L’italiano generato dai modelli linguistici: una (nuova) varietà di lingua?
According to a recent hypothesis, texts generated in Italian by language models (such as those belonging to OpenAI’s GPT family) are claimed to represent a language variety that could be termed IA-taliano ‘AI-Italian’. This paper, which is exploratory in nature, provides initial empirical support for this hypothesis based on an analysis of two comparable corpora representative of the short biography genre: one consisting of texts drawn from the Treccani online encyclopedia, and the other consisting of texts produced by GPT-4o. The features selected for analysis span all linguistic levels and include both phenomena already discussed in relation to generated texts and phenomena that are salient for characterizing Italian varieties. The objective of the study is to quantitatively test the existence of correlation patterns among features, using a multivariate analysis technique – PCA (Principal Component Analysis) – and to qualitatively verify which features appear only in texts generated by GPT-4o, while also highlighting the different uses of a single feature. This initial survey shows that texts produced by GPT-4o are indeed characterized by a different configuration of features than those written by humans, as well as by certain distinctive phenomena. Furthermore, while the corpus of generated texts is characterized by a significant number of features that are gaining ground in the “neo-standard” and/or characteristic of a more “modern” use of Italian, the corpus of human written texts is marked by features characteristic of the literary standard and formal writing, or in any case typical of a more “traditional” standard usage.
Authors
- Anna-Maria De Cesare
- Massimo Cerruti
Publication Details
- Journal
- Linguistik Online
- Published
- 2026-11-14
- DOI
- https://doi.org/10.13092/vm9ets09
- Primary Topic
- Linguistic research and analysis
- Type
- article
- Field-Weighted Citation Impact
- 0.00