Generated Text as a Locus of Language Contact: Exploring the Impact of Generative AI on English and German

Abstract

This thesis investigates the impact of machine-generated language on human linguistic behavior, addressing its implications for language production, translation practices, and contact-induced language change. In the current era of rapidly advancing language technologies, including machine translation systems and large language models (LLMs) such as GPT-4, this research introduces a novel framework for understanding how exposure to generated text influences natural language use. The thesis is multidisciplinary approach, integrating corpus linguistics, psycholinguistics, translation studies, contact linguistics, and computational linguistics. The corpus-based analyses contrast the linguistic properties of generated texts – including machine translations and open-ended outputs from LLMs – with human translations and human-authored writing. Findings reveal significant differences in lexical diversity, syntactic monotonicity, and lexical density, highlighting the distinctive characteristics of generated text across languages and domains. Furthermore, the psycholinguistic experiments demonstrate that exposure to machine-generated texts influences human syntactic choices, with strong priming effects observed among L2 learners of German. These effects persist across translation and production tasks, suggesting that generated texts may shape natural language patterns in contexts where they are heavily relied upon. The thesis also situates machine-generated language within the paradigm of language contact, proposing the framework of Language Contact Through Generated Text (LCTGT). This conceptual model extends existing theories of language change through translation, recognizing generative AI as an influential linguistic input capable of reshaping human language patterns. As language technology becomes increasingly common, it will be important to continue investigating the linguistic implications of exposure to generated text. It is hoped that LCTGT can help in this future line of research.