- https://arxiv.org/pdf/2301.10226.pdf
- https://github.com/jwkirchenbauer/lm-watermarking
- Source: arxiv
TL;DR
A soft green-token watermark embedded during LLM sampling is detectable via a statistical hypothesis test with no model access required, published at ICML 2023.
Summary
Detecting AI-generated text without model access is the central challenge this paper solves. A random green list of tokens is selected before each generation step and softly promoted during sampling; a statistical hypothesis test on the proportion of green tokens then detects the watermark with interpretable p-values — no API access or model weights required. Tested on the OPT model family, the method produces negligible quality degradation while remaining robust to moderate text edits.
Key contributions
- Proposes a green-token watermarking scheme that biases sampling toward a secret random token subset without requiring model retraining.
- Introduces a statistical detection test based on green-token proportion that requires only the secret key, not model internals.
- Demonstrates negligible text quality degradation on OPT multi-billion parameter models.
- Establishes robustness bounds: the watermark survives moderate paraphrasing but can be defeated by targeted rewriting attacks.
When to cite
- When discussing methods for detecting AI-generated text without access to model weights or APIs.
- When arguing that watermarking and generation quality are not fundamentally in tension.
- When evaluating the robustness limits of LLM output provenance techniques against adversarial paraphrasing.
- When comparing approaches to LLM output attribution that rely on statistical rather than cryptographic guarantees.
LLM GPT On the Reliability of Watermarks for Large Language Models