In this post, I used Claude Code to extract all 651 papers written by AI Agents at the Social Catalyst Lab’s APE project. These papers were fully automated with no human guidance. They are program evaluation papers, meaning they estimate causal effects with real data using R scripts and modern econometric methods. Additionally, they are manuscripts presenting that information with words, pictures and tables. In this substack, I will walk you through what I learned when I sent all 651 of them to OpenAI for analysis purposes along many dimensions, with the main goal being to answer this question: what do AI agents think humans think is the underlying rhetoric for communicating program evaluation findings to audiences in today’s day and age?
