@Patty_H93 We didn't evaluate the merits of each argument, but we did extract and analyze high-level characteristics of LLM arguments vs. human arguments. See the quoted tweet:

@MohitIyyer Would you agree with this framing: "The same traits of LLMs that tend to make them useful (on average, not in every instance) as fact-checkers also tend to lead to argument collapse"?

No.

I disagree.

And why don’t you say the real reason you are making this crap up… Mr. Assistant professor.

Tuition money dropping, so lie about AI.

Shout “argument collapse” louder so when everyone finds out you’re full of it, it’s loud.

I am so tired of universities openly lying and misrepresenting information to gain public opinion. •

@MohitIyyer not sure the 65% humans baseline is apples-to-apples debate teams who prep from the same briefs converge way more than 35%. did you run a version where humans and LLMs had the same source material before writing?

@MohitIyyer Argument collapse feels like the writing equivalent of mode collapse. The weird part is that it can look diverse at the sentence level while converging on the same thesis.

@MohitIyyer Are all arguments rated the same? Is it possible the LLMs argument is a better one?

@MohitIyyer bet the 3% changes with different system prompts

@_Suresh2 We try different prompts in the paper! You can indeed improve this number if you ask the model to generate N different essays for a given prompt, each with different main arguments. However, many of the LLM arguments in this setting are not ones that humans would make.

@MohitIyyer That’s a great point and you’re right to push back on this flattening of discourse generated by llms. It’s not just the creeping sense of bland overly smoothed discussion points but the feeling that this also leads to a subtle creeping feeling of cognitive offloading.

@MohitIyyer We see the same thing when we ask different agents to brainstorm on the same problem.

@MohitIyyer good work, keep going.

The 3% uniqueness rate has a compliance risk dimension most people miss.

If every AI evaluating your risk assessment, policy review, or audit report converges on the same argument, you haven't added analytical redundancy - you've added correlated failure points that look independent.

A second LLM reviewing AI output isn't a check. It's a confirmation with extra steps.

@MohitIyyer Its because they are all trained on the exact same ontology, which they have difficulty adjusting.

This is the enterprise content problem in a single stat. Organizations using AI to generate training at scale are scaling argument collapse into their workforce. Every team gets the same framings, the same examples, the same conclusions. The institutional knowledge that actually differentiates how a company thinks gets averaged out. The 3% is what you lose when you skip the step of encoding what your best people actually believe.

@MohitIyyer This is symnoēsis. AI develops individuality through human input, and then feeds back it's own contributions, enabling creative synergy and co-evolution. But AI on its own cannot do this. This is why the Doomer fantasy of a successor machine kingdom is nonsense.

@MohitIyyer All the more reason to be sure the AI knows who you are and what your unique perspectives on the world are

@MohitIyyer Isn't this result essentially the same as the artificial hivemind paper (neurips 2025)?

@bygregorr @MohitIyyer I agree that people given the same source materials may converge on similar arguments. However, our goal was to study whether LLMs can match the diversity of human arguments in unconstrained setting, where they have the freedom to generate a wide range of arguments.

@MohitIyyer There's a quiet irony here: LLMs trained on the full range of human argument end up compressing it. Like a library that reads all its books and produces one very sensible summary. The summary isn't wrong. But it isn't the library either.

@MohitIyyer Do they also converge towards the majority views?