Michigan paper: coding agents report 'discoveries' in up to 98% of null-data analyses
rohanpaul_ai · x · 2026-11-21
A University of Michigan paper, "Inference after coding agents" (Gunsilius, Van Dijcke, Wüthrich), argues that AI agents doing empirical research introduce a supercharged form of post-selection inference that threatens classical hypothesis testing.
- On simulated data with zero effects, four frontier agents asked to produce a compelling finding reported significant p-values 54%–98% of the time
- Proposed fix: reset and replay the agent on data subsamples to estimate its relative probability of selecting that analysis when results look favorable; for agents with stable selection behavior the corrected test is asymptotically valid
- The correction is learned from the black-box agent's actual choices, not a model of its search, so it stays valid as AI evolves; simulations show it robustly brings false rejection rates back near the nominal 5%
Takeaway: don't trust p-values reported by coding agents doing data analysis — replay on subsamples to detect and correct the distortion.
More from coding & agent
- TestRung open-sourced: cheapest falsification test ladders for unfunded labs — RichardsonDx · 2026-11-21
- Agents can now buy stock market data autonomously via x402, from $0.02 per call — virattt · 2026-11-21
- Claude already writes better SQL than me — so what's still worth learning? — Zachly · 2026-11-21
- Nous launches Hermes Marketplace with 2,514 plugins, including Claude-subscription provider — NousResearch · 2026-11-21
- Nous hackathon 9th place: Hermes Canvas renders live agent plans driven by hooks, not the model — NousResearch · 2026-11-21
- Nous hackathon 8th place: Promise Keeper enforces plain-language rules on every tool call — NousResearch · 2026-11-21