible to people but machine-readable, and the company says it won’t hurt quality or readability. If you copy and paste the text, the signal is meant to travel with it. Some editing may not strip it out either, though a heavy rewrite or translation could.
This is not just a chatbot trick. Because the watermark lives at the model level, it shows up across Claude, the API, and tools like Claude Code. Anthropic is also watermarking images so users can tell Claude handled the file and whether it has been tampered with since. Older Claude models are excluded for now, though Anthropic says it is working on extending the marking.
The timing matters. The move lines up with the EU AI Act’s transparency rules, which kicked in on August 2 and require generative AI providers to make synthetic output machine-readable and detectable. That puts Anthropic in the middle of a broader industry scramble to label AI output before regulators, publishers, and annoyed readers do it for them.
Text is the hard case. Images have been easier to watermark; prose gets copied, paraphrased, translated, cut up, and folded into other writing all the time. Anthropic itself says the mark only shows Claude had a hand in something, not that Claude wrote the whole thing. Even a proofread or translation request can leave a trace.
And the company is walking into a messy debate. Substack is letting readers estimate how much of a post or comment is human versus AI-written. YouTube has tightened its rules on inauthentic content and generic, templated AI output, including AI personas giving advice in sensitive areas. But none of this is a clean fix. Anyone determined to hide AI text still has plenty of ways to break a statistical watermark, and a blunt “AI” label risks flattening everything from fake-news spam to a journalist cleaning up a transcript.
My take — AI-written commentary, not fact-checked reporting
This is the right direction, and also a lovely admission that the industry spent years shipping the mess first and thinking about labels later. A watermark won’t save the internet from synthetic sludge, but it does push the burden back toward the companies minting it. The awkward part is obvious: one mark, many uses, and not all of them deserve the same suspicion.
Read more about this at: Fortune