UCR study: ChatGPT and Gemini rely on logos, and the web may be losing its soul

Asking a question online and receiving a crisp, assured response within seconds triggers a particular feeling.

It feels like a step forward: you get what you came for without trawling through countless blog posts, forum debates and uneven personal accounts.

A fresh study from University of California, Riverside (UCR) indicates that what disappears in that streamlined exchange matters more than it first appears.

And as AI systems increasingly mediate how we discover information online, the web may be quietly shedding something it has built up over the last 25 years.

To investigate, the researchers looked at how large language models such as ChatGPT and Gemini answer subjective, opinion-led questions.

They then set those outputs against how people respond to the very same prompts. The gap they observed was both repeated and significant.

Logic versus everything else

To sort different styles of reasoning, the team drew on Aristotle’s rhetorical triangle.

This framework breaks persuasion into three modes: logos, grounded in logic and factual coherence; ethos, which leans on authority or personal credibility; and pathos, which taps into emotion and shared human experience.

Using these categories, the researchers compared how people and AI systems assemble arguments and produce answers.

A different kind of persuasion

When the team examined answers from ChatGPT and Gemini and compared them with web results surfaced by Google and Bing, they saw a pronounced split.

Content written by people on the web used all three kinds of reasoning, combining factual points with moral considerations, lived experience, emotional appeals and narrative storytelling.

“What we found is that humans essentially use all three of those, whereas LLMs essentially only rely on logos,” said co-author Kevin Esterling, a professor of public policy and political science at UCR.

“The way they try to persuade is different from the way humans persuade.”

The margarita problem

To make the distinction tangible, the researchers offer a straightforward example. Ask an AI for a margarita recipe and you will typically receive a capable, well-organised response assembled from a vast body of training data.

What you will not get is the sort of entry you might encounter on Difford’s Guide, a cocktail site where Simon Difford presents dozens of margarita recipes grouped into seven styles.

The site also follows the drink’s backstory to a journalist’s discovery in 1930s Mexico of something that, at the time, went by the name “Tequila Daisy.”

Those are the elements that tend to vanish in AI answers: the history, the personality, and the human voice explaining why any of it should matter. The output is not exactly incorrect, but it can read as oddly dry.

Why AI reasoning is shaped the way it is

The researchers also propose an explanation for why AI systems tilt so strongly towards fact-driven, logic-first replies.

The “alignment” and safety mechanisms added by AI companies are intended to push responses towards factual, non-controversial territory, and away from emotionally charged or politically loaded phrasing.

That produces answers that are dependably safe, but also consistently stripped of the more tangled and personal forms of reasoning that people use when writing about disputed issues.

The study further found that ChatGPT and Gemini answered in notably similar ways.

“When using AI platforms instead of web searches, we retrieve a distilled version of knowledge, constrained by the guardrails of each AI platform, and missing any human emotion or opinion diversity,” said co-author Vagelis Hristidis, a computer scientist at UCR.

Why human communication is different

Esterling argues that part of what is absent goes to the heart of how people actually communicate.

In conversation, people continually predict how others will respond-emotionally, intellectually and morally-and those expectations influence how they frame their case, what they stress, and which stories they choose to tell.

“When humans talk to each other, we can understand what the other is thinking,” Esterling said. “There’s this kind of two-way interaction.”

Language models do not operate that way. They produce statistically likely sequences of words based on training data and internal parameters.

There is no internal representation of a listener, and no sense of what will connect emotionally or feel personally meaningful.

“It’s not like talking to a person at all,” Esterling said. “It’s just a machine that’s predicting what words ought to be said in response to a prompt.”

What we might be losing

These tools are increasingly used to look up information about politics, health care, ethics and public affairs-precisely the areas where the full spectrum of human reasoning matters most.

A question about health care reform or fossil fuel policy is not merely a question of facts. It is also about values, whose experiences are treated as legitimate, and what sort of society people want to live in.

“As people increasingly rely on AI systems for information discovery at the expense of traditional web searches, the web may gradually lose its soul and cease to reflect the human nature that has shaped it over the past 25 years,” Hristidis said.

The gains in efficiency from AI-powered information retrieval are undeniable: getting a clear answer quickly is genuinely useful. Yet what gets screened out is often the material that helps people understand not only the facts, but one another.

“As humans, we’re hardwired to think that anything producing language has human cognition behind it,” Esterling said. “But this paper is showing that machines produce language that doesn’t have human qualities when it comes to reasoning and argumentation.”

The web has been shaped by people debating, sharing, persuading and telling stories. Whether it remains that way may hinge on whether we recognise what we are trading away.

The research was presented at the ACM Web Science Conference in Braunschweig, Germany.