Politeness Strategies in Human–AI Interaction: A Comparative Pragmatic Analysis of Arabic Responses Produced by ChatGPT and Human Participants (Published)
Even though conversational large language models are now often involved in socially relevant communications; it is still unclear whether their polite presentation is tuned to the context in a similar way as human language. The present study contrasts the pragmatic implementation of politeness in Arabic responses elicited by a set of 15 written discourse-completion tasks which differ in social distance, relative power, and degree of imposition, between human participants and ChatGPT. There were 660 valid human answers and 1 standardized ChatGPT answer per scenario in the human corpus, and one standardized ChatGPT answer per scenario in the exploratory AI corpus. Analysis is based on Brown’s and Levinson’s politeness theory and speech-act coding along with directness of request, supportive moves, modal realization, and length of response in comparison. Conventionally indirect forms were most frequent (53.7%) in 352 human requests, followed by direct requests (32.7%), hints (7.7%) and non-target realizations (6.0%); Cochran’s Q(7)=60.66, p<.001. ChatGPT employed the conventionally indirect form in each of the 8 request situations and averaged out more supportive/mitigating devices per request (3.75 vs. 2.49). In all 15 scenarios the AI realization was found to be congruent with the modal human category in 9 cases (60.0%) and longer than the mean of the corresponding human scenario in all scenarios. The results show a pragmatic convergence, but systematic preference for elaboration, indirectness, and risk-averse mitigation in the responses. Human politeness was less rigid and contextually efficient. Note that the AI corpus consists of one generation per scenario, so the contrast with the human-generated one is an exploratory one and shouldn’t be interpreted as a fixed distribution of the model’s behaviour.
Keywords: Arabic, ChatGPT, Pragmatics, discourse completion test, human–AI interaction, large language models, politeness