Stable Language Variation in the Age of AI

Can Large Language Models (LLMs) impact the pace and direction of change by altering stable patterns in language? Singular vs. Plural variation is commonly observed cross-linguistically, particularly in contexts with a collective subject such as ‘majority’ that agrees with a verb and/or adjective in a phrase as in Spanish la mayoría está/están de acuerdo en ‘the majority agrees/agree’. In many languages, this variation is quite stable; for example, in Persian it is documented over a period of 24 centuries. Do LLMs reproduce stable variation, or do they exaggerate biases, potentially propelling language change in an environment where texts are machine-generated and human beings are communicating with chatbots? We examine examples and data from Russian, Ukrainian, Norwegian, Persian, Spanish, and North Sámi.