The Limits of ChatGPT’s Pragmatic Interpretation: An Experimental Study of Conversational Implicature in Indonesian Language Conversations

Abstract

This study evaluates ChatGPT’s interpretative accuracy and contextual sensitivity in understanding conversational implicatures in Indonesian. A computational experimental design employed 80 minimal stimulus pairs, yielding 160 responses across five communication themes and four pragmatic phenomena. Each pair contained identical utterances presented in different situational contexts. Data were collected using GPT-6 at a ‘High’ reasoning level via Temporary Chat (unpersonalised). Two pragmatics experts independently assessed the responses using a 0–2 ordinal rubric. The analysis comprised Interpretation Accuracy (IA), Pairwise Joint Accuracy (PJA), descriptive comparisons between conditions, and qualitative examination of pragmatic inferences. IA reached 98.13% for Rater 1 and 91.25% for Rater 2, while PJA reached 96.25% and 82.50%, respectively. Exact agreement was 93.13%, with a weighted Cohen’s kappa of 0.332, indicating limitations in scoring consistency. According to Rater 2, pairwise accuracy was highest for indirect criticism or evaluation (95.00%) and lowest for indirect refusal (65.00%). Accuracy was descriptively higher in Condition A (96.25%) than in Condition B (86.25%). Qualitative analysis identified instances of inferences exceeding contextual support and insufficient use of relevant contextual information. These findings highlight the importance of evaluating both interpretative accuracy and contextual justification, while acknowledging the sensitivity of performance estimates to rater judgments.