AI text watermarking & removal
What is AI text watermarking & removal?
Developers are building tools to detect AI-generated text through watermarks, but security researchers are simultaneously creating methods to strip those watermarks away, creating an escalating cat-and-mouse game.
If watermarks can be easily removed, the entire strategy for tracking and attributing AI-generated content collapses, making it harder to combat misinformation, detect plagiarism, and maintain transparency about machine-generated material.
References
- Google figures out how to watermark AI-designed proteins — Ars Technica
- Watermarkable Multi-Draft Speculative Sampling via Poisson Processes — ArXiv
- LLMs respond differently to harmful prompts when AI watermarking is used — Ars Technica
- Claude Watermarks Explained: How They Work, What They Reveal, and What You Can Remove — Medium: Large Language Models
- How text watermarking works — Medium: LLM
- Minimax bounds for watermarked and masked recursive discrete distribution estimation — ArXiv
- D2C-Routing: Dimension-to-Composition Evidence Routing for Mixed-Origin AI-Generated Text Detection — ArXiv
- Marketing AI Pulse Brief (August 2026): Claude's Watermark, CMO Gaps, AI Agent Ads & Fooling ChatGPT — YouTube