THE LINGUISTICS OF AI: UNDERSTANDING HUMAN LANGUAGE PROCESSING IN MACHINES:

Abstract

This study focuses on the qualitative linguistic analysis of the discourse produced by artificial intelligence systems, which process, decode and generate human language. It speaks to a current debate on whether language models show actual linguistic understanding or for example just build patterns if they are trained on huge amounts of text. The study builds on the theories within the range of interest including Cognitive Linguistics, Systemic Functional Linguistics, Pragmatic Theory, and Discourse Analysis, and provides an evaluation of AI scores on four dimensions – syntax, semantics, pragmatics and cultural interpretation, all of which are interconnected. The largest number of texts was generated with a purposive sampling, using controlled texts, and then analysed with discourse analysis and thematic linguistic interpretation. This study compares three prevalent AI systems: ChatGPT (GPT-4), Gemini and Claude 3, on challenging tasks such as (complex) sentence creation, idiomatic translation, conversational politeness, and cultural symbolism. The results reveal that in general all three systems have good syntactic ability, a more or less precise semantic comprehension and a limited pragmatic and cultural competence. While most of these systems produce grammatical structures that are acceptable, meaning levels of common usage are correct, and expressions are socially acceptable, such responses are often formulaic and deficient in more dense contextual, experiential, historical, and cultural support. Based on this, the study points to the seeming inability of AI since the language performance does not seem to amount to human-level concept matching, but to sophisticated linguistic simulation. The study encompasses the use of known linguistic theories to analyze the discourse produced by AI and their implications for the future of language technologies, providing contributions for the recent field of AI linguistics and stressing the need for interpreting, culturally sensitive, yet understandable, and responsible development in language technologies.