Anthropic’s attempt to make AI-written text easier to identify has barely begun, yet developers are already racing to make its markers disappear. The clash pits transparency advocates against users who fear a hidden label could follow even lightly edited human work.

The company began embedding an “imperceptible watermark” in text from supported Claude models released since August 2, saying the statistical pattern in word choice would persist through copying and pasting. Anthropic has tied the move to its commitments under the EU AI Act, while supporters see such labeling as a needed check on synthetic content.¹

Within days, Paris tech founder Guillaume Meyer released the open-source Watermarks Remover project. He says it strips hidden data and repeatedly rewrites text while preserving its meaning, disrupting the statistical signal behind a watermark. The first version took roughly five hours; the project later drew more than 14,000 GitHub stars, though its effectiveness is not guaranteed.¹

Meyer’s case is that watermarking mistakes assistance for authorship. “I am all for content attribution,” he said. “I am against the watermarking technique, and that's a very significant distinction.”¹ He has warned that a writer using AI merely to proofread, translate or change a single line could still be tainted by a label. “I think it’s the wrong answer to a real problem,” he said.²

Other developers quickly followed with browser-based and Claude-focused removal tools, reflecting rising demand but also the fragility of the underlying approach. ETH Zurich researcher Thibaud Gloaguen put the technical reality plainly: “There will always be ways to remove the watermark,” including rewriting the text.¹

The legal picture remains unsettled. EU guidance tells AI providers to make labels resilient against removal and modification, but does not explicitly bar third parties from building scrubbers. The sharper risk lies with users who deploy them to pass AI output off as human work—a practice Anthropic’s policy prohibits.¹