Educational Implications of Lexicographic Misalignment

AI-generated Definitions of English Loanwords in L2 Context

Abstract

By replicating the experimental design of Balenović and Proroković (2025), on the same lexical dataset, the present research investigates whether the GPT-5 model yields more accurate, contextually appropriate, and linguistically differentiated lexicographic output when it comes to senses and usages of English loanwords in Croatian. Additionally, the aim is to determine how GPT-5’s definitions, contextual examples, and sense partitioning compare qualitatively and quantitatively to those produced by GPT-4o. Despite improvements, the findings indicate that the GPT-5 model continues to display systematic vulnerabilities in L2 contexts, particularly for low-frequency loanwords. Though not exclusively, as in GPT-4o, less attested items elicit a higher rate of L1-to-L2 overgeneralization, leading to unverified or semantically implausible uses presented with high confidence. The study concludes that, although GPT-5 represents a significant step forward in the lexicographic endeavor, its performance remains inconsistent for L2 loanword interpretation. In other words, there still seems to exist the continued need for critical human oversight and the educational importance of AI literacy. Model's misinterpretation of prompt design and intent entails that recognizing and correcting such errors requires not only user awareness but language competence as well.