Anthropic’s watermark survives copy-paste, but not the real dev workflow
Anthropic
Anthropic's AI watermark system survives copy-paste but fails to hold up against real-world developer workflows, according to a report by The New Stack. The watermarking technology is designed to identify AI-generated text but proves ineffective when developers modify code or use it within their normal processes.
Anthropic has developed a watermarking technique to tag AI-generated text, which remains intact after copy-pasting. However, testing shows the watermark does not survive the actual development workflow: when developers edit, reformat, or integrate the AI output into their codebase, the watermark is lost, making it impossible to trace the text back to the AI model. The New Stack report highlights the limitation, suggesting that while watermarking may deter casual copying, it is insufficient for robust provenance tracking in professional software development.
Source: Anthropic (GNews) —
original