The Science of Politics Can AI ‘vibe research’ replace social science?
6 snips
Feb 4, 2026 Andy Hall, a Stanford political scientist who used AI tools like Claude Code to extend and audit a vote-by-mail study. He describes using agents to plan and run modular replications, where they shine at data collection and coding but sometimes miscoded or missed cases. He explores best practices, risks of overconfident AI reviewers, dynamic papers that reveal fragility, and what massively faster research could enable.
AI Snips
Chapters
Transcript
Episode notes
Force Agents To Test Their Code
- Test every piece of AI-written code before trusting it to avoid long copy-paste debug loops.
- Maintain an evergreen clod.md with explicit rules so agents self-check and reduce errors.
AI Extension Of Vote‑By‑Mail Paper
- Andy Hall used Claude Code to extend his vote-by-mail paper and it replicated core results while adding new data.
- Claude collected most new county data but missed some elections and mis-coded one county year.
Agents Can Act, Not Just Chat
- Claude Code acts as a 'doer' on your machine: it can create files, fetch web data, run code, and produce outputs.
- That makes it powerful for building entire empirical projects end-to-end if properly supervised.