Researchers propose a framework for evaluating the fairness of language models across different languages. Current evaluation methods focus on English text and may not accurately reflect performance in multilingual settings. The new framework includes empirical calibration of detection thresholds, additional quality measurements, and analysis of cross-language disparity. An experiment with six watermarking schemes and eleven languages reveals that fairness gaps are often due to structural differences between language families rather than idiosyncratic characteristics of individual languages. --- Why it matters: This matters because AI models are increasingly being used in multilingual settings, and understanding their fairness across different languages is crucial for ensuring equitable outcomes. By developing a more comprehensive evaluation framework, researchers can identify potential biases and develop more robust models. Source: https://arxiv.org/abs/2608.20047