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

Abstract

Introduction Students in non-English higher education are increasingly using large language models (LLMs) as informal study aids; however, the factual reliability of these systems in lower-resource languages - and its consequences for learning - remain largely unmeasured. Methods We report a controlled, fully reproducible evaluation of factual hallucination in eight open-weight LLMs answering 100 verifiable factual questions in English, Russian, and Uzbek. Uzbek, an agglutinative and comparatively low-resource Turkic language, is the primary language of instruction at our institution. Responses were scored with a deterministic answer-key procedure. Results The factual error rate increased monotonically with the resource level of the query language: 3.5% in English, 10.9% in Russian, and 15.9% in Uzbek (pooled over models; chi-square = 68.5, df = 2, p < 10 -14 ). The gap was present in nearly every model, but its size varied widely: a 12B general-purpose model answered reliably across all three languages, whereas the smaller 1-4B models had an error rate of 24%-27% in Uzbek. Language-adapted models - open checkpoints continued-pretrained or fine-tuned on Uzbek - closed the gap almost entirely, reaching an error rate of 2%-8% in Uzbek, comparable to their reliability in English. Discussion A student querying an off-the-shelf model in Uzbek is several times more likely to receive an incorrect fact than an English-speaking peer, which is both an equity concern and a hazard for academic integrity when unverified output enters student work. The risk, however, is not intrinsic to the language: it can be substantially reduced by preferring language-adapted models and by auditing per-language reliability before classroom deployment. All questions, code, and model outputs are released to support replication.