Anthropic’s plan to mark Claude-generated text has quickly turned into a test of how durable invisible AI watermarking can be. Soon after the company confirmed that Claude models would globally embed machine-readable watermarks into AI-generated content, developers began publishing ways they believe can remove them.
The result is an early clash between AI transparency requirements and the practical limits of technical enforcement. The core question is simple: if a watermark is hidden inside word choices and phrasing, what happens when the text is rewritten?
A fast response from developers
Within four hours of Anthropic confirming the watermarking plan, developer Guillaume Meyer had published an override. His code, designed to remove watermarks from Claude-generated text, has since spread widely on GitHub.
According to the source article, the project has been bookmarked more than 20,000 times on X and has attracted more than 100 contributors. Many others have also started incorporating the technology into their own projects.
Meyer told WIRED that people are approaching the issue for different reasons. Some oppose the idea that all AI-generated content should be labeled that way. Others, including Meyer, are interested in the technical challenge itself.
The audience for these tools is not limited to programmers. Meyer says freelance content writers and social media creators have contacted him for help using the code.
Why Claude is adding watermarks
Anthropic announced last week that Claude would adopt watermarking to comply with the European Union’s AI Act. The new rules, which came in earlier this month, require providers such as Anthropic and OpenAI to label synthetic audio, image, video, or text so machines can detect that it is AI-generated.
Providers that do not comply can face fines of up to 3 percent of annual turnover. The rules also say providers cannot market circumvention tools, but they do not place the same legal restriction on independent tools.
That distinction matters because the workaround ecosystem can grow outside the companies that are required to label synthetic content. Anthropic can add marks to Claude output, but independent developers can still try to build methods that strip those signals away.
Meyer says he supports transparency and content attribution, but does not see watermarking as the right approach. He is concerned about false positives and about systems that may fail to distinguish between light AI assistance and heavy AI generation.
That concern is personal as well as technical. As a native French speaker, Meyer says he often uses Claude and other AI tools like Grammarly to edit his writing. In his view, a detector that only produces a probability could be misused by employers or researchers if treated as stronger evidence than it is.
How the watermark can be weakened
Anthropic’s watermark is invisible to a human reader. It works by placing a detectable pattern in Claude’s word and phrase choices. A machine that knows what to look for can identify the pattern, while a person reading the text would not see anything unusual.
The technique is called SynthID. It was developed by Google, which has used it to watermark AI-generated content since 2023. Computer scientist Scott Aaronson proposed a similar method while working at OpenAI, but says the firm did not deploy it because it was concerned watermarks would put customers off the product.
Meyer’s removal method uses a large language model that does not add watermarks. It generates multiple rewrites, swaps in synonyms, and slightly reorganizes content. The logic is that changing enough of the wording may disturb the pattern that made the watermark detectable.
That approach depends on access to models that do not apply their own watermarks. The source article notes that this may not remain a safe assumption, because 190 organizations, including OpenAI, Microsoft, and Meta, have signed the EU’s transparency code of practice.
It is still unclear how many laboratories will implement watermarks. The source article says watermarks must be included in all new models released from August and integrated into existing models by December.
Other workarounds raise the same issue
Meyer is not the only developer testing the limits of Claude watermarking. Software engineer Erik Hughes built a tool with Claude in 15 minutes that removes invisible and look-alike characters, reorders sentences within paragraphs, and replaces several words with synonyms.
Leon Chlon, a Visiting Fellow at the University of Oxford, says watermarks can also be removed by condensing Claude’s response, translating it into a dialect like Arabic, and then translating it back. The point is not that every method has been proven against Anthropic’s own detector, but that major edits can change the signals a detector relies on.
Anthropic itself acknowledged that heavily edited, paraphrased, or translated content might not carry a watermark. That admission places an important limit on what watermarking can promise: it may help identify unmodified or lightly changed AI output, but it may not survive substantial transformation.
Wayne Pan, chief technology and cofounder at Silicon Valley-based sovereign AI startup Haimaker, incorporated Meyer’s open-source tool into his platform. He said there is no certainty the tool works until Anthropic releases its detection software, but that the basic SynthID-text approach gives developers reason to think the method may work.
The test is still coming
Anthropic told WIRED that it is adding marking to Claude’s output to comply with the EU AI Act and that other labs are taking similar steps. The company also said text from supported Claude models, including Claude Code, will carry an invisible watermark, and that the mark does not change the meaning, quality, or readability of Claude’s responses.
The company also plans to ship a text-detection API, which would let users do more of this detection themselves. Anthropic says it is still working out how to implement watermark detection for text and plans to release a tool soon.
That release will be the meaningful test. Until developers can compare their workarounds against Anthropic’s own detector, claims about removing Claude watermarks remain partly theoretical.
Still, the early response shows why text watermarking is difficult. Unlike an obvious label, an invisible pattern must survive the ordinary ways people edit, translate, condense, and rewrite language. If those changes are enough to break detection, watermarking may be useful as one signal, but not as a complete answer to AI content attribution.