Meet CBI fellow Chelsea Wang

Qiaosi Wang (Chelsea) recently received her Ph.D. in Human-Centered Computing at the Georgia Institute of Technology, where she was advised by Ashok K. Goel. Her research broadly lies in human-AI interaction, cognitive science, and responsible AI. Specifically, she studies the human-centered design of AI systems that exhibit seemingly human-level cognitive and social capabilities, especially Theory of Mind (ToM)-like capabilities, through her proposed theoretical framework of Mutual Theory of Mind (MToM). She is particularly interested in designing AI systems that can responsibly account for people’s changing social perceptions of AI during human-AI interactions.

In recent years, AI systems are increasingly being equipped with equivalents of humans' ToM capabilities to make inferences about people’s mental states (e.g., emotions, beliefs, intentions) as they assume a variety of social roles in our society. As these AI systems exhibit human-like social and cognitive capabilities to read and adapt to our preferences, beliefs, and intentions, people also begin to expect these machines to perform social functions at a human level. While failures in meeting those expectations can lead to user frustration or abandonment of the AI system, exceeding those social expectations can lead to greater harms, such as overreliance on AI, self-disclosure of sensitive information to AI, and even emotional attachment to AI. Managing and accounting for people’s perceptions and expectations of AI systems performing at various social capacities thus becomes a critical challenge for improving user experience and mitigating potential harms in human-AI interaction.

To address this challenge, Wang’s work aims to design AI systems with ToM-like capabilities that can responsibly account for humans’ changing perceptions of AI during communication, inspired by the Mutual Theory of Mind in human-human communication. Such AI systems could help replicate MToM in human-AI communication, where both the human and the AI can use their ToM-like capabilities to construct, recognize, and respond to how they are interpreted by the other party. This would facilitate accurate interpretations of each other’s roles and capabilities. Moving toward this vision, her work has provided theoretical implications through her proposed MToM framework for human-AI communication. This framework breaks down the iterative communication process into analyzable stages, highlighting the key elements involved in each stage that are critical to forming MToM. Her work has also contributed empirical and design implications by building AI’s ToM-like capability to automatically construct people’s perceptions of AI and by examining people’s changing perceptions of AI across successful and failed human-AI communication outcomes.

Several obstacles lie ahead as Wang and her colleagues work toward designing MToM in human-AI communication:

1. developing techniques to enable AI’s capability of continuously recognizing and updating its interpretation of humans’ perceptions of AI
2. designing human-centered evaluations of AI’s ToM-like capabilities in real-world human-AI communication
3. translating the MToM theoretical lens into practical design guidelines for AI systems
4. responsibly designing AI’s ToM-like capabilities while considering potential harms

With the generous support provided by the Carnegie Bosch Institute and guidance from her faculty hosts at HCII, she is excited to tackle some of these obstacles during her CBI postdoc.

During her CBI fellowship, she aims to develop a research agenda that establishes the theoretical, design, and evaluation foundations of MToM in human-AI communication. Along with her faculty hosts, Dr. Hong Shen and Dr. Jodi Forlizzi, she will focus on translating the MToM theoretical framework into concrete design guidelines with industry AI practitioners and examining the potential of building an MToM evaluation situated in real-world human-AI communication. The CBI fellowship provides an incredible opportunity for her to not only pursue academically interesting problems, such as theory development for MToM, but also to gain insights into industry AI design practices through connections and resources provided by industry mentors—a key part of her research agenda. She is excited to embark on this journey and establish further collaborations with the broader human-computer interaction and computer science community at CMU.