A Viral Open-Source Tool Built to Strip AI Watermarks
Guillaume Meyer, a Paris-based technology entrepreneur, responded to Anthropic's recent decision to add invisible AI watermarks by building a tool that strips them. He published the first version of "Watermarks Remover" on GitHub within roughly five hours, drawing on years of work with open-source AI models.
The project exploded after Meyer shared it on X on August 11, generating more than two million impressions and spreading across LinkedIn and other networks. The tool works by checking whether a piece of text or an image carries the statistical pattern used for watermarks, then generating subtle variations that preserve the meaning while disrupting the mark. It repeats this process until the watermark is no longer detected.
Meyer says he is not opposed to content attribution. His concern is that the statistical technique Anthropic and others are using can produce false positives. As a non-native English speaker who relies on Grammarly's AI-powered proofreading throughout the day, he argues that a watermark tied to AI assistance could label everything he writes as AI-generated.
The project remains a technical, fast-moving experiment. Meyer acknowledges it is not a finished product and expects adjustments will be needed when Anthropic and other vendors release detectors.
Why a Paris Founder's Workaround Exposes the Limits of AI Watermarking
Why Meyer Argues Anthropic's Watermark Is the Wrong Fix
Meyer's central objection is not to transparency but to the method. Text watermarks based on statistical word-choice patterns are inherently probabilistic. That means they can flag human writing, especially when an author uses AI-assisted proofreading, translation or editing. In his case, Grammarly runs throughout his writing process, so a system that labels AI-assisted text could mislabel all his work.
How the Remover Exploits a Structural Weakness
The tool does not claim to crack Anthropic's exact algorithm. Instead, it iterates: it checks for a watermark, introduces small semantic-preserving changes, and checks again until the signal disappears. Applied to images, the same logic operates on pixels rather than words. This approach works against statistical watermarks more broadly, which is why the first version arrived so quickly.
The False-Positive Problem Could Outrun the Watermark's Benefits
The risks are most concrete for researchers, students and professionals who use AI writing assistance on a portion of a document. Meyer points to a scenario in which a researcher uses AI to change one line at the end of a ten-page paper and the entire document gets labeled AI-generated. For trust systems meant to identify disinformation, that kind of error could undermine the very confidence they are meant to protect.
Where Open Source Leaves the Conversation
The viral response, which brought contributors and overwhelmingly positive feedback alongside a small share of criticism, suggests wider discomfort with AI watermarks than the technical debate alone. The project's fast spread also signals a practical reality: once a watermark scheme is public and detector logic can be inferred, countermeasures can appear quickly.
What the Watermark Fight Means for AI Labels, Rules and Users
For companies, researchers and users watching the AI provenance debate, the project offers a clear set of practical signals:
- Treat statistical text watermarks as a screening signal, not proof. Meyer's tool disrupts the signal by iterating through small meaning-preserving changes, so any attribution system built on statistical patterns alone can be bypassed cheaply.
- Separate AI-assisted editing from AI-generated content. Meyer's Grammarly example shows that a binary watermark that flags AI-assisted text would mislabel routine professional writing, not just synthetic disinformation.
- Plan for detector updates, not a one-time fix. Meyer explicitly says the project will need adjustments when Anthropic releases detectors, meaning both sides of the watermark race are still evolving.
- Regulators and platform operators should assess whether watermark mandates can meet their own false-positive tolerances. The tool's five-hour build time suggests the technical barrier to removal is low, not months of specialized effort.
- For the open-source community, expect the legal picture to remain unsettled. Meyer says he may commercialize a simpler version but is not certain whether that is legally possible, making legal clarity a near-term question rather than a settled rule.
Risk & Opportunity Assessment
| Commercial Risk | Medium | Open-source removal tools could undermine the commercial value of AI watermarking and provenance products, weakening the assurance Anthropic's planned watermark is meant to provide. |
| Competitive Risk | Medium | Anthropic and other vendors such as Gemini already using statistical watermarks face a common bypass; the low entry barrier reduces any first-mover advantage in watermark technology. |
| Regulatory Risk | Medium | If regulators rely on statistical watermarks for AI transparency rules, a readily available remover and the threat of false positives could undermine compliance and trust. |
| Reputation Risk | Medium | The creator reports 99.9% positive feedback but also criticism from people who want reliable AI detection, and he must repeatedly explain that the intent is not impersonation or theft. |
| Technology Disruption | High | An effective countermeasure emerged within hours of the watermark announcement, showing that statistical watermark detection can be disrupted quickly by iterative semantic-preserving changes. |
| Commercial Opportunity | Medium | Meyer is weighing commercialization of a simpler version of the remover, and the viral traction suggests demand, though legal and usability questions remain unresolved. |
Comments 0