Building Generative AI Products Responsibly at Snap Inc.

Generative AI models offer new opportunities and challenges for creating AI products responsibly. Existing literature in the field that focuses on the study of harmful or biased outcomes from AI and how to build AI responsibly has focused on discriminative models, where model performance can be evaluated across different demographic groups on objective measures. Generative AI models produce novel images and free-form text which require innovative methods for ensuring responsible product development and deployment. In this talk we discuss the approach we take at Snap for analyzing and correcting outcomes to create experiences that are high quality and comfortable for our community. We'll cover best practices including developing experiments to analyze and modify products before launch and ways to monitor product outcomes after launch using Snapchat's new chatbot, My AI, as a case study.