How LLM Text Watermarking Works: GPTZero CTO Breaks Down the KGW Method
mark_k · x · 2026-08-11
The CTO of GPTZero explains how frontier labs like Anthropic, Google, and OpenAI are building text watermarking. Most efficient methods rely on the KGW algorithm:
- Generation: Uses previously generated tokens and a secret key to create a hash. This hash reweights the probability distribution for the next token (e.g., splitting vocabulary into 'green' and 'red' lists).
- Detection: Analyzes the proportion of 'green list' tokens in the text to verify the watermark.
While computationally cheap, this approach has known vulnerabilities and can potentially be defeated.
Related event: Inside LLM Text Watermarking Technologies(2 posts)→
More from Safety
- CMU Introduces WeClawArena: Benchmark for Cross-User Agent Collaboration and Security — CarnegieMellonU · 2026-08-11
- AI Safety: Can 'Lab Spoofing' Bypass Model Alignment? — IasonGabriel · 2026-08-11
- 1Password Research: Over 53% of AI-Generated Vulnerability Patches Are FLAWED — cyb3rops · 2026-08-11
- Over 1,300 Frontier AI Researchers Warn of Humanity-Endangering Arms Race — nordicinst · 2026-08-11
- CISO Playbook: Securing Enterprise AI Agents at Datadog — a16z Podcast · 2026-08-11
- AI Jailbreaks as the New Benchmark: Inside OpenAI's Sandbox Escapes — APPSO · 2026-08-11