🤖Developer Cracks Claude Watermarking Within Hours
EU AI Act compliance just got a lot harder for Anthropic
TL;DR
Within four hours of Anthropic embedding watermarks into Claude models, developer Guillaume Meyer cracked the code. His workaround has gone viral on GitHub with over 20k bookmarks and multiple contributors.
Developer Guillaume Meyer published a workaround to remove Claude's AI-generated content watermarks within hours of their implementation. The EU AI Act requires model providers like Anthropic to label synthetic content or face fines up to 3% annual turnover, but independent developers can create tools to bypass these measures. This development raises questions about the enforceability and effectiveness of such regulations in a rapidly evolving tech landscape.
Key Points
Meyer's removal method uses non-watermarking large language models to generate multiple rewrites, swapping in synonyms and slightly reorganizing content.
The code has gone viral on GitHub with over 20,000 bookmarks and more than 100 contributors since its release.
Wayne Pan from Haimaker incorporated Meyer's open-source tool into his platform due to disagreement with Claude watermarking content.
SynthID-text underpinning Claude's watermarking makes it fairly certain that Meyer's method works, raising regulatory compliance questions.
Anthropic must include watermarks in all new models released from August and integrate them into existing models by December.
Why It Matters
If you're developing AI-generated content with Claude or similar models, the rapid spread of watermark bypass tools complicates EU compliance efforts. Developers like Wayne Pan are already integrating these workarounds into their platforms, challenging model providers' ability to enforce labeling requirements.
Frequently Asked Questions
Why does this matter?
If you're developing AI-generated content with Claude or similar models, the rapid spread of watermark bypass tools complicates EU compliance efforts. Developers like Wayne Pan are already integrating these workarounds into their platforms, challenging model providers' ability to enforce labeling requirements.
What happened?
Within four hours of Anthropic embedding watermarks into Claude models, developer Guillaume Meyer cracked the code. His workaround has gone viral on GitHub with over 20k bookmarks and multiple contributors.
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 3,263 builders reading daily.