How inclusive large language models can be? The curious case of pragmatics

This article provides a conceptual and critical analysis of the role of generative artificial intelligence (GenAI), particularly Large Language Models (LLMs), in supporting the development of pragmatic competence in language education, with a specific focus on the Italian language. While GenAI tools demonstrate remarkable capabilities for personalized feedback and interactive instruction, their development is marked by a significant paradox: the very mechanisms that enable personalization are rooted in vast, centralized training corpora that are predominantly English-centric. This linguistic mediation introduces biases that risk distorting pragmatic norms in other languages, threatening communicative authenticity and linguistic inclusivity. This paper explores the implications of such biases for second language (L2) learning, highlighting potential risks to sociocultural communication norms and cognitive development. Grounded in postdigital and socio-material frameworks, and drawing on theories of cognitive extension, this analysis first problematizes the pragmatic profile of LLMs in Italian by critically reviewing the existing empirical landscape, including language-specific benchmarks. Identifying a crucial gap in pedagogically oriented research, the study then proposes a rigorous, multi-phase research agenda. This agenda aims to guide the co-design and validation of a GenAI tool that is ethically informed, pedagogically robust, and linguistically attuned to the nuances of Italian pragmatics. The ultimate contribution is a pathway toward ensuring that GenAI enhances rather than impoverishes learners’ communicative capacities, avoiding the potential for cognitive deflation and fostering a more equitable integration of AI in education.