Designing for Verifiability in Machine Learning Systems


Generative AI has impressive capabilities, but can be difficult to use efficiently and reliably to reach a specified goal, especially if users can’t directly verify outputs. Why is it so difficult to design interfaces to these systems that support verifiable interaction? How can we do better? In this talk I introduce Cephalo, a system I am building that supports more reliable and verifiable translation in the context of hospital discharge instructions. I explain why it’s important to build shared intent specification languages to support better interactions with generative AI, and how that applies beyond machine translation.

Jun 23, 2023