Re-running 8 famous social science findings overnight with Claude Code + Codex: only one holds up
beglen · x · 2026-10-27
The author used Claude Code + Codex to re-test eight famous social science findings on newer data with high academic standards, completing the work overnight with summaries and publishable one-page infographics.
Early results:
- Only 1 of 8 findings holds up (with caveats); one holds only weakly
- Violent films reduce violent crime (Dahl & DellaVigna 2009): fails in 8 US cities, 2005–2025 — +0.3% assaults per extra million viewers (95% range -0.3% to +0.9%), opposite the original -1.3%
- Upset football losses and domestic violence (Card & Dahl 2011): +1.9% (-5.4% to +9.8%) in Chicago/Kansas City/LA police records, but the measure differs so it can't confirm or refute
- Hit TV characters change baby names: weaker than claimed
A demonstration of using coding agents for large-scale replication work — and a fresh warning about the replicability of classic findings.
More from coding & agent
- A daily two-checkpoint routine for managing AI agents at work — RachelVT42 · 2026-10-27
- "Is it time to fork Zed?" Community pushback over the editor's AI direction — gnukeith · 2026-10-27
- Reddit hands-on: Qwen3.5 9B struggles at agentic work while 35B A3B handles it — Nautisop · 2026-10-27
- JetBrains gifts researcher a flickering agent tracker, cooking his work-life balance — mervenoyann · 2026-10-27
- PM uses Claude Code and Codex for 74 analyses to spot a possible exoplanet in NASA data — CodeByPoonam · 2026-10-27
- Developer uses Claude Code and Codex to find a possible planet — CodeByPoonam · 2026-10-27