Biases in Generative AI

Speaker:

Dr. Francesco Sovrano, ETH Zurich

Abstract:

This presentation examines the promises and pitfalls of using generative AI in different settings. The talk highlights significant challenges: inherent biases in large language models can lead to reduced critical thinking, overreliance, and even propagation of errors when handling out-of-distribution inputs. Through empirical experiments, including analyses of programming education tasks and in-file vulnerability localization, the presentation demonstrates how such biases emerge and impact output quality. Emphasizing the role of explainable AI, Dr. Sovrano advocates for systematic detection and mitigation of these biases to responsibly integrate AI into the real world.

Biography:

Dr. Sovrano is a computer scientist and data science researcher specializing in responsible artificial intelligence. His work focuses on explanations as a way to enhance transparency, learning, and information acquisition. During his PhD, Francesco developed a computational theory of explanations, resulting in innovations like LLM-based user interfaces for enhancing textbook explanations and ensuring legal compliance of AI systems.

Currently an early-career fellow at ETH Zurich’s Collegium Helveticum, Dr. Sovrano is working on tools to better understand and address biases in LLM-generated outputs. His postdoctoral research at the University of Zurich included applying theories of explanation to software development, EU regulatory compliance, and machine learning for code. He received his PhD in Data Science and Computation from the University of Bologna in 2023.

Host: Prof. Marc Langheinrich