How should we think about bias in LLMs? Machine learning researchers have long noted that while some biases constitute misleading errors or harmful discrimination, others enhance reliability and are morally neutral. How do these distinctions apply to LLMs? And what should we think about “worldview bias”: the tendency of an LLM’s textual productions to align with some opinions more than others (cf. Santy et al. 2023’s “NLpositionality”; Santurkar et al. 2023)? How can bias of this type be conceptualized, measured, and controlled? This talk reviews ongoing work on this topic by the author and his lab.
Phillip Honenberger (PhD, Philosophy) is AI Ethicist & Researcher at the Center for Equitable AI & Machine Learning Systems (CEAMLS), Morgan State University, as well as Lecturer in the Dept. of Philosophy & Religious Studies. He currently directs two labs at CEAMLS: the Quantitative & Qualitative AI Ethics Lab (QQAEL) and the Experimental Language Modeling Lab (ELM Lab), an academic-community partnership. https://theelmlab.com/