Developers are already testing ways around Claude’s invisible AI-content watermarks, according to WIRED, just days after Anthropic said its models would embed machine-readable signals into generated content for European Union AI Act compliance. WIRED reports that developer Guillaume Meyer published an override within four hours of Anthropic confirming that Claude models would globally watermark AI-generated content. The code has since spread quickly: WIRED says it has been bookmarked more than 20,000 times on X and attracted more than 100 contributors on GitHub, with other projects incorporating the approach. The immediate response exposes the practical tension behind text watermarking. The EU rules, which WIRED says came into force earlier this month, require providers such as Anthropic and OpenAI to label synthetic audio, image, video, or text so that machines can detect it as AI-generated. The reported penalty for noncompliance can reach up to 3% of annual turnover. WIRED also notes an important boundary: providers cannot market circumvention tools, but the rules do not bar independent tools. Anthropic’s watermark is not a visible label. WIRED describes it as a pattern in Claude’s word and phrase choices that a human reader would not notice but that a detector could identify if it knows what to look for. Some users worry that shaping model output this way could affect response quality, while Anthropic says that will not be the case. The technique is called SynthID, according to WIRED. It was developed by Google, which has used it to watermark AI-generated content since 2023. WIRED also reports that computer scientist Scott Aaronson proposed a similar method while working at OpenAI, but that OpenAI did not deploy it because of concerns that watermarking would deter customers. Meyer’s stated objections are partly technical and partly social. WIRED reports that he is not opposed to transparency or attribution, but worries that watermarking can create false positives or fail to distinguish between light editing assistance and heavy AI generation. As an example, WIRED notes that Meyer, a native French speaker, uses tools including Claude and Grammarly to edit his writing. Those concerns matter because detection systems can be used outside the narrow compliance setting. Meyer told WIRED that employers could reject candidates or researchers could face inflated accusations if a detector flags text as likely touched by Claude. WIRED also reports that freelance content writers and social media creators have contacted Meyer for help using the code. The technical details remain early in this cluster. WIRED says Meyer’s removal method uses a non-watermarking large language model to produce multiple rewrites, but the evidence here does not independently validate the tool’s reliability across Claude outputs or future Anthropic changes. The strongest supported fact is narrower: public developer workarounds appeared rapidly and drew significant attention. Who benefits: Users who want to avoid AI-generated-content labels may benefit from public circumvention code. Developers researching watermark detection and removal also gain a visible test case. Who's exposed: Model providers relying on invisible text watermarking are exposed to rapid public attempts at removal. Employers, schools, or research institutions that treat detector outputs as decisive evidence are also exposed to the false-positive concerns raised in WIRED’s report.