METAPHORICAL FRAMINGS IN NEWSPAPER ARTICLES GENERATED BY CHATGPT-5.6 AND GEMINI, DEALING WITH THE WAR IN IRAN

Abstract

The paper aims to explore the range of metaphorical framings in newspapers articles generated by ChatGPT-5.6 and Gemini, dealing with the war in Iran. We have compiled a small specialized corpus with 42 articles and the total corpus length of 19,758 words. LLMs were given topically related article titles and introductory paragraphs chosen randomly from various online sources (e.g., Al Jazeera, Reuters). Based on the supplied information, the LLMs were instructed to generate three different versions of the article, from the following three viewpoints: (i) neutral, (ii) pro-Iranian, and (iii) pro-Israeli. All articles were tagged manually for instances of metaphorically used words, and prepared for subsequent analyses using WordSmith Tools 6.0. Metaphor identification was conducted using the MIPVU (Steen et al., 2010). Quantitative analysis showed a significant main effect of LLMs in the neutral condition (p=.01), with a significantly higher overall mean density of metaphorically used words in articles generated by Gemini. The difference in metaphor density in the two biased-framing conditions did not reach significance. Qualitative analysis revealed differences in the range and types of metaphorical framings between the two LLMs, where the discourse generated by Gemini showed more affective components compared to articles generated by ChatGPT-5.6. The discourse generated by ChatGPT-5.6 was dominated by containment, force, and journey metaphors. While these metaphor groups were also quite frequent in the discourse generated by Gemini, the latter LLM also generated metaphorical expressions corresponding to fire, hydra, cancer, disease, and snake metaphors. Through training and fine-grained instructions, the researched LLMs can pose as useful tools for researching possible framings and their rhetorical functions in discourse.