Face, formulas, and the machine

Pragmalinguistic competence without sociopragmatic calibration in LLM speech act production across Indonesian digital interaction norms

Bibliographic Data

ID24237160
AuthorsTressyalina Tressyalina (0000-0002-4823-213X), Bima Mhd Ghaluh (0000-0002-6005-2246), Ella Wulandari (0009-0008-9426-8433)
Year2026
Volume266
Pages91-113
Publication date2026-12-01
Open AccessNo
TypeARTICLE
VenueJournal of Pragmatics (JOURNAL)
Journal identifiersISSN: 0378-2166 • E-ISSN: 1879-1387
PublisherElsevier BV (PUBLISHER)
DOI10.1016/j.pragma.2026.09.009
OpenAlexW7222840338
LanguageEN
References cited21

The rapid integration of large language models (LLMs) into multilingual digital environments has raised specific questions about contextual appropriateness, sociopragmatic calibration, and interactional adaptation across culturally situated norms. While benchmarks have evaluated LLM performance across dozens of languages, these assessments remain confined to classification tasks and are not designed to assess productive sociopragmatic competence in underrepresented ecologies such as Indonesian. This study compares speech act production between three flagship LLMs (GPT-5, Claude Opus 4.1, Gemini 3 Pro) and 150 Indonesian native speakers. Data were elicited using a 16-scenario Discourse Completion Task spanning requests, refusals, complaints, and apologies across varied power (+/-P) and distance (+/-D) configurations. Responses were coded using an adapted Cross-Cultural Speech Act Realization Patterns (CCSARP) framework and digital-pragmatic categories capturing emoji, phatic markers, colloquial particles, and code-mixing characteristic of Indonesian online discourse. Results indicate that while all three LLMs produced grammatically well-formed and strategically recognisable speech acts, significant divergences emerged in face-work calibration, scalar modification, and text-based digital-pragmatic signalling. All models systematically preferred negative politeness strategies and used basa-basi (phatic communion), colloquial particles, code-mixing, and emoji less frequently, while deploying honorifics with less contextual specificity than human responses. Claude Opus 4.1 exhibited the most conservative profile, while Gemini 3 Pro showed the greatest contextual sensitivity, though all remained distinct from the Indonesian native-speaker comparison group. These findings provide an early production-based empirical comparison of LLM pragmatic alignment in Indonesian digital interaction, contributing to computational sociopragmatics with implications for AI safety, cross-cultural NLP evaluation, and sociopragmatically aware language technologies.

Code-mixing · emojis · facework · honorifics · Indonesian language · large language models · LLM benchmarking · negative politeness · politeness strategies · power and distance · sociopragmatic competence · speech act production · Language, Discourse, Communication Strategies · Linguistics and Language Analysis · Speech and dialogue systems

Tools

Open DOIOpen Access