Embarrassingly Simple Text Watermarks

TL;DR

Easymark introduces a family of embarrassingly simple text watermarks that leverage Unicode codepoints with identical appearances to watermark text without altering meaning or degrading quality. It provides three variants—Whitemark for digital text, Variantmark for CJK text, and Printmark for printed text—along with a formal Impossibility Theorem that perfect watermarking is unattainable, underscoring the value of simple, robust approaches. Empirical results show Easymark outperforms prior watermarks in detection reliability while maintaining BLEU and perplexity, and enabling user-side deployment without reliance on LLM providers. The method offers a practical baseline for watermarking in real-world flows and a foundation for further theoretical and applied work in machine-origin tracing of text. It emphasizes that watermarks should be used as a starter or complementary tool, given the inherent vulnerability to erasure.

Abstract

We propose Easymark, a family of embarrassingly simple yet effective watermarks. Text watermarking is becoming increasingly important with the advent of Large Language Models (LLM). LLMs can generate texts that cannot be distinguished from human-written texts. This is a serious problem for the credibility of the text. Easymark is a simple yet effective solution to this problem. Easymark can inject a watermark without changing the meaning of the text at all while a validator can detect if a text was generated from a system that adopted Easymark or not with high credibility. Easymark is extremely easy to implement so that it only requires a few lines of code. Easymark does not require access to LLMs, so it can be implemented on the user-side when the LLM providers do not offer watermarked LLMs. In spite of its simplicity, it achieves higher detection accuracy and BLEU scores than the state-of-the-art text watermarking methods. We also prove the impossibility theorem of perfect watermarking, which is valuable in its own right. This theorem shows that no matter how sophisticated a watermark is, a malicious user could remove it from the text, which motivate us to use a simple watermark such as Easymark. We carry out experiments with LLM-generated texts and confirm that Easymark can be detected reliably without any degradation of BLEU and perplexity, and outperform state-of-the-art watermarks in terms of both quality and reliability.

Figures (7)

• Figure 1: A screenshot of Whitemark on Jupyter noteook. [2]: The implementation of the Whitemark algorithm. [3][4]: The original text and watermarked text look the same. [5][6]: The watermarked text can be detected by the detect_watermark function. [7][8]: The Sacrebleu library identifies the original and watermarked texts.
• Figure 2: A list of whitespaces. The red bars indicate the lengths of the whitespaces. These codepoints can be used for steganography and watermarks. *: no-break space.
• Figure 3: Variant characters and variation selectors. Both (a) and (b) have the meaning of mackerel. Although these characters are exchangeable in most scenarios, Unicode supports distinguishing them by variation selectors for some special purposes. The right table shows the list of Unicode sequences for the characters. These variations can be used for watermarking.
• Figure 4: An illustration of Variantmark. (a) A watermarked text has alternating patterns. A red circle indicates a character with a variational character, and a blue circle indicates a character without a variational character. (b) The patterns remain even if the text is edited to some extent.
• Figure 5: A screenshot of Variantmark for Japanese texts. Although the original text and watermarked text look the same, the Unicode sequences are different.

Theorems & Definitions (9)

• Proposition 3.1
• proof
• Proposition 3.2
• proof
• Theorem 3.3: Impossibility Theorem, Informal
• Theorem 3.4: Impossibility Theorem, Formal
• proof
• Theorem B.1
• proof