AI-generated text detection for Slavic languages

Abstract

This is a dataset of news in original form with two AI-generated versions in parallel for three languages: Slovenian, Czech, and Slovak. The current design targets the more precisely defined task of distinguishing human-written from machine-generated journalistic text.To reduce differences caused by fundamentally different types of publishers, public-service news organizations were selected as the initial human-text sources for each language. The resulting sources are: RTVSLO (\url{https://www.rtvslo.si/}) for Slovene, ČT24 (Česká televize) (\url{https://ct24.ceskatelevize.cz/}) for Czech, Správy STVR (\url{https://spravy.stvr.sk/}) for Slovak. For each source article, two synthetic versions were generated using DeepSeek (DeepSeek-V4-Flash) and ChatGPT (GPT-5.6 Sol). Both generators received the same lead-conditioned generation instruction. Additionally, for Slovene, the GAMS (GaMS3-12B-Instruct) LLM was used. The generation procedure uses the summary or lead associated with a human-written news article as input to a large language model (LLM). The LLM is then instructed to generate a complete news article based on this information.

More by Jernej Vičič

• Dataset of Uzbek base words: extraction and data analysis based on the school corpusSocial Sciences & Humanities Open · 2026 · Citations: 2
• Dataset of Uzbek base wordsRepository of the University of Primorsk · 2026
• Uzbek 65 million web corpus lemmatized Word2Vec modelZenodo · 2026
• Uzbek 65 million web corpus Word2Vec modelZenodo · 2026