Anthropic will watermark text generated by its AI models, including Claude, the company has said.
The move comes as Anthropic seeks to comply with the EU AI Act’s Transparency Code which took effect earlier this month.
Despite the change coming off the back of an EU code, it will roll out globally, including in Australia.
The new EU Code requires AI companies to mark AI-generated or edited content in a way other systems can identify them.
Anthropic has said that all models released after 2 August will automatically have tech to watermark both computer-generated text and files.
Through means that we can only begin to fathom, the watermark will travel with text when users copy and paste it.
The company will extend support to older models, too.
“When a supported Claude model generates text, it weaves an imperceptible watermark directly into the text itself. You won’t see it, and it doesn’t change the meaning, quality, or readability of Claude’s response,” the company wrote in on a support page.
“Because the watermark is part of the text, it will travel with the text when it’s copied and pasted elsewhere, and may persist through some editing. Watermarking will be applied at the model level, which means it will be present no matter which Claude product or surface the text comes from.
“We’re also working to enable users and other third parties to detect Claude’s embedded watermarks and provenance metadata. Detection checks whether a piece of text or a file carries a supported Claude mark. If a supported mark is found, it indicates that the content may have been processed by Claude.”
The other large AI firms, including Google, Meta, Microsoft and OpenAI, have also committed to complying with the code.
It isn’t clear how much editing the copy will need by users to remove the watermark.
What is clear, however, is that it’s very bad news for many, many professionals.
For publishers — including B&T — the move is to an extent confounding and troubling.
Will watermarked text be readable and potentially penalised by search engines?
What protections are in place to prevent watermarked copy generated by Claude and other models being viewed as the source, rather than a synthesising of other copyrighted materials generated and shared online?
What safeguards are in place for publishers when copying and pasting information from press releases (standard practice for many)?
Much to ponder. Perhaps you could ask Claude?
