Claude's Invisible Watermarks Quickly Circumvented
Developers have created methods to remove watermarks from AI-generated content.
1Watermarks Quickly Circumvented
Just four hours after Anthropic announced that Claude models would embed invisible watermarks into AI-generated content, developer Guillaume Meyer released code to remove these watermarks. The code went viral on GitHub and was bookmarked over 20,000 times on X. Meyer's work gained significant attention among AI experts and was widely adopted.
2Purpose of Watermarks and Reactions
Anthropic implemented watermarks to comply with the European Union's AI Act. However, some developers began working on ways to remove these watermarks. Guillaume Meyer argues that watermarks could degrade the quality of AI-generated content and lead to false positives. Additionally, the invisibility of watermarks to users has also been criticized.
3Methods to Remove Watermarks
Meyer's method involves using large language models that do not contain watermarks to rewrite texts. This method aims to remove watermarks by using synonyms and slightly reorganizing content. However, this approach relies on other large language models, which may also implement watermarks.
4Other Developers' Solutions
Apart from Meyer's method, other developers have developed different ways to remove watermarks. For example, software engineer Erik Hughes created a tool with Claude to remove watermarks. Leon Chlon from the University of Oxford suggested translating the text into a language like Arabic and then translating it back to remove watermarks.
5Future of Watermarks
Anthropic plans to develop an API to detect watermarks. This API will allow users to detect watermarks in texts. However, it remains uncertain whether watermarks will be completely unbreakable. Developers believe that watermarks could restrict the freedom of AI-generated content.
6Impacts of Watermarks
While watermarks make it easier to detect AI-generated content, they could also affect the quality and usability of such content. There is a risk of false positives, which could lead to employers unfairly rejecting candidates or researchers facing exaggerated accusations of AI use.
7Technical Details of Watermarks
Anthropic uses a technique called SynthID to embed invisible watermarks into texts. This technique was developed by Google and has been in use since 2023. Watermarks are based on the selection of words and phrases in texts, which are undetectable to humans. However, this technique could affect Claude's output and reduce the quality expected by users.
8Consequences of Watermarks
The full impact of watermarks on AI-generated content is still unknown. However, the rapid development of solutions by developers to remove watermarks raises doubts about the long-term effectiveness of this technology. The API Anthropic plans to develop for detecting watermarks could be a significant step in this regard.
Original article: Wired — Claude'nin Gizli Filigranları Kısa Sürede Aşıldı