Evaluating GPT-5 Literary Metaphor Translation from English into Arabic through Recognition, Interpretation, and Rendering

Abstract

Research has indicated that artificial intelligence (AI) systems can produce fluent literary translations but may struggle to preserve metaphorical meaning and imagery. However, most previous studies have evaluated completed translations, providing limited insight into whether AI systems recognize and interpret metaphors more successfully than they render them. This study extends previous research by examining GPT-5’s recognition, interpretation, and Arabic rendering of 202 literary metaphors identified through MIP/MIPVU in the first two chapters of Kazuo Ishiguro’s The Buried Giant. GPT-5’s Arabic renderings were also compared with those produced by DeepL and Kholoud Amr’s published translation. Each dimension was evaluated using a 0–3 ordinal scale, with high inter-rater reliability (weighted Cohen’s κ = .89–.93). The results showed that GPT-5 performed better in metaphor recognition (M = 2.30) and interpretation (M = 2.29) than in rendering (M = 1.37). DeepL achieved higher rendering scores (M = 1.68), while the published translation performed substantially better (M = 2.55). Qualitative analysis identified recurring patterns involving lexical equivalence, discourse-level consistency, contextual disambiguation, and rhetorical recreation. The findings suggest that successful metaphor recognition and interpretation do not necessarily correspond to effective figurative rendering in Arabic, highlighting the value of evaluating these dimensions separately.