Claude AI Watermarks Cracked Hours After Launch

Anthropic's new AI watermarks for its Claude platform were reportedly cracked just hours after their launch. The company introduced these invisible watermarks to meet upcoming EU transparency regulations. However, developers quickly shared methods online to strip or bypass these digital signatures. Anthropic had announced the system last week, aiming to embed watermarks into all content from its Claude language models. This was intended to help identify AI-generated text. But almost immediately, coders on GitHub and Reddit posted workarounds. Some methods involved simple text changes, while others exploited how watermarks degrade when content is paraphrased or translated by another AI model. One developer on GitHub even posted a script claiming a 95% success rate in removing the watermarks. This rapid circumvention highlights a core challenge in authenticating AI content. The EU AI Act requires technical solutions for identifying synthetic content to combat issues like deepfakes and disinformation. Companies could face large fines for non-compliance. This situation shows the difficulty of implementing effective content authentication in the fast-evolving world of AI.