Original title: Language, Language Models, and What We're Talking About
Malvina Nissim · September 2026
Language modelsComputer systems that generate or predict text based on patterns learned from large datasets. learn to talk from the data people feed them during training. But many systems learn from translated or artificial data instead of natural, everyday speech. They are also tested on equally unnatural data. Using Italian language models as evidence, the author asks if these tools truly represent real Italian or actual human language, and whether natural language processing still cares about language at all. This creates a bigger question: what kind of words do we actually want these systems to make? The author argues we must clearly separate models built as commercial engineering products from models built to study how human language works. Making this distinction shows the future might not be as pessimistic as we fear.
Imagine learning to cook a family recipe from a cookbook written in a language you don't speak, so you use an online translator. The translated instructions might miss subtle steps or ingredient details, so the dish never tastes quite right. In this paper, language models trained on translated or computer-made text similarly miss the authentic features of the language, making their outputs feel off. The researchers argue we need to separate models meant for practical use from those meant to understand language itself.
If AI tools for translation or education are built on models that don't reflect real language, they could produce unnatural or misleading text. This affects anyone relying on these tools for communication or learning in their native language.
The abstract doesn't provide empirical data or specific solutions; it's a conceptual argument about NLP research directions. It doesn't test hypotheses or measure model performance. The focus is on Italian models, so it doesn't establish generalizability to other languages.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Language, Language Models, and What We're Talking About — Malvina Nissim
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.