Understanding LLM Watermarking
As AI-generated content becomes more prevalent, the ability to distinguish it from human-written text is crucial for academic integrity, misinformation prevention, and copyright. LLM Watermarking is a technique designed to embed invisible signals into the generated text.
How It Works: The “Green List” Method
Most watermarking techniques work during the token selection phase of text generation:
-
When generating a token, the model’s vocabulary is split into a “green list” and a “red list” based on a hash of the previous token.
-
The model is biased to select tokens from the “green list.”
-
A human writing naturally would select a random distribution of green and red tokens, but a watermarked AI will have a statistically improbable number of green tokens.
Detection
Detecting a watermark doesn’t require access to the model, just the hash function and the seed used for the green list. By analyzing the frequency of green-list tokens, a detector can calculate a “z-score” indicating the likelihood that the text was AI-generated.
Limitations
- Editing: Significant editing or paraphrasing by a human can “wash away” the watermark.
- Translation: Translating the text to another language and back often removes the statistical bias.
- Quality Trade-offs: Aggressive watermarking can sometimes lead to less diverse or creative text generation.