Maryland Researchers Find LLMs Collapse Public Debate Into 19x Fewer Arguments
A UMD study finds LLMs converge on the same arguments 19x more than humans, even when told to diversify
Takeaways
- LLMs from five frontier labs converge on the same main arguments 19x more than humans: only 3.4% of LLM arguments are unique vs. 65.3% for humans in the same debates.
- The paper introduces argument collapse, measured across 1,039 human essays, 448 longer-form essays, and 23,384 LLM-generated essays from GPT, Claude, Gemini, DeepSeek, and Minimax.
- Asking models to explicitly generate diverse responses only recovers ~50–55% of distinct human arguments; much of the added variation falls outside the human argument space.
- LLMs favor abstract, hedged, unfalsifiable sub-arguments; humans use concrete cases, named statistics, and topic-specific proposals.
- Structural collapse is also documented: LLMs follow a rigid thesis→support→proposal arc far more than humans, even under diversity or persona-guided prompting.
- Code and data are available at github.com/mungg/argument_collapse.
Ask five different frontier AI models to weigh in on a contested public debate, and you might expect five different perspectives. What researchers at the University of Maryland found instead is something more troubling: the models almost always say the same thing. A new paper, "Argument Collapse: LLMs Flatten Long-Form Public Debate", introduces a formal framework for measuring this convergence and shows it runs far deeper than surface-level repetition.
The invisible consensus machine
The core problem is what the authors call argument collapse , the tendency of independently generated LLM essays to converge on the same small set of plausible arguments, regardless of which company built the model. This isn't about style or word choice. It's about the substance of the ideas themselves: the main claims, the supporting reasons, and even the structural arc of how an essay is organized.
The stakes are real. LLMs are now common aids in public-facing argumentative writing, including opinion essays and policy memos, but model suggestions can alter what claims writers make, how those claims are supported, and how much of the writer's own voice remains present. If everyone is reaching for the same AI co-author, and that co-author always reaches for the same arguments, the range of ideas circulating in public discourse quietly narrows.
The dataset: 24,000+ essays, two corpora
The researchers built a large-scale comparison using two real-world debate corpora:
- NYT Room for Debate: 1,039 human essays across 195 debates (median ~352 words each), covering 97 binary (yes/no) and 98 open-ended questions.
- Boston Review forums: 448 human essays across 61 longer-form forums (median ~1,150 words each).
For debates in both corpora, the team generated corresponding responses from five frontier LLMs , GPT, Claude, Gemini, DeepSeek, and Minimax , under three generation conditions: vanilla, diversified, and position-guided. In total, they collected 23,384 LLM-generated essays. To avoid confounding results with training cutoff drift, they excluded debates whose answers depend heavily on fast-changing events.