Event description: Large language models (LLMs) can hold remarkably fluent conversations—but they also make things up. These “hallucinations” (or more accurately, confabulations) are one of the biggest challenges to building trustworthy AI systems. In this talk, Ash will explore why these errors happen, how we can spot them, and what can be done to reduce them. Ash will introduce VISTA Score, a new method for checking factual consistency across multi-turn conversations, and show how it outperforms existing tools in identifying misleading claims. She also shares practical strategies—from better prompts and retrieval methods to fine-tuning with both human and synthetic data—that can make smaller models nearly as reliable as their larger counterparts. The goal: to understand not just how these systems go wrong, but how we can make them more transparent, responsible, and aligned with the truth.
Speaker: Ash Lewis, Ohio State University
This event is sponsored by the Big Ten Academic Alliance (BTAA).
UW–Madison’s Love Data Week 2026 runs from Monday, February 9, to Friday, February 13, and provides an opportunity to explore, share, and celebrate all things data. The Office of Data, Academic Planning & Institutional Research (DAPIR) and the University Libraries at UW-Madison proudly sponsor UW–Madison’s Love Data Week, which includes a partnership with the BTAA.