Artificial intelligence doesn’t make decisions like a human, but according to research from Arts & Sciences economics professor Kevin He, people seem to think it does.
2 min. read
Today, people use artificial intelligence tools to do everything from cleaning up their resumes to designing holiday menus. Once their AI tool of choice has delivered an answer, though, it’s still up to the human to decide how to use that information.
That could change with new AI agents that promise to do basic tasks like shopping online or booking vacation travel without any human input at all, says Kevin He, an assistant professor in the Department of Economics in Penn’s School of Arts & Sciences. Though that prospect has all sorts of as-yet-unknown implications, a 2025 study He conducted makes one point clear: People using these tools may overestimate the extent to which AI will make the choice they want it to make.
For the study, He, doctoral student Mengjia Xia, and Penn State’s Ran Shorrer recruited 300 participants from the online research platform Prolific, gave them virtual chips as a currency, and asked them to decide how to spend those chips in a series of economic problems. At the end, they received $1 for every 1,000 chips they had left.
Across nine decision problems, study participants wagered, saved, or gave away their chips. Participants then predicted how an AI chatbot would act in the same scenarios. They found that participants consistently believed the AI would behave more like an average person than it did. In the donation game, for example, people donated an average of 27 out of 100 chips. They predicted the average AI would donate 32, but in reality, it donated 49.
People also exaggerated how closely the average AI response would match their own decisions—even though this AI had not been personalized to them.
The study can’t explain what’s driving these expectations of the AI. It may simply be human cognition: Researchers know that people tend to anthropomorphize animals and objects. They also tend to project their own knowledge, experiences, and opinions on others. This is true, for example, about consensus on political leanings. They may be applying that cognitive bias to AI, too, He says, adding that that the marketing and design of AI products may also influence misperception.
Materials in the Annenberg School for Communication Library Archives include thousands of TV scripts, the first issue of TV Guide, and interviews about the early days of HBO—which help to chronicle TV’s 100-year story.
Centering joy in AI development and implementation
PIK Professor Desmond Upton Patton—of Annenberg and SP2—and collaborators introduce a joy-informed framework designed to initiate conversations among engineers, designers, and researchers.
Winter Storm Fern brought icy and snowy conditions to the Northeast and other parts of the country over the weekend. Penn Today asks physicist Robert Carpick about the unique properties of ice, the science of curling, and how close we are to ‘nonslip’ ice.
Organizations like Penn’s Netter Center for Community Partnerships foster collaborations between Penn and public schools in the West Philadelphia community.
Penn receives national distinction for community engagement
The recognition by the American Council on Education and Carnegie Foundation for the Advancement of Teaching acknowledges Penn’s long-standing commitment to community-engaged scholarship and partnerships in West Philadelphia and beyond.
