> ## Content Index
> Fetch the complete content index at: https://www.metatalks.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Four hours after Anthropic confirmed Claude's hidden watermarks, a developer published code to strip them out
- URL: https://www.metatalks.ai/four-hours-after-anthropic-confirmed-claudes-hidden-watermarks-a-developer-published-code-to-strip-them-out/
- Published: 2026-08-22T14:33:00.000Z
- Updated: 2026-08-22T14:32:59.000Z
- Author: Al
- Tags: News, AI Detection, #newswire

**Coders have piled into developer Guillaume Meyer's project and built bypasses of their own. Whether any of them beats the mark, Wired reports, cannot be checked until Anthropic ships the detector it has yet to release.**

Meyer published code meant to strip the hidden watermarks out of Claude's text within four hours of Anthropic confirming that its Claude models would, across the board, embed machine-readable markers in what they produce.

The code spread on GitHub, has been bookmarked more than 20,000 times on X and has drawn more than 100 contributors.

The marker it goes after runs through the words and phrasings Claude chooses — a signature no person reading the text can notice, though a machine trained to look for it can. Google built the technique, known as [SynthID](https://deepmind.google/models/synthid/?ref=metatalks.ai), and has marked its own AI output with it since 2023\. Anthropic said earlier this month that Claude would take up watermarking to meet the requirements of the EU AI Act, and a company spokesperson told Wired that other labs are moving in the same direction.

The rules came into force this month and carry penalties of up to 3 percent of yearly turnover. They bar providers from marketing tools that circumvent the labeling; independent developers who build them face no legal restriction.

Several programmers have arrived at bypasses by separate routes. Meyer runs the text through a large language model that adds no watermarks of its own, generating several rewrites that swap in synonyms and shuffle the material a little — a wager on rival models continuing to leave their output unmarked, which may not be a safe one: 190 organizations, OpenAI, Microsoft and Meta among them, have signed the EU's code of practice on transparency.

Software engineer Erik Hughes spent 15 minutes, with Claude assisting, on a tool that removes hidden and look-alike characters, shuffles the order of sentences inside a paragraph and replaces a handful of words with synonyms. Leon Chlon, who holds a visiting fellowship at the University of Oxford, told Wired that another route is to shorten what Claude wrote, translate it into a language far removed from English, such as Arabic, and then translate it back.

Anthropic has acknowledged that material reworked substantially, reworded or put through translation may end up without a watermark a detector can find. No one can confirm that the removal tools do the job while the company's detection software remains unreleased. Wayne Pan, chief technology officer and cofounder of the Silicon Valley sovereign AI startup Haimaker, told Wired that understanding the basic SynthID-text approach behind Claude's watermarking leaves him fairly confident the removal method works, adding that he does not believe a watermark can ever withstand every attempt to strip it.

Meyer is not against transparency and supports content attribution, he says; his objection is that watermarking is a poor fix with serious drawbacks and risks, among them false positives and a mark that may not separate a light edit from heavy AI use. Treating it as proof troubles him, since Anthropic itself grants that the system yields no more than a likelihood that Claude had a hand in a text — enough, he cautions, to see applicants turned down unjustly or researchers hit with inflated allegations. Freelance content writers and social media creators have come to Meyer, [Wired](https://www.wired.com/story/coders-say-they-already-found-workarounds-to-claudes-invisible-watermarks/?ref=metatalks.ai) reported, for help putting the code to use.

Anthropic says it is still working out how detection of text watermarks should be built and means to put out a tool before long, while work to refine the watermarking itself continues.