AI-generated text is everywhere. But how can we tell whether a piece of text actually came from an AI model? And more importantly — **how does watermarking actually work under the hood?** I built **The Green List** to make that question interactive. Instead of another blog post with static diagrams, you can actually experiment with LLM watermarking: 🟢 See how a secret key creates a “green list” of tokens 📊 Generate text and inspect z-scores, p-values and per-token evidence 🧪 Test paraphrasing, translation, truncation and adaptive attacks 🧠 Explore entropy, capacity and multi-bit payloads 🎲 Understand Gumbel-max and tournament sampling 🔬 Compare watermarking schemes side-by-side ⚡ Run post-hoc selection watermarking against real models The interesting part: **the numbers aren't fake.** The algorithms run directly in the browser, so the charts, statistics, KL divergences, decoded payloads and attack curves are calculated from the actual algorithms rather than pre-rendered illustrations. I built this because LLM watermarking is going to become increasingly important — but most explanations make the underlying ideas harder to understand than they need to be. My goal was simple: **Make the research understandable by letting people play with it.** 🔗 Try it: https://aianytime.github.io/llm-... 💻 Open source: https://github.com/AIAnytime/llm... Would love to hear from researchers, AI engineers, and educators: **what part of LLM watermarking is hardest to understand?**

The Green List

An interactive explainer app for LLM watermarking

An interactive explainer app for LLM watermarking