Study on Cognitive Biases of Large Language Models from a Cross-Lingual Perspective: A Comparative Analysis of English, Hebrew, and Russian

A groundbreaking cross-lingual study focuses on the cognitive bias patterns of large language models (LLMs) in English, Hebrew, and Russian contexts. It fills the gap in bias research for non-English languages (especially morphologically complex ones), provides important data support for understanding the multilingual fairness of AI systems, and helps build more fair and inclusive multilingual AI systems.

With the global widespread application of LLMs, whether they exhibit consistent cognitive biases across different languages has become a key issue. Traditional studies mostly focus on the monolingual English environment, and there is a lack of systematic understanding of bias patterns in morphologically complex languages such as Hebrew and Russian. This study fills this gap and provides an empirical basis for building more fair and inclusive multilingual AI systems.

The study constructs a parallel dataset covering English, Hebrew, and Russian, following strict psychological experimental paradigms, and includes classic cognitive bias types such as confirmation bias, anchoring effect, and availability heuristic. Each language version is proofread by native speakers to ensure semantic equivalence and cultural adaptability, which is a key quality control step for cross-lingual research.

The same model shows significant differences in bias intensity across different languages: As the language with the most abundant training data, English makes the model more "confident" in certain bias tasks but prone to falling into reasoning traps; Due to the scarcity of corpus, models in Hebrew and Russian are more conservative in reasoning and exhibit bias characteristics of different structures. This language-dependent bias pattern poses new challenges to the fairness evaluation of multilingual AI.

Although current multilingual alignment technologies can achieve basic semantic alignment, at the level of deep cognitive reasoning, different languages still maintain relatively independent bias characteristics. Simple translation or transfer learning cannot eliminate cross-lingual bias differences; developers need to introduce targeted debiasing strategies during the training phase, especially for low-resource languages.

When enterprises deploy multilingual LLMs, they cannot assume that the performance in the English environment can be transferred to Hebrew or Russian markets. They need to conduct specialized bias audits for target languages and establish language-specific content review mechanisms. This dataset provides developers with a standardized testing benchmark, promoting the progress of multilingual AI fairness in the industry.

Research Limitations: The three language samples are representative but do not cover global linguistic diversity; the cultural variability of cognitive biases has not been fully explored. In the future, it can be extended to more languages to deeply explore the mechanism of cultural factors' influence on AI biases.