Anthropic is taking heat from developers over a plan to stamp invisible watermarks into the code and text that its Claude models generate.

The pushback followed a post on Tuesday by Thariq Shihipar, a member of the technical staff on Anthropic's Claude Code team, who confirmed the feature and framed it as a compliance move.

"This is part of working with the EU AI Act, other labs are adding similar watermarking," he wrote on X, in a post that has drawn more than 365,000 views. "It's hard to identify AI-generated text, and this gives people better tools to do that."

Anthropic says it signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content, and lays out the mechanics in a help-center article. The company describes the marking as imperceptible by design. When a supported model generates text, it "weaves an imperceptible watermark directly into the text itself," one that travels when the text is copied and pasted and can survive light editing. Images get separate signed metadata.

That an unseen artifact would be threaded through working code is exactly what set some developers off. "Claude adding invisible watermarks? inside MY codebase??? total b*llshit," the developer Nick Dobos wrote in a post that drew more than 105,000 views and over a thousand likes. "Diabolical precedent to be setting." Others worried that a hidden signal woven into their text or code could quietly degrade its quality.

The backlash over watermarking touches the recent brouhaha raging through Silicon Valley and Washington over whether U.S. policy should promote open-source AI models – programs that are available to anyone to download, inspect and run – or if they should be kept proprietary and in some cases subject to review by government officials. Most big tech firms have signaled a preference for open-source, but Anthropic, whose Claude is a closed-source model, has been a notable holdout.

Analysts have begun asking whether open weights make text watermarking useless altogether. The presence of a secretive mechanism within one of the most popular AI platforms could, by itself, prompt some users to route their queries to other providers, such as decentralized computing networks that could generate content and code using open-source models.

Watermark-strippers?

Developers have already answered with countermeasures, some perhaps more tongue-in-cheek than bona fide. At least two experimental tools have surfaced on the code-sharing site GitHub that purport to perturb and strip Claude's statistical watermark, one pitched as a multi-vendor remover of AI provenance marks, another as a critical demonstration that the mark can be defeated by simply rewriting the text. Both concede the job is best-effort without the vendor's own detector.

The thread also turned on Shihipar's premise that AI text is hard to spot, with one commenter wryly noting that it's pretty obvious "when you've got slop written all over" it.

Anthropic acknowledges that the presence of a watermark, or lack thereof, could send false signals.

According to Anthropic's documentation, a detected watermark "provides a signal that content was processed by Claude, but is not fully conclusive," and a missing mark does not mean content was human-written.

The watermarking applies to Claude models launched after Aug. 2 across Anthropic's products, and the company says it will ship a text-detection API.

Under Anthropic's plan, the coding assistant doing the writing does not know it is leaving the traces.

"Agent will not be aware," Shihipar said, pointing to Google's SynthID as a model for how the technology works. He added that developers could use the marks to "check if a PR was generated by Claude Code." PR stands for "pull request," which is a snippet of code changes that developers are proposing to make to a computer program.

HOW AI WAS USED IN THE PRODUCTION OF THIS PIECE: The first draft of this story was written on Claude Desktop using a custom skill designed to turn company announcements on X into news stories, and filed directly to our DistroVerse web app using our Distro Publisher MCP connector. I then did a full edit on our built-in (non-AI) DistroVerse story-editing interface, and I personally checked all the facts prior to publication. I used ChatGPT in a custom project to create the story illustration in SVG format.