ASSET Seminar: “How can we enable LLM auditing?”

Oversight and auditing of AI systems is becoming increasingly difficult as people use systems in a wide variety of ways, with instructions expressed in natural language prompts. We can no longer use readily quantifiable metrics like accuracy or statistical parity to understand model performance and potential impacts. Instead, we need ways of conducting open-ended analyses of models and usage data that do not infringe on user privacy. In this talk, I will discuss ways we are working towards these goals, beginning with an in-depth analysis of LLM usage in a specific domain: AI for querying astronomy literature. While manual analysis of usage data and follow-up interviews with astronomers offer an in-depth look at how astronomers interacted with an LLM-powered system, manual evaluation does not scale to the large volume of usage data in other contexts. Thus, I will next discuss methods for automated inductive coding, which offer more scalability, and finally, leveraging synthetic data to enable increased oversight of model usage and development without compromising privacy.

Seminar Recording