Machine Speech, Human Speech
Raya Mezeklieva

Abstract:
As large language models (LLMs) increasingly populate everyday writing, certain words and stylistic habits have come to be perceived as telltale signs of machine authorship, reportedly prompting some writers to avoid them. This thesis investigates this oppositional pathway of LLM-driven language change, in which human writers adapt their language to not resemble LLM output but to be read as unmistakably human. Because LLMs themselves learn from human language which may include these oppositional changes, we hypothesized that successful differentiation strategies would not remain exclusively human for long, making this an iterative process of adaptation and counter-adaptation. To test how this may affect our linguistic behavior over time, we designed an online experiment repurposing the chain-generation structure of the Iterated Learning paradigm. Participants iteratively edited a news article to make it read more convincingly as human-written and unlike a simulated, state-of-the-art LLM that had itself incorporated the editing strategies of previous participants. We assessed (1) the sustainability of this process, based on participants’ own judgments of whether their edits improved the text’s signaling of human-authorship; (2) whether the strategies underlying participants’ edits changed systematically across generations; and (3) whether such changes were reflected in changes in the lexical diversity, syntactic complexity, and orthographic and grammatical correctness of the texts themselves over time. We found that participants could keep finding means to differentiate text from machine-generated output across several iterations. Editing rationales showed no consistent linear change across generations, though exploratory analyses suggested a possible non-monotonic, reaction-and-counterreaction pattern along some dimensions. At the level of the texts, lexical diversity declined steadily across generations, while syntactic complexity showed no comparable systematic change and error rate only increased slightly in early generations. These findings provide some of the first experimental evidence that individuals can repeatedly adjust their linguistic behavior to differentiate it from machine-generated text, while also suggesting that, in the absence of mechanisms for coordinating such adaptations, this process may tend toward lexical simplification rather than the lexical innovation more typically associated with language use for identity signaling.