Abstract

High-profile cases of AI-induced harms motivate the need for transparency in human-chatbot interactions. What if there were standardized disclosure frameworks for safe human-chatbot interactions, like nutritional labels on food? In this talk, Sheer Karny presents work toward building such labels by looking under the hood of AI models to surface their behavioral tendencies before users ever interact with them. His findings show how people systematically misjudge how AI models will behave, and that model behaviors we treat as simple, like sycophancy, are more complex than they appear internally. Additionally, he will demonstrate a new interface that tracks how these behavioral tendencies shift across a live conversation. This talk will explore the need for for new tools that help people form accurate expectations of AI, opening discussion on bridging cognitive science and AI transparency to make human-AI interaction safer.

Speaker Bio

Sheer Karny is a multidisciplinary researcher studying human-AI interaction advised by Dr. Pat Pataranutaporn. A through-line of his research is creating collaborative AI systems by designing assistants, agents, and interfaces that complement human cognition, decision-making, and well-being. His recent work has been about bridging mechanistic interpretability -- the ability to make sense of neural-level activity in LLMs -- and user interfaces to create safer human-chatbot interactions, helping dub the term Neural Transparency. This connects to his contributions to a recent pre-print, "My Boyfriend is AI", underscoring his desire to understand the individual and societal-level implications of the increasing prevalence of human-AI companionship. This work has been covered by MIT Tech Review, The Guardian, and The New York Times, among others.

His work has been published in journals and conference proceedings such as Nature Machine Intelligence and the ACM Collective Intelligence Conference.

To stay up-to-date on his work, check out his website and google scholar profile.