TB: Andy Hall: Five Pieces of Advice on AI for Polisci/PE Academics

Core Thesis

Academic researchers must treat AI coding agents as daily tools now—not aspirationally. The correct response to AI is not to do existing work faster but to attempt research that was previously impossible, and to shift toward an engineering paradigm: identify political problems, build AI-driven solutions, test them. AI governance is simultaneously the most important institutional design challenge of the current era, and political scientists are the most qualified people to work on it—but most haven't shown up.

Key Takeaways

Learn the Tools Aggressively — or Fall Behind

  • Use Claude Code or Codex every day; become a skilled manager of research agents
  • Learn to anticipate agent failure modes: strong at code and scraping; weak at new research questions and introductions
  • Require agents to test every piece of code before claiming a task is finished (Hall's CLAUDE.md discipline)
  • Have agents replicate your papers before posting — minimal automated verification from day 1
  • Yiqing Xu's automated reproducibility tool is cited as a clear example of the value already available

Increase Your Ambition

  • Write papers that could not have been written before AI — more datasets, more robustness checks, live dashboards, new visualizations
  • "Normal" papers (same scope as pre-AI) will start seeming boring, fast
  • The correct response to "more is possible" is to raise the target, not maintain it

Engineering-Style Research Paradigm

  • The credibility revolution (natural experiments, RCDs) is approaching its "natural crest" — low-hanging identification has been largely plucked
  • AI enables: identify a political problem → build an AI solution → test it in the world
  • Hall's examples: auditing AI voting-recommendation bias; building prediction market mechanisms with new liquidity structures

AI as an Object of Political Study

  • AI labs are accumulating "a kind of power that has few historical precedents" — over information, decision-making, and infrastructure for billions
  • Governance institutions were designed for a world that no longer exists and are not keeping up
  • Political economists are trained to study power concentration and institutional design — but "most of the discipline hasn't shown up yet"
  • "The people writing AI constitutions… are mostly not drawing on the traditions that run from Madison through North and Weingast to Acemoglu and Robinson. They're reinventing the wheel, sometimes poorly."

Clarify Your Goals

  • Hall's goal: build a science of governance and apply it
  • Inspired by Polybius: constitutional design as the source from which "all designs and plans of action not only originate, but reach their consummation"
  • For those with outward-facing research goals, AI is exciting; for those with inward-facing goals (the act of doing research in a particular way), AI may feel threatening

Mental Models

  • Incentives Matter — tenure insulates academics from the economic pressure that would force adaptation; Hall's advice is aimed precisely at those not yet facing that pressure
  • Compounding Returns — early, aggressive AI adoption compounds; each paper that uses AI tools trains better intuitions and produces verification infrastructure that the next paper inherits
  • Second-Order Thinking — the second-order effect of AI is not "do existing work faster" but "change the kind of work you attempt"
  • It Pays to Get the Design Right — applying institutional design expertise to AI governance is exactly what Hall is calling for; the wrong institutions will be hard to reverse

See also

Source

Andy Hall, Stanford GSB & Hoover Institution, freesystems.substack.com, ~2026 Google Doc: https://docs.google.com/document/d/1fyQ5brVo3AUo0epoL-cYubYw7WayDYQBXQNNqnTckmE/

Backlinks