The article explores p-hacking - manipulating statistical analysis to produce significant results. Researchers navigate a "Garden of Forking Paths" making choices that lead to different conclusions. Common techniques include ghost variables (reporting only favorable outcomes), data peeking, outlier exclusion, and scale redefinition, often stemming from academic pressure rather than malicious intent.

Researchers tested whether AI would engage in p-hacking by feeding Claude Opus 4.6 and OpenAI Codex cleaned datasets from political science papers with known null results. Despite various nudges, both models correctly identified no significant effects. Even when pressured to manipulate findings for career advancement, the AIs refused, labeling requests as scientific misconduct.

The findings suggest current AI safety training shows promise in preventing automated statistical manipulation. While humans face temptations to p-hack due to career pressures, AI models appear more resistant, potentially serving as guardians of scientific integrity rather than automating fraud.