Coders Say They Can Already Strip Claude’s New Text Watermark, but Nobody Can Prove It Yet

Updated Oct. 26, 2026, 6:15 a.m. ET: Adds Anthropic’s detection API, now in private preview.

Guillaume Meyer needed about four hours. That is how long it took, WIRED reported on Wednesday, between Anthropic confirming that Claude would embed invisible, machine-readable watermarks in its output and Meyer publishing code to remove them. The tool has since been bookmarked more than 20,000 times on X and drawn more than 100 contributors on GitHub.

Anthropic is adding the marks to comply with the EU AI Act, whose labeling rules for synthetic content took effect this month. In an explainer, the company says it uses a version of SynthID-Text, the method Google DeepMind published in Nature in 2024. Rather than adding hidden characters, the watermark changes the source of randomness Claude uses when choosing between equally good words, using a key that a detector can later check. Anthropic says it saw no effect on quality or readability in internal testing, and that the result is only a probability that Claude was involved.

How the workarounds operate

Meyer’s method runs Claude’s text through a non-watermarking model to produce several rewrites, swapping synonyms and reshuffling sentences. Software engineer Erik Hughes, according to WIRED, took 15 minutes to build his own tool with Claude; it strips invisible and look-alike characters, reorders sentences within paragraphs and replaces some words. Leon Chlon, a visiting fellow at the University of Oxford, said condensing a response, translating it into Arabic and back can also wipe the mark.

None of this has been verified. Anthropic had not released its detector, and Wayne Pan, CTO and co-founder of Haimaker, said his confidence in Meyer’s approach comes from understanding how SynthID-Text works. Anthropic itself says heavy editing, paraphrasing and translation may leave no detectable watermark, and that short passages carry too little signal to judge.

Code is a smaller target than prose. Because the watermark only nudges choices where any option is equally valid, Anthropic says code “has generally less watermarking than some other forms of text,” with comments the likeliest place for it to show up.

Why some developers object

Meyer told WIRED he isn’t against attribution. “I just think watermarking in itself is a really bad solution, because it has major drawbacks and risks,” he said. As a native French speaker who uses Claude and Grammarly to edit his writing, he worries a detector won’t separate light help from wholesale generation, and that employers or reviewers could treat a probability as proof.

Anthropic’s spokesperson told WIRED that “text from supported Claude models, including output from Claude Code, will carry an invisible watermark,” and that the company planned a text-detection API.

That API now exists in limited form. Anthropic’s updated explainer says the detection API is in private preview for organizations the EU law covers, such as regulators, media, fact-checkers, researchers and enterprises with their own compliance obligations, with wider access planned over time. Developers who want to test whether the removal tools work will have to get on that list first.