Research Reveals Using AI Results In Weaker Understanding, Less Original Insights

The research indicates that those relying on AI believe it to be smarter than they are and that it will discern all of the relevant information about the topic. They are therefore les motivated to learn and less capable of useful insights because they have not had to do the work themselves. 

The concerns are not just cognitive but behavioral. Surrendering control to AI in this way weakens the ability to learn and to reason, while giving control to those who own and train the AI being used. JL

Jackie Snow reports in the Wall Street Journal:

In experiments involving 4,500 participants at the University of Pennsylvania’s Wharton School, people who used LLMs to research everyday topics demonstrated weaker understanding of those topics afterward and produced less original insights than people who looked up the same topics using Google. The research is resonant with similar studies: Students who use AI tools to complete assignments tend to do better on homework—but worse on tests. “They’re getting the right answers, but they’re not learning.” The findings suggest believing information came from an LLM makes people learn less. "They think the system is smarter than them, so they stop trying. That’s a motivational issue, not just a cognitive one.”

ChatGPT and other “large language models” promise to make learning easier than ever. But new research suggests that lessons learned so easily are less likely to stick.

LLMs are a form of generative artificial intelligence that communicate like humans, using language. But in a series of experiments involving more than 4,500 participants at the University of Pennsylvania’s Wharton School, people who used LLMs to research everyday topics demonstrated weaker understanding of those topics afterward and produced less original insights than people who looked up the same topics using Google.

The findings raise concerns about how people search and learn, says Wharton marketing professor Shiri Melumad, first author of the research. “It is like the Google Effect on steroids,” she says, in a nod to earlier research suggesting people tend to remember less when information is easy to look up. With LLMs, she says, “We’re shifting even further away from active learning.”

Four experiments

Across four experiments, Melumad and co-author Jin Ho Yun randomly assigned participants to use either Google Search or an LLM to research an ordinary topic and then to write advice based on what they learned.

 

The first experiment asked more than 1,100 participants to use either Google or ChatGPT to research how to plant a vegetable garden. Compared with the ChatGPT users, Google users spent more time searching, reported greater effort, and wrote longer, more-detailed responses. Natural language analysis showed their advice also contained more unique phrasing and factual references.

To rule out the possibility that differences in the information itself—not just the way it was presented—was driving the results, a second experiment showed nearly 2,000 participants the same seven gardening tips, either as a single AI-style summary or split across six mock webpages, the way google search results are typically displayed. The Google users again engaged more deeply, retained more and wrote more thoughtful and original advice than the LLM users.

Two other experiments produced similar results.

A motivation problem

Daniel Oppenheimer, a professor of psychology and decision sciences at Carnegie Mellon University, says the research is resonant of what he sees in similar studies he does in his lab: Students who use AI tools to complete assignments tend to do better on homework—but worse on tests. “They’re getting the right answers, but they’re not learning,” he says.

Oppenheimer says the findings suggest that simply believing information came from an LLM makes people learn less. “It is like they think the system is smarter than them, so they stop trying,” he says. “That’s a motivational issue, not just a cognitive one.”

Oppenheimer cautions against rejecting AI altogether, however. He has seen GPT help students learn when they use it the right way—say, by critiquing a draft produced by an LLM or asking it probing questions. “AI doesn’t have to make us passive. But right now, that’s how people are using it,” he says.

Melumad, too, has concerns about future effects of AI, particularly in educational settings or professions that depend on critical thinking. But, like Oppenheimer, she sees the opportunities that AI offers as well.

“Younger people are increasingly turning to LLMs first,” Melumad says. “But if we don’t teach them how to synthesize and interpret information themselves, we risk deskilling the ability to learn deeply at all.”