Cross-linguistic patterns of cognitive biases in large language models: a comparative study in English, Hebrew, and Russian

Large Language Models (LLMs) are being increasingly incorporated into decision-support systems. Nonetheless, a lack of clarity remains with reference to their reasoning processes, particularly in multilingual contexts. This uncertainty extends to cognitive biases - systematic errors in judgment, similar to those documented in human cognition. Existing research on cognitive biases in LLMs has focused primarily on English-language settings and a limited range of model families, leaving open the question of whether bias manifestations differ across input languages and distinct architectures. The current study investigated the ability of three widely used LLMs (ChatGPT, Claude, and Gemini) to solve cognitive tasks targeting availability heuristics and confirmation bias, comparing their performance to a human control group. The tasks were administered in English, Hebrew, and Russian, representing Germanic, Semitic, and Slavic linguistic contexts. The analytical dataset comprised 2, 028 observations: 507 human responses, collected via dedicated online questionnaires, and 1, 521 LLM-generated responses, obtained through API interfaces. Statistical analyses implemented Pearson’s Chi-Square tests, post hoc comparisons with Bonferroni correction, logistic regression, and Firth penalized logistic regression to compare correctness patterns across models, tasks, and languages, with human performance serving as a baseline. The results revealed a “cognitive gap”: LLMs consistently outperformed human participants on the rule-based deductive task, yet exhibited bias-mimicking error patterns in heuristic reasoning-based tasks. The observed effects varied significantly across languages and models, challenging the expectation of uniform multilingual performance and suggesting that LLM architecture interacts with linguistic structures in unpredictable ways. Overall, the findings indicate that cognitive bias expression in LLM outputs is not merely a technical constraint but a language-dependent phenomenon with practical implications for deployment in multilingual environments. The study emphasizes the need for cross-linguistic evaluation when assessing the reliability of LLM-based decision-support systems, particularly in domains where biased reasoning may affect judgment and decision quality.