Humans overrely on overconfident language models, across languages

As large language models (LLMs) are deployed globally, it is crucial that their responses are calibrated across languages to accurately convey uncertainty and limitations. Previous work has shown that LLMs are linguistically overconfident in English, leading users to overrely on confident generations. However, the usage and interpretation of epistemic markers (e.g., ‘It’s definitely’, ‘I think’) can differ sharply across languages. Here, we study the risks of multilingual linguistic (mis)calibration, overconfidence, and overreliance across five languages to evaluate the safety of LLMs in a global context. We find that overreliance risks are high across all languages. We first analyze the distribution of LLM-generated epistemic markers, and observe that while LLMs are cross-linguistically overconfident, they are also sensitive to documented linguistic variation. For example, models generate the most markers of uncertainty in Japanese and the most markers of certainty in German and Mandarin. We then measure human reliance rates across languages, finding that while users strongly rely on confident LLM generations in all languages, reliance behaviors differ cross-linguistically: for example, users rely significantly more on expressions of uncertainty in Japanese than in English.

Introduction. A critical component of safe and reliable human-AI interaction is the ability for agents to clearly express their epistemic states, meta-information about their certainty in the knowledge they communicate. One common approach is to have a large language model (LLM) linguistically express its confidence using epistemic markers like ‘It’s definitely’ or ’I think’ (Kadavath et al., 2022; Tian et al., 2023; Liu et al., 2023; Xiong et al., 2024; Tanneru et al., 2023; Mielke et al., 2022; Lin et al., 2022; Stengel-Eskin et al., 2024). Recent work (e.g. Zhou et al., 2024b) has emphasized the need to calibrate model confidence with human reliance. In other words, models should express confidence in a way that triggers the right human reliance behaviors. Past work has found that English language models are systematically overconfident—that is, they generate epistemic markers expressing high certainty even when incorrect, with high frequency. Compounding this, English speakers tend to systematically overrely on language model outputs (Zhou et al., 2024b).

Discussion / Conclusion. Here, we studied the risks arising from miscalibrated multilingual language models. We found that, while multilingual LLMs do adhere to documented linguistic norms around the production of epistemic markers, they are still systematically overconfident across languages. Further, we showed that users tend to overrely on LLMs in all languages, and that overreliance risks may actually be worse in languages like Japanese, where expressions uncertainty are more common but have diminished function as markers of epistemic state. More Uncertainty Generations, Less Perceived Uncertainty Our findings illustrate that although languages might differ in their generations of epistemic markers, this does not guarantee a reduction in LLM overreliance.