Discursive representation of the time concept in AI systems: A comparative analysis of temporal framing in Uzbek-language AI-generated content

This paper investigates how different artificial intelligence systems – namely ChatGPT (GPT-4o), Gemini 1.5 Pro, Qwen, and the Uzbek-language platform Salom-ai.uz – encode temporal concepts when generating Uzbek-language texts. Drawing on a purpose-built elicitation corpus of 24 AI-generated responses to six prompt categories targeting evidentiality, aspect contrast, cyclic time, habitual present, epistemic modality, and conceptual metaphor, this study conducts a systematic discourse analysis using an original seven-category error taxonomy. The findings reveal a hierarchy of temporal fidelity across systems, with Gemini demonstrating the strongest preservation of Uzbek evidential markers, while ChatGPT and Salom-ai.uz exhibit systematic TIME IS MONEY metaphor calquing from English. A paradoxical finding – that the Uzbek-targeted system Salom-ai.uz performed worst on evidentiality retention – points to the insufficiency of the target-language labelling without morphologically informed training. The study contributes an original error taxonomy and the concept of 'evidential erasure' to the emerging field of computational linguistics applied to low-resource agglutinative languages.