Two University of Maryland researchers have received seed funding for an interdisciplinary project that combines natural language processing and sociolinguistics to examine how people understand and evaluate AI in everyday life.
Supported by a one-year, $100,000 grant from the Institute for Trustworthy AI in Law & Society (TRAILS), the project brings together Julia Mendelsohn, an assistant professor in the College of Information with an appointment in the University of Maryland Institute for Advanced Computer Studies (UMIACS), and Charlotte Vaughn, an associate research professor in the Language Science Center.
The researchers are addressing a fundamental challenge in modern AI: understanding how the public perceives—and decides whether to trust—the technology.
“As AI becomes embedded across sectors and everyday communication, decisions about its adoption and regulation hinge on public trust,” Mendelsohn said.
Existing research relies heavily on surveys that ask people directly about their trust in AI. Extending beyond that, the UMD team will focus on indirect cues through a complementary, bottom-up approach, studying how trust emerges in natural conversation. By measuring linguistic cues that convey a person’s position on a topic, the researchers plan to identify implicit expressions of trust or distrust in AI systems.
“Looking more deeply at AI-related discourse online, I’ve kept noticing how people would signal (dis)trust in ways that aren’t captured by survey instruments alone,” said Mendelsohn, who brings expertise in natural language processing and computational sociolinguistics to the project. “I don’t think natural language can replace validated surveys at all, but it allows us to develop a more flexible model of trust that can capture a broader range of stances, and is more closely tied to people’s lived experiences.
Vaughn will apply her expertise in sociolinguistics to connect those computational models with human behavior. A central element of the project is the use of sociolinguistic theories of “stancetaking”—the ways people use language to position themselves in relation to a topic—to better understand people’s relationships with AI.
“We hypothesize that patterns in constructing stance when talking about AI are windows into how people understand and evaluate AI in everyday life,” Vaughn said.
Linguistic markers such as hedges that indicate uncertainty (“kind of”), modal verbs that express possibility (“might” or “could”) and adverbs that signal certainty (“definitely”) could offer clues about a person’s epistemic stance, or attitude toward knowledge, she said.
A major part of the project will involve face-to-face data collection at the Planet Word museum in Washington, D.C. Vaughn is the founder and director of Language Science Station, a pop-up research lab at the museum that invites visitors to participate in studies about language.
“Planet Word visitors are a diverse cross-section of folks from all over the region, country, and world, and the setting also offers a chance to interact with both adults and children,” Vaughn said. “We really aim to engage with each museum visitor who participates about the topic of the study, which has led to such insightful conversations with members of the public that would be hard to imagine having otherwise.”
Students will play a crucial role in the project, Mendelsohn said. During the coming year, undergraduate and graduate students from the College of Information will be trained to conduct in-person interviews and engage with museum visitors. A graduate research assistant will also help Mendelsohn and Vaughn design the study and analyze its data.
The researchers aim to learn not only how people make judgments about AI, but also how they define and conceptualize the technology itself.
“Even though I’ve been referring to ‘AI’ as a single umbrella, there isn’t a single unified definition or interpretation of ‘AI,’ and I expect that even the same person will have widely differing attitudes towards different kinds of technology or use cases,” said Mendelsohn, who is also a member of the Computational Linguistics and Information Processing (CLIP) Lab. “For AI literacy interventions to be successful, they need to be able to meet people where they’re at, and our work helps with this.”
Over the next year, the team will use its indirect measures of trust to study public discourse about AI across different sectors and sociodemographic groups. The researchers plan to use findings from the seed project to pursue larger collaborative grants from organizations including the National Science Foundation and the Russell Sage Foundation.
—Story by Diya Sharma, UMIACS communications group