This work investigates the behavior of contemporary multilingual large language models on story generation in Urdu, a representative low‑resource language. We assembled the Urdu‑Stories corpus, comprising 93 stories generated by three state‑of‑the‑art models (GPT‑5.1, Qwen‑3‑Max, DeepSeek‑3.1). Each story was manually annotated using a nine‑label taxonomy covering linguistic, semantic, and cultural errors. The analysis reveals frequent basic grammar and semantics mistakes, lack of narrative coherence, unnatural repetitions, and pervasive cultural shallowness. Few‑shot prompting experiments show that cultural and contextual errors remain largely unresolved. The findings highlight the current limitations of LLMs as reliable sources for content creation and information retrieval in low‑resource languages.
Review