Title: The Homogenization of Discourse: How LLMs Dilute Long-Form Public Debate

Recent research suggests that the growing integration of Large Language Models (LLMs) into public discourse risks diminishing the diversity of debate. By repeatedly deploying polished and plausible rhetoric, these systems may inadvertently flatten the complexity of public argumentation. This study investigates "argument collapse," a phenomenon where essays produced by various LLMs converge upon a limited repertoire of core arguments, secondary points, and structural frameworks.

To analyze this trend, the researchers compared a dataset of 1,039 human-written responses from 195 New York Times (NYT) debates and 448 human contributions from 61 extended Boston Review (BR) forums against 23,384 essays generated by LLMs. The data reveals a stark contrast in originality: within the NYT corpus, 65.3% of main arguments presented by humans were unique to their respective debates, whereas only 3.4% of LLM-generated main arguments were distinct.

While prompting models to produce diverse outputs introduces some variation, the study found that typical models recover only about half of the distinct main arguments observed in human writing. Furthermore, much of the additional variation introduced by LLMs falls outside the boundaries of the human argument space.

This collapse extends to sub-arguments as well. In essays sharing the same primary thesis, 41.0% of human sub-arguments were unique, compared to just 9.1% in LLM responses. Qualitative analysis indicates that LLMs tend to recycle generalized and cautious sub-arguments, while human writers favor more concrete, topic-specific points. Structurally, LLM-generated essays adhere to a rigid narrative arc, typically beginning with a direct claim and rapidly transitioning to proposals. These patterns persisted in the longer-form essays from the Boston Review, confirming that argument collapse is not limited to short-form interactions but extends to more substantial written exchanges.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC