Research Summary: Do large language models hallucinate more in Uzbek? A cross-lingual measurement of factual reliability and its implications for non-English higher education

A cross-lingual evaluation published in Frontiers in Education reveals that large language models (LLMs) exhibit significantly higher factual hallucination rates in lower-resource languages. Evaluating eight open-weight models across 100 verifiable factual queries demonstrated that error rates rose monotonically from 3.5% in English to 10.9% in Russian and 15.9% in Uzbek, highlighting acute factual reliability risks for non-English higher education.

Why it matters

This disparity matters because baseline performance metrics established in English cannot be assumed to apply across multi-lingual environments. Organizations operating in or serving lower-resource language populations face elevated misinformation and academic integrity risks when integrating generative AI. Consequently, unvetted AI adoption threatens educational quality and equity across international and non-English contexts.

Key insights

• Factual hallucination rates scale inversely with language resource availability: 3.5% in English, 10.9% in Russian, and 15.9% in Uzbek.
• The study evaluated eight open-weight large language models across 100 verifiable factual questions using deterministic scoring.
• Students in non-English higher education institutions are increasingly relying on LLMs as informal study aids without verified reliability benchmarks in their primary instruction languages.

Source

Frontiers in Education — https://www.frontiersin.org/articles/10.3389/feduc.2026.1929327