Something happens when you ask a question on the internet and get a clear and confident answer within seconds.
I feel like I’m making progress. You don’t have to sift through a dozen blog posts, forum threads, and personal testimonials of varying quality to get what you need.
A new study from the University of California, Riverside (UCR) suggests that what’s left out in that exchange matters more than you might think.
What’s more, as AI systems take over the way we find information online, the web may be quietly losing what it has spent 25 years accumulating.
The researchers compared how large-scale language models such as ChatGPT and Gemini respond to subjective, opinion-based questions.
They then compared this to human responses to the same questions. This difference was found to be consistent and meaningful.
logic and everything else
To categorize different types of reasoning, researchers turned to Aristotle’s rhetorical triangle.
This framework divides persuasion into three categories. Logos depends on logic and factual consistency. A spirit that appeals to authority and personal trust. and pathos, which is based on emotion and common human experience.
The team used these categories to compare how humans and AI systems structure arguments and answer questions.
a different kind of persuasion
When the team analyzed responses from ChatGPT and Gemini along with web results from Google and Bing, they found clear discrepancies.
Human-created web content utilized all three modes of reasoning, combining factual arguments and moral concerns, personal experience, emotional appeals, and storytelling.
“What we found is that humans essentially use all three of these, whereas LLMs rely essentially on logos alone,” said co-author Kevin Starling, a professor of public policy and political science at UCR.
“Their method of persuasion is different from the way humans use persuasion.”
margarita problem
The researchers use a concrete example to explain the difference. Ask an AI for a margarita recipe and you’ll get a competent, well-organized answer derived from vast amounts of training data.
What it doesn’t offer you is something like the Difford guide. This is a cocktail website where Simon Difford offers dozens of margarita recipes divided into seven styles.
The website traces the drink’s history back to Mexico in the 1930s, when journalists discovered what was then called the “Tequila Daisy.”
A human voice advocating such details—history, personality, and why these things are important—is exactly what the AI filters through. That’s not exactly wrong, but it seems dry.
Why does AI reasoning look like this?
Researchers have hypothesized why AI systems rely so heavily on fact-based, logic-centered responses.
The “tuning” and safety systems that AI companies overlay on their models are designed to steer responses away from emotionally or politically charged language and toward factual, indisputable grounds.
The result is an answer that is both reliably safe and systematically stripped of the messier, more personal kinds of reasoning that human writers bring to controversial questions.
The study also found that ChatGPT and Gemini are very similar in how they answer questions.
“Using AI platforms to replace web search results in a distilled version of knowledge that is constrained by the guardrails of each AI platform, loses the diversity of human emotions and opinions,” said co-author Vagelis Fristidis, a computer scientist at UCR.
Why is human communication different?
Part of what’s missing, Stirling argues, is fundamental to how humans actually communicate.
When people talk to each other, they always anticipate how the other person will react emotionally, intellectually, and morally. Those expectations shape how they discuss, what they emphasize, and what stories they tell.
“When humans talk to each other, we can understand what the other person is thinking,” Estherling said. “There’s this two-way street.”
Language models can’t do that. They generate statistically probable sequences of words based on training data and internal parameters.
I don’t have a model for the listener, I don’t know what will resonate emotionally, what will resonate personally.
“It’s completely different than talking to people,” Estherling said. “It’s just a machine that predicts what words to say in response to a prompt.”
what we might lose
These tools are increasingly used by people to find information about politics, health, ethics, and public affairs. It is precisely in these areas that all human reasoning is paramount.
Questions about health care reform and fossil fuel policy are not just questions of fact. It’s a question of values, whose experiences matter, and what kind of society people want to live in.
“As people come to rely on AI systems for information discovery at the expense of traditional web search, the web may gradually lose its soul and no longer reflect the human nature that has shaped it over the past 25 years,” Fristidis said.
The efficiency gains from AI-powered information retrieval are real, and getting clear answers quickly can be really useful. But what is being filtered is something that helps people understand each other, not just facts.
“As humans, we are hardwired to think that anything that produces language has human cognitive capabilities,” Stirling says. “However, this paper shows that when it comes to reasoning and argumentation, machines generate language that has no human qualities.”
The web was built by people arguing, sharing, persuading, and telling stories. Whether it lasts probably depends on whether we realize what we are giving up.
The research was presented at the ACM Web Science Conference in Braunschweig, Germany.
—–
Like what you read? Subscribe to our newsletter for fascinating articles, exclusive content and the latest updates.
Check out EarthSnap, a free app from Eric Ralls and Earth.com.
—–
