Where an AI Watermark Can Hide in Plain Text

This article discusses Anthropic's announcement to watermark AI-generated text from Claude, focusing on the technical challenges and potential methods. It explains that while file metadata (C2PA) is easy to strip, an 'imperceptible watermark' in plain text is more complex. The author argues that since pure ASCII text has no hidden byte space, the watermark must lie in statistical word choice biases, similar to Google's Gemini. The piece also explores how such watermarks could be circumvented through editing or paraphrasing, making it a relevant read for those interested in AI security and content provenance.