How developers are trying to remove Anthropic’s AI text watermarks

AI developers are developing methods to bypass watermarks that are being added to machine-generated text by companies like Anthropic. This comes as regulators push for more transparency around AI-generated content. Anthropic announced it would embed invisible watermarks in future Claude models, but within days, developers began sharing tools to remove these watermarks.

One such tool, created by Guillaume Meyer, has gained popularity on GitHub. However, Anthropic has stated that its watermarks only indicate that Claude was 'likely' involved in processing the content at some point, not that it was exclusively generated by the AI. This has led to concerns that watermarking could result in false positives, causing employers to unfairly reject candidates or researchers being accused of using AI.

The watermarking method used by Anthropic is based on Google DeepMind's SynthID-Text approach, which leaves a pattern in the choice of words and phrases that is undetectable to humans but identifiable by machines. Early attempts to remove these watermarks have involved re-writing the text using another AI model that doesn't insert watermarks, swapping synonyms, and reorganising the content.

However, the effectiveness of these workarounds is uncertain, as it depends on Anthropic releasing the software they use for watermark detection.