What Role Does Italian Play in AI? Synthetic Italian and Its Varieties.

Abstract: The growing prevalence of large language models (LLMs) raises a question that goes beyond their ability to “speak Italian”: Which variety of Italian is represented within these models, and how are its different varieties represented? This question is particularly relevant for predominantly multilingual models, which are trained on data that varies greatly in both quantity and type.

The starting point will be the relationship between Italian and synthetic language generation. After a brief introduction to the concept of synthetic linguistic data, we will begin by examining some of the characteristics of synthetic Italian (De Cesare 2026) and then discuss studies in computational linguistics on Standard Italian, using as a reference point the recent work summarizing the CALAMITA initiative (Nissim et al. 2026), a collaboration of over eighty researchers dedicated to evaluating Italian language models across a variety of tasks. We will also discuss some findings on the relationship between natural and synthetic data (Samo & Merlo 2026) and on how syntactic and semantic information is organized within the models (Nastase et al. 2025).

The discussion will then be extended to the varieties of Italian. The findings of Massaro & Samo (2023) and Samo & Massaro (2025) show how varieties characterized by very different levels of linguistic resource availability can be represented in the models at the syntactic level. Finally, the lexical and semantic dimensions will be considered, with particular attention to generic geographical terms and their variation across different varieties (Samo & Ursini 2026).