Back to Tools
AI Search

LLM Watermark Playground

Generate, detect, and attack statistically watermarked AI text.

The playground runs a real 125M-parameter language model on a free cloud instance. If it has been idle it may take a minute to wake up.

Open full screen

How to Use

  1. 1 Start in the Detect tab: load the bundled watermarked sample for an instant positive result, then paste your own writing and watch it score at exactly the chance rate
  2. 2 Generate your own watermarked text in tab 2 and compare it against an unwatermarked generation from the same prompt, with every green-list token highlighted
  3. 3 Attack the watermark in tab 3: truncation, deletion, and replacement sweeps, a dilution demo that hides a watermarked quote inside ordinary text, and a paraphrase attack you run with any chatbot
  4. 4 Change gamma or the seeding scheme in the sidebar after generating to see detection collapse: without the exact key there is nothing to test

What This Demonstrates

This is the watermarking scheme from Kirchenbauer et al. (2023), A Watermark for Large Language Models, and its robustness follow-up, On the Reliability of Watermarks. At each generation step, a hash of the preceding tokens pseudorandomly favours a "green list" of tokens. Whoever holds the key can count green tokens in any text and get a z-score with a calculable false positive rate: a hypothesis test, not a guess.

What This Is Not

It is not an AI detector. Text from ChatGPT, Claude, or Gemini will correctly score "not detected" here whatever marks it carries: Google watermarks Gemini output with SynthID-Text, and Anthropic now watermarks new Claude models under the EU AI Act, but detecting any vendor's mark requires that vendor's own key and scheme. A negative result is never evidence a text is human. Your own writing scores at the chance rate, which is the false-positive guarantee commercial AI detectors cannot offer.

Need something built for your business?

I build custom scripts, data pipelines, and full apps for SEO and product data problems that off-the-shelf tools don't solve.

Book a Call