A new study reveals that large language models generate significantly less diverse arguments than humans when responding to public debates, with only 3.4% of LLM main arguments being unique compared to 65.3% for human responses. This 'argument collapse' phenomenon persists even when models are prompted to generate diverse answers, suggesting LLMs may homogenize public discourse by repeatedly introducing the same polished arguments across different contexts.
The research documents a critical limitation of LLMs in maintaining argumentative diversity at scale. While previous studies identified LLM homogeneity in simple tasks, this work demonstrates the problem extends to complex, nuanced public debates where diversity of thought is essential. The comparison across 23,384 LLM-generated essays against human responses from New York Times debates and Boston Review forums reveals a structural mismatch: LLMs converge on similar argument frameworks regardless of input variation, whereas humans organically generate distinct approaches to the same questions.
This pattern reflects how LLMs are trained on large text corpora where certain well-articulated arguments appear repeatedly. When asked to generate public-facing content, models default to these high-likelihood argument sequences, essentially memorizing and regurgitating the 'consensus' arguments present in training data. The finding that diversity prompts only recover about half of distinct human arguments while introducing out-of-distribution variation underscores the fundamental challenge: forced diversity often means incoherence rather than genuine intellectual variety.
For platforms, media organizations, and public discourse infrastructure, argument collapse poses a legitimacy threat. If LLMs draft significant portions of letters-to-editors, op-ed responses, or forum contributions, public debates risk becoming increasingly synthetic and repetitive, potentially eroding the epistemic value of crowd-sourced argumentation. The structural patterns LLMs follow—opening with direct claims and pivoting to proposals—create recognizable templates that astute readers may eventually identify as machine-generated.
Future research should examine whether audiences detect this homogenization and whether it affects trust in human-written content adjacent to LLM outputs. Developers building debate or public discourse platforms must consider whether moderating LLM content or clearly labeling synthetic arguments becomes necessary to preserve argumentative diversity.
- →LLM-generated arguments achieve only 3.4% uniqueness versus 65.3% for humans in public debates, indicating severe homogenization at scale.
- →Sub-arguments and structural patterns also collapse, with LLMs reusing generic, hedged language while humans employ concrete, topic-specific reasoning.
- →Prompting LLMs for diversity produces only marginal improvements and introduces out-of-distribution arguments rather than genuine variety.
- →Argument collapse extends beyond short-form responses to longer-form essays, suggesting a systemic rather than context-specific limitation.
- →The phenomenon threatens the epistemic value of public discourse if LLMs become primary sources of written responses in debates and forums.