LobstersWednesday · August 12, 2026FREE

Text Watermarking for Non-Academics

Anthropic has announced how Claude marks AI-generated content, bringing text watermarking into immediate practical discussion. The article explains that plain text appears to offer no obvious place for a watermark, unlike images or digital files, because copying text discards metadata and formatting. However, natural language is redundant, offering choices in word selection, contractions, and sentence structure. A language model makes similar choices when generating text, selecting from probable next tokens. A watermarking system can influence these selections in a controlled way, creating a statistical pattern across a long passage that a detector can test for later. This pattern survives ordinary copying and pasting because it lives in the chosen words themselves. An individual choice reveals little, but many coordinated choices provide evidence. The article notes that Anthropic's announcement gives the subject new practical relevance, but the technique is broader than Claude, and public descriptions do not provide a complete implementation specification. The principles describe the technology generally, not attributing a particular design to Anthropic.