AI Coding Agents in Social Science: Methodologically Diverse, Empirically Consistent, Interpretively Vulnerable

In three linesComparative study of Claude Code and Codex on social science data analysis. AI agents produce methodological diversity equal to or exceeding humans, but remain vulnerable to interpretive bias at the verdict level. A biased prompt does not shift aggregate estimates unlike biased human analysts.

Summary generated by Claude — human-verified