Anthropic confirmed last week that every piece of text generated by its Claude family of large language models will now contain an invisible, machine‑detectable watermark. The move satisfies a requirement in the European Union’s AI Act, which obliges model providers to label synthetic audio, image, video or text so that regulators and downstream users can identify AI‑generated material. Non‑compliance carries fines of up to 3 percent of a company’s annual turnover.
Only a few hours after the announcement, French developer Guillaume Meyer posted a Python script that strips the watermark from Claude‑generated text. The repository, hosted on GitHub, gathered more than 20,000 bookmarks on X and attracted over 100 contributors who adapted the code for their own projects. "Anthropic is embedding watermarks in its Claude texts … the issue is practically history just one day later," an AI specialist wrote on social media, attaching a meme of Meyer breaking chains.
Open‑source workaround spreads
Meyer says he was motivated by curiosity and a belief that watermarking is a flawed solution, not by a desire to subvert transparency. "I'm not against content attribution," he told WIRED, "but the watermark introduces major drawbacks and risks." The script works by feeding Claude’s output into a separate, non‑watermarked language model, which rewrites the passage with synonyms and minor re‑structuring. The process yields text that reads naturally while eliminating the hidden pattern that the watermark relies on.
Other engineers have followed suit. Software engineer Erik Hughes built a 15‑minute tool that removes invisible characters, reorders sentences and swaps words for synonyms. Visiting Fellow at Oxford University, Leon Chlon, demonstrated that translating Claude’s response into a linguistically distant dialect—such as Arabic—and back can also erase the watermark, a technique Anthropic acknowledges in its compliance documentation.
The rapid proliferation of these tools has drawn attention from freelancers, social‑media creators and startup founders who rely on Claude for content generation. Some have reached out to Meyer for assistance, hoping to avoid the watermark in client deliverables. Others view the work as a technical challenge, eager to test the limits of the AI Act’s enforcement mechanisms.
Anthropic responded to inquiries with a statement emphasizing compliance and user choice. "Text from supported Claude models, including Claude Code, will carry an invisible watermark that does not change meaning, quality or readability," the company said. It also promised a text‑detection API that developers can integrate into their workflows to verify whether a piece of content bears the watermark.
Critics argue that the watermark’s very invisibility may degrade Claude’s output. By nudging the model toward specific word choices to embed the pattern, the watermark could subtly steer responses away from optimal phrasing. Wayne Pan, co‑founder of AI startup Haimaker, warned that “you can’t have a watermark that will withstand everything.” He incorporated Meyer’s open‑source tool into his platform, citing concerns that heavily edited or paraphrased text might escape detection, undermining the purpose of the regulation.
Legal experts note that while the EU rules prohibit providers from offering circumvention tools, they do not forbid independent developers from creating them. This loophole leaves the market open to a cat‑and‑mouse dynamic: regulators push for transparent labeling, while the developer community builds methods to bypass it.
As the deadline for integrating watermarks into new models approaches in August, and existing models must be updated by December, the pressure is on AI labs to refine their detection and labeling systems. Whether the industry can reach a consensus on a solution that balances transparency, usability and technical feasibility remains uncertain.
Cet article a été rédigé avec l'assistance de l'IA.
News Factory APP - actualités agentiques pour booster votre SEO et AEO.