Is Multilingual LLM Watermarking Truly Multilingual? A Simple Back-Translation Solution

Summary

This paper investigates the effectiveness of current methods designed to secretly mark text generated by large language models (LLMs) so it can be traced back to its source, specifically when that text is translated into different languages.

What's the problem?

Existing watermarking techniques for LLMs claim to work across many languages, but they haven't been thoroughly tested on languages that aren't widely used like English or Spanish. The researchers found that these methods become unreliable when the generated text is translated into medium- or low-resource languages – languages with less available data for things like translation and language modeling. This happens because the way the text is broken down into smaller pieces (tokens) doesn't work as well for these languages, leading to the watermark being lost or becoming undetectable.

What's the solution?

To fix this, the researchers developed a new method called STEAM. STEAM uses back-translation – translating the text *back* into the original language – to strengthen the watermark that might have been weakened during the initial translation. Importantly, STEAM isn't tied to any specific watermarking technique, meaning it can be added to existing methods to improve their multilingual performance. It also works well with different ways of breaking down text into tokens and is easily adaptable to new languages.

Why it matters?

This research is important because it highlights a fairness issue in LLM watermarking. If watermarking only works reliably for a few popular languages, it creates a bias and could unfairly impact users or applications that rely on less common languages. STEAM offers a simple and effective way to make watermarking more consistent and reliable across a wider range of languages, promoting fairer and more trustworthy use of LLMs globally.