Text Watermark Demo
Checking whether this browser can run the demo...
STEP 3
Review the watermarked text
Highlights mark small pieces of text that received the watermark's extra preference. A piece can be a whole word, part of a word, punctuation, or a space joined to a word.
Generated text will appear after a run.
Run the experiment to see which selected tokens received the watermark preference.
STEP 4
Check for a watermark
How one token is chosen
The model writes one token at a time. At each position, it scores the tokens that could come next. Those scores become probabilities, and sampling selects one token. A likely token has a better chance, but the top choice is not always selected.
Before sampling, the watermark uses your keyphrase and the four most recent tokens to choose a changing favoured set. It adds a small boost to those token scores. The checker later counts how often a favoured token was selected.
- PredictScore possible next tokens
- FavourBoost the keyed subset
- SampleSelect one adjusted token
- CheckCount favoured choices
RESULT
Waiting for textGenerate text to see whether it contains the pattern made with this keyphrase.
The two bars use the same scale, from zero to all checked positions.
- Positions checked
- N/AAt least 32 are needed
- Evidence score
- N/AStrong evidence starts at 4.00
How to read the evidence score
The score measures how far the found count sits above or below the no-watermark expectation. A score of 4.00 or more is strong evidence once 32 positions have been checked.
The approximate chance of a result this high if the watermark is absent is N/A. This is not the chance that AI wrote the text.
The final chance estimate appears when generation is complete.
STEP 5
Edit the text and check again
Edit the generated text and watch the score change. It shows how much evidence remains for a watermark made with this keyphrase.
The score updates shortly after you stop typing.
WAITING
Generate a passage before editing and checking it.
Length compared with originalN/A
About this demo
How it works
How does this demo add a watermark?
While the language model writes, the watermark gives a small boost to a secret, changing set of tokens. The checker uses the same keyphrase and rules to see whether those favoured tokens appear more often than expected. This creates a statistical pattern; it does not insert a hidden message.
What does the checker need?
It needs the same keyphrase, exact text-mapping rules and watermark settings used during generation. In this demo, one quarter of the possible text pieces is favoured, the boost is 2, four previous pieces provide context, and repeated contexts are counted once. The checker does not need the model or its original sampling settings.
What happens if I edit the text?
The watermark is added while the model writes, not to finished text. Small edits may leave enough of the pattern to detect, while shortening, paraphrasing or heavily rewriting can weaken or remove it. The result updates as you edit so you can see the effect.
Limits and results
What do the result and chance estimate mean?
Strong evidence means that the checker found substantially more keyphrase-matching tokens than it would expect without this watermark. The demo needs at least 32 checked positions and a score of 4.00 or higher. The chance estimate describes how often a result this high could occur without the watermark; it is not the chance that AI wrote the text. A lower score does not prove that the text is unwatermarked.
Can this prove AI authorship?
No. It tests only for this watermark, keyphrase and configuration. It cannot detect every kind of AI-written text, identify the person or model that wrote the text, or prove authorship. Editing, paraphrasing, different text-mapping rules or a wrong keyphrase can also weaken or invalidate the result.
Method and privacy
What research is this demo based on?
The watermark is based on A Watermark for Large Language Models by Kirchenbauer and colleagues.
Does my text leave this browser?
The model and its text-mapping file are downloaded when you first use the demo. Your prompt, keyphrase, generated text and watermark checks are processed in this browser; the demo does not send them to a text-generation service.
Why do the highlights split some words?
The model reads text in small pieces rather than only whole words. A piece can be a whole word, part of a word, punctuation, or a space joined to a word. The highlights follow those boundaries.