Abstract
Automatic fake news detection has become an important applied problem in natural language processing because digital news and social media content spread rapidly across languages. For Kazakh, the task is especially challenging due to the limited availability of open labeled corpora and the weak adaptation of ready-made models to local media contexts. This paper evaluates classical TF-IDF-based machine learning models on a balanced Kazakh-Russian fake/real dataset of 1,808 texts, with 452 fake and 452 real documents in each language. The experiments include bilingual training, Kazakh-only and Russian-only evaluation, as well as Kazakh-to-Russian and Russian-to-Kazakh transfer. Word-level and character-level TF IDF features are tested with Logistic Regression, Linear SVM, and Complement Naive Bayes. The monolingual and bilingual settings achieved Macro-F1 around 0.985. Cross-lingual evaluation revealed a clear directional asymmetry: Kazakh-to-Russian transfer produced Macro-F1 = 0.654, whereas Russian-to-Kazakh transfer reached Macro-F1 = 0.926. The findings are interpreted as an explainable baseline for Kazakh-Russian fake/real classification rather than as a production-ready fact checking system.