As artificial intelligence makes headlines across the world, one School of Information Science associate professor researches how to keep AI human-centered, unbiased and accessible to everyone. Souvick “Vic” Ghosh joined the University of South Carolina’s College of Information and Communications just two months ago, but he quickly earned rare access to a government supercomputer needed to advance his AI research.

In 2020, Ghosh began studying AI with a focus on human-computer interaction and conversational systems – more commonly known as virtual assistants or chatbots designed to interact with humans using language. 

“I am pro-AI of course, but I do believe we must still handle its continued development with care,” said Ghosh. “I teach my students that we must approach the topic with a healthy amount of skepticism and explore the pros and cons of AI and its uses. My research centers on building human-centered AI.”

To build human-centered AI, Ghosh says AI must positively impact society and equitably represent all users, no matter their background. He said his research aims to accomplish that goal by answering these questions.

• How do people interact with conversational AI, and what do they need from it?
• How do AI systems engage with people of different identities, and how do they cause harm?
• Do our evaluation methods reflect the people these systems are meant to serve?

“To test conversational AI, you can use real people or simulated users. My latest research focuses on the effectiveness of using simulated users, which has become a standard way to test conversational search engines,” said Ghosh.

In his study, Ghosh captured data using both groups of real people and simulated users. When using people, he found the process slower, as finding and scheduling participants was cumbersome compared to collecting “thousands of conversations fast and cheap with simulated users.” However, Ghosh remains skeptical that data collected from simulated users truly reflects everyone who might need AI.

“We found that simulated users differed from real users in several ways such as having a higher reading level and wider vocabulary. So if these simulated users are used to advance conversational search engines for real people with lower reading proficiencies or someone searching in a second language, won’t they be underserved by the very technology supposedly designed to help them?” questioned Ghosh. “That is when the use of simulated users in testing of AI becomes a problem."

Ghosh realizes, however, that using simulated users rather than people remains the quickest, most cost-effective way to collect “conversations” needed to hone AI systems. Therefore, he must ensure the conversations, or data, he collects are effectively analyzed and used. Ghosh recently earned an allocation from the National Science Foundation, which provides him computing time and storage on a high-performance supercomputer housed at the Pittsburgh Supercomputing Center in Pennsylvania.

Ghosh plans to use this supercomputer to “measure the validity of large language models as simulated users in conversational search evaluation” – his pilot project. He hopes his latest research ultimately allows AI to eventually better serve people of all socio-economic, educational, cultural and racial backgrounds.

“One of the main reasons I decided to come to USC this fall was because the CIC Dean and other university leaders wanted to be one of the first educational institutions in the country to offer AI education,” said Ghosh. “So not only were they supportive of my research and application for this NSF allocation, but also they saw the importance of teaching the next generation about AI’s capabilities and our responsibilities as humans to proceed ethically with AI’s advancement.”