When AI Adapts, Teaches, and Trains: Paradoxes and Open Challenges in Human–AI Interaction

Bio: Eugenia Rho is an Assistant Professor of Computer Science at Virginia Tech whose research examines communication breakdowns—how people talk to one another, and how technology can enhance or disrupt these interactions. Bridging Natural Language Processing (NLP), Human-Computer Interaction (HCI), and Statistics, her work uses language as a lens for understanding the real-world impact of words, identifying factors that shape conversations, and exploring how emerging AI language technologies can foster more effective and empathetic communication.

At Virginia Tech, she leads the Society + AI & Language (SAIL) Lab, which develops computational models and design interventions that connect linguistic behavior to social outcomes. Her research team builds and applies NLP pipelines powered by large language models to predict conversational trajectories and measure the real-world consequences of how people communicate across diverse contexts. Integrating HCI principles and design thinking, her work also focuses on creating human-centered AI tools that help individuals understand, anticipate, and prevent communication breakdowns in everyday and high-stakes settings.

Her interdisciplinary collaborations span psychology, public health, law enforcement, and clinical practice, with a particular emphasis on developing ethical and context-sensitive language technologies. Her research has been featured by NPR, CNN, Forbes, PBS, Newsweek, Scripps News, and Scientific American, among others. Before joining Virginia Tech in 2021, she was a postdoctoral researcher in the Stanford NLP Group at Stanford Computer Science, working with Dan Jurafsky. She holds a PhD from the School of Information & Computer Science at the University of California, Irvine, and a bachelor’s in Political Science from Columbia University.

Abstract: AI language technologies promise to adapt to us, support our work, and help us grow. Yet these promises also raise fundamental challenges in how such systems are designed and evaluated. In this talk I will discuss three dimensions of this challenge through ongoing studies in my lab. The first is a paradox in LLM personalization. When autistic users disclose their identity to language models, the same mechanism meant to make AI more responsive can trigger stereotype-driven advice. We find that disclosure often steers models toward risk-averse recommendations, creating what we call the safety-opportunity paradox. The second is a tension in learning with AI. In our pair programming study, we compare human-AI (HAI) pairs with human-human-AI (HHAI) triads. When participants worked alone with AI, they relied heavily on AI-generated code. But when working with a human peer alongside AI, they became far more selective, using almost no AI-generated code in their final coding solutions. These contrasting experiences between HAI and HHAI conditions situate our findings within the long-standing HCI discussion on augmentation versus automation. The third involves open questions in AI-mediated communication training. Systems like CommCoach use conversational role-play to help managers practice difficult workplace interactions. But critical questions remain about whether AI can accurately recognize user intent in sensitive contexts, whether people will sustain practice over time, and whether better communication skills actually transfer to real workplace conversations. Taken together, these studies show that the deeper challenge is not whether AI performs well but whether it supports human agency and growth.

This seminar is both online and in-person:
Zoom Link