UCR study finds AI could drain the web of its human soul

A particular feeling comes with asking a question online and receiving a neat, assured answer within seconds.

It can seem like real progress: you find the information you want without having to sift through a dozen blog articles, forum discussions and personal accounts of uneven quality.

However, a new study by the University of California, Riverside (UCR) indicates that the material removed in this process may matter more than it first appears.

As AI systems increasingly shape the way people look for information online, the web could also be gradually shedding something that took 25 years to build.

The researchers examined how large language models, including ChatGPT and Gemini, answer subjective questions based on opinion.

They compared those answers with human responses to identical questions, and identified a difference that was both meaningful and consistent.

Logic versus everything else

To sort the various forms of reasoning, the researchers drew on Aristotle’s rhetorical triangle.

This framework separates persuasion into three types: logos, based on logic and factual coherence; ethos, which invokes authority or personal trustworthiness; and pathos, which appeals to emotion and common human experience.

Using these categories, the team assessed how people and AI systems formulate arguments and respond to questions.

A different kind of persuasion

After examining answers from ChatGPT and Gemini alongside Google and Bing web results, the team identified a clear split.

Web content written by people used all three modes of reasoning, combining factual claims with moral issues, lived experience, emotional appeals and 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

The researchers offer a practical example of this distinction. Ask an AI for a margarita recipe, and it will produce a capable, clearly structured response based on a vast volume of training data.

Yet it will not provide the sort of material available on Difford’s Guide, a cocktail website on which Simon Difford presents dozens of margarita recipes across seven styles.

The site also follows the cocktail’s history to a journalist’s discovery in 1930s Mexico of a drink then known as the “Tequila Daisy.”

That sort of information - its history, personality and a human voice explaining why it matters - is precisely what AI removes. It is not exactly incorrect, but it can feel dry.

Why AI reasoning is shaped the way it is

The researchers propose a reason that AI systems depend so strongly on factual, logic-led answers.

The “alignment” and safety mechanisms that AI firms add to their models are intended to guide answers towards factual, non-controversial territory and away from emotional or politically charged phrasing.

Consequently, responses are dependably safe, but are consistently deprived of the more complicated and personal forms of reasoning that human authors apply to disputed issues.

The study further found that ChatGPT and Gemini answered questions in closely 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 contends that some of what is absent is central to the way people communicate in the first place.

In conversation, people continually anticipate the other person’s emotional, intellectual and moral reaction. That expectation influences their arguments, the points they stress and the stories they choose to share.

“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 in this manner. Instead, they create statistically likely word sequences using training data and internal parameters.

They have no representation of the listener, nor any awareness of what may connect emotionally or seem 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

People are increasingly turning to these tools for information on politics, health, ethics and public affairs - precisely the areas in which the full spectrum of human reasoning is most important.

A query about healthcare reform or fossil fuel policy is not solely about facts. It also concerns values, whose experiences are considered and the type of society people wish to inhabit.

“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 efficiency benefits of AI-driven information retrieval are genuine: receiving a clear answer quickly is truly helpful. Yet the process filters out the elements that enable people to understand one another, rather than facts alone.

“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.”

People built the web through debate, sharing, persuasion and storytelling. Whether it remains that way may depend on recognising what is being surrendered.

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