LLM Output Watermarking — Detecting AI-Generated Content

Why it matters

Distinguishing AI from human writing is regulatory and practical need. Understanding shapes deployment.

The architecture

Green-list watermarking: at each generation step, bias sampling toward specific token subset. Detectable via statistical test.

How it works end to end

Robustness: sophisticated watermarks survive paraphrasing, translation, some editing. Simple ones don't.

Deployment: OpenAI, Google researching; not always enabled.

Limitations: attackers can strip via distillation or heavy editing.