Mark My Words: Analyzing and Evaluating Language Model Watermarks

Julien Piet, Chawin Sitawarin, Vivian Fang, Norman Mu, David A. Wagner

Abstract

As large language models (LLMs) proliferate, distinguishing machine-generated text from human writing has become a critical safety requirement. This paper introduces MARKMYWORDS, a new open-source benchmark that systematically evaluates LLM watermarking schemes. By decomposing watermarking into a modular design space, the authors evaluate current methods across diverse tasks and metrics, including quality, detection efficiency, and tamper resistance. The findings reveal that while current methods are practical for natural language—detecting AI text in under 100 tokens—they struggle with low-entropy tasks like code generation and sophisticated paraphrasing attacks.

Introduction

The rapid rise of LLMs has brought valid concerns about their misuse, from automated propaganda and social engineering to academic dishonesty. "Watermarking"—embedding a subtle, mathematical signal into the model's output—is a promising defense. However, the field has lacked a unified way to compare various proposed schemes, such as Kirchenbauer’s distribution-shifting or Aaronson’s exponential sampling. MARKMYWORDS provides a rigorous, empirical framework to evaluate these methods' efficiency, impact on text quality, and robustness against attacks.

The MARKMYWORDS Framework

The authors decompose watermarking into a modular "design space" consisting of three components: the randomness source, the sampling algorithm, and the score function. This modularity allows for a granular analysis of how different design choices affect performance.

The selection and verification procedures in the MARKMYWORDS framework.

To evaluate these components, the benchmark proposes three core metrics:

  • Quality (Q): Uses an "LLM-as-a-judge" (Llama 3 8B) to rate how much the watermark degrades the text's utility compared to unwatermarked text.
  • Watermark Size (E): The median number of tokens required to detect the watermark with a low (2%) false-positive rate.
  • Tamper Resistance (R): A score representing how well the watermark survives eight common perturbations, such as synonym swapping and paraphrasing.

Analysis of Watermarking Schemes

Several key insights emerged from the evaluation across diverse datasets, including creative writing, legal research, and code generation:

  • The "Distribution-Shift" Advantage: The scheme proposed by Kirchenbauer et al. (2023) emerged as the most practical for natural language. It is highly tunable, allowing developers to trade a tiny bit of text quality for much faster detection.
  • The Entropy Wall: Watermarking relies on "entropy"—the model having multiple plausible choices for the next word. In low-entropy tasks like code generation, forcing a watermark often breaks logic and degrades quality.
  • Indistinguishability vs. Utility: While some watermarks are theoretically "indistinguishable" from natural text, "distinguishable" ones like distribution-shifting often perform better in practice without human-perceivable quality loss.

Results and Discussion

For natural language, current schemes can detect a watermark in 60 to 100 tokens (roughly 40-70 words) with almost no perceivable loss in quality. Detection is extremely fast due to custom CUDA implementations, making it feasible for high-throughput applications.

Number of tokens needed to detect various watermark schemes across different sampling temperatures.

However, sophisticated paraphrasing remains a significant challenge. Paraphrasing by another LLM (like GPT-3.5) can remove up to 50% of the watermark signal from even the best schemes. This suggests that while we have effective tools to label AI-generated prose today, the arms race between watermarking and sophisticated paraphrasing is just beginning.

Conclusion and Future Work

MARKMYWORDS establishes a new standard for evaluating LLM watermarks. Future research should focus on addressing adaptive attackers who might iteratively scrub watermarks, improving performance on low-entropy tasks like programming, and defending against prompt-injection attacks designed to bypass watermarking logic entirely.

Summary updated

BibTeX
@inproceedings{piet2025mark,
  author = {Julien Piet and Chawin Sitawarin and Vivian Fang and Norman Mu and David A. Wagner},
  title = {{Mark My Words: Analyzing and Evaluating Language Model Watermarks}},
  booktitle = {SaTML},
  year = {2025},
  doi = {10.1109/satml64287.2025.00012},
}
Figure