When Jonathan Gratch arrived at USC in 1995 as a researcher in computer science, he didn’t expect to spend the next three decades asking how computers could model emotion – and what those models could teach us about human beings.
Now, as he prepares to leave USC to join Vanderbilt University, Gratch reflects on a career that has helped shape affective computing; the study of how computational systems recognize, model and respond to human emotion.
The evolution of virtual humans
As research professor of computer science and psychology at USC Viterbi’s Thomas Lord Department of Computer Science within USC Stevens School of Computing and Artificial Intelligence, Gratch directs the Affective Computing Lab at the USC Institute for Creative Technologies (ICT).
“ICT was formative for my research and gave me the freedom to pursue questions that few people were asking,” Gratch said. He describes ICT as a “Renaissance place,” where researchers with different expertise can approach shared problems from multiple angles. As a Department of Defense-sponsored University Affiliated Research Center (UARC), ICT brings academic and military expertise together with the technological and creative capabilities of the film and gaming industries, for the purpose of advancing technologies for simulation and training.
“The applied mission allowed us to build the research infrastructure around virtual humans,” Gratch said. “Combining AI and psychology with advances in graphics, immersive technologies and animation allowed us to create very realistic, interactive characters and use them to study how people respond.”
Modeling emotion
Gratch’s early research at USC focused on computational models of human cognition. With longtime collaborator Stacy Marsella, former USC researcher and now professor of computer science at Northeastern University, he began investigating appraisal theory, which explains emotion through how people evaluate events in relation to their goals. Anger, fear or guilt depend not simply on what happens, but on assessments including whether an event helps or hinders an important goal, who caused it and how much control a person perceives.
“The key insight was that the processes described by appraisal theory were already fundamental to AI reasoning,” Gratch said. “An intelligent agent has goals. It has to assess whether events support those goals and, when something goes wrong, whether it can recover.” The work helped establish a computational account of how emotions arise and influence behavior, a contribution Gratch identifies as one of the most important of his career. It also exemplified an approach that has remained central to his research: using computational models of human behavior to make psychological theories concrete enough to test.
Gratch’s influence extended to establishing affective computing as a research community. As founding editor-in-chief of IEEE Transactions on Affective Computing and former president of the Association for the Advancement of Affective Computing, he helped create institutions connecting research in AI, psychology, emotion and social science.
“I’m proud of the scientific contributions, but also of helping establish affective computing as an interdisciplinary field,” Gratch said. “A large part of that has been creating places where people studying AI, emotion and social science can learn from one another.”
Virtual humans as experimental partners
At ICT, computational models of emotion became part of a larger effort to create virtual humans.
“We connected our models of emotion to virtual bodies so the characters could determine what emotion made sense in a particular situation and how to respond,” Gratch said. “Once we had a character that could show emotion, my research shifted toward another question: How do people react to an emotional machine?”
Virtual humans became experimental instruments. Unlike human “confederates” traditionally used in psychology experiments, their behavior could be systematically controlled. Researchers could manipulate an expression of anger or another social signal and measure its effects on negotiation, cooperation and perceptions of an interaction partner.
“The advantage was that we knew exactly what was generating the behavior,” Gratch said. “We could systematically control what the virtual human did and study how a particular expression or behavior changed the interaction.”
One of Gratch’s most influential findings emerged from a DARPA-funded mental-health project with ICT colleagues. The team developed a virtual interviewer that combined emotion detection with empathy and active-listening behaviors to encourage military personnel to discuss symptoms associated with post-traumatic stress.
“We found that people felt more comfortable and less judged when they believed they were talking to a fully automated system,” Gratch said. In a study involving a National Guard unit deployed to Afghanistan, returning soldiers disclosed twice as many PTSD symptoms to the virtual system as through standard military intake forms. “Knowing it was a computer changed what people were willing to reveal.”
From virtual humans to LLMs
Large language models (LLMs) have transformed the scale at which Gratch can investigate social interaction. Earlier experiments sometimes required a researcher hidden behind an elaborate interface to control a virtual human, making studies expensive and restricted to particular scenarios. LLMs now allow researchers to conduct thousands of dialogues and study emotion in negotiations and conflicts with measurable outcomes.
“Human-AI interaction has suddenly become a major research focus, but ICT gave us the ability to study many of these questions years ago,” Gratch said. “We could build interactive characters when very few groups had the infrastructure to do it. Some of those papers are now becoming foundational references for researchers exploring similar questions with LLMs.”
His current research also returns to appraisal theory, investigating whether internal representations in foundation models correspond to psychological concepts involved in emotion and whether manipulating representations such as blame can change a model’s response.
“In a sense, we’ve come full circle,” Gratch said. “We started by asking whether psychological theories of emotion could be implemented computationally. Now we can ask whether models trained on enormous amounts of human language have developed internal representations that resemble those same theories.”
The social stakes have also expanded. People now use AI for companionship, relationship advice, negotiation and difficult conversations. Gratch is studying what may happen when machines mediate interactions ordinarily shaped by emotional feedback between people.
“AI can be very good at validating how you feel, but that isn’t always what you need,” Gratch said. “Sometimes growth comes from reappraising the situation – challenging your initial interpretation rather than simply reinforcing it. If these systems only tell us what we want to hear, we risk losing an important part of that process.”
Published on September 24th, 2026
Last updated on September 24th, 2026