When does artificial intelligence improve human decision-making? Ruogu Fang, a widely recognized AI researcher who will be joining Vanderbilt’s esteemed Department of Biomedical Engineering on August 16 as the Flowers Family Dean’s Faculty Fellow in Engineering, aims to address this question and more through a recent award she received from the National Science Foundation.
Fang’s overall research addresses fundamental and translational questions at the intersection of neuroscience and artificial intelligence. It integrates AI, neuroscience, and biomedical imaging to quantify brain dynamics, enable earlier diagnosis of neurodegenerative diseases through advanced imaging, predict personalized treatment outcomes, and design precision interventions. She also leverages principles from neuroscience to develop next-generation trustworthy and adaptive AI systems.
Through the more than $890,000 NSF award, Fang seeks to understand how humans and AI can make better decisions together. AI performance is often viewed in terms of accuracy, said Fang, but accuracy is only part of the story. She said equally important is whether the AI knows when it is likely to be right or wrong, and whether it communicates that confidence effectively.
“We want to understand how the way AI expresses confidence influences human trust, reliance, and ultimately the quality of joint decision making,” said Fang, who is the grant’s co-principal investigator and site-PI at Vanderbilt. Brian Odegaard, assistant professor of psychology at the University of Florida, is PI.
For example, in her first experiment, human participants perform a visual perception task, deciding which of two briefly presented images contains a brighter target. After making their own choice and indicating how confident they are, they see the recommendation and confidence level of a simulated AI partner. If they disagree, participants can either keep their original answer or change it based on the AI’s advice.
The AI’s overall accuracy is held constant across all experimental conditions, Fang said. The only thing that changes is how informative the AI’s confidence is. In one condition, the AI’s confidence is essentially random. In another, it is moderately informative. In the third, the AI’s confidence is highly reliable, meaning that when it says it is confident, it is usually correct. Fang said this allows isolation of the role that confidence plays in effective human-AI collaboration.
She said by comparing how people interact with these different AI partners, questions can be answered such as: When do people appropriately rely on AI? What kinds of AI explanations and confidence signals lead to better teamwork? And how should future AI systems communicate uncertainty so that humans and AI together outperform either working alone?
“More broadly, while this first experiment focuses on controlled visual decision making, the overall NSF project seeks to establish general principles for designing AI teammates that communicate uncertainty effectively, adapt to individual users, and support better human-AI decision-making in applications ranging from healthcare to other high-stakes domains,” Fang said.