Psychology's p-hacking crisis is repeating in AI benchmarking

Moritz Hardt, a computer scientist at the University of Maryland, has concluded that AI testing is a dark art dressed up as science.

The benchmarks — standardized datasets measuring whether new AI models outperform old ones — are not rigorous at all. Researchers cherry-pick metrics, report results selectively, tweak models until they get the headline number they want, and call it done.

Nobody is enforcing standards because nobody agreed there was anything to enforce. The field moved so fast and stakes felt so high that rigor got left behind like luggage at an airport. This has happened before, almost exactly, and we have the receipts.

For twenty years, psychology labs ran experiments with p-hacking. Researchers made ad hoc choices about what counts as meaningful, stopped experiments early when numbers looked good, tested multiple hypotheses and reported only winners. Switched analysis plans midstream if the original didn't work. The field knew these choices were possible. Researchers were honest people making them. Nobody was lying. The journals kept accepting papers, citations climbed, and reviews said things like, this seems reasonable. That was the whole problem. Reasonableness in a system without enforcement is not a standard, it is permission to do whatever makes your hypothesis stick. The replication crisis of the 2010s revealed that somewhere between 36 and 47 percent of published psychology findings could not be reproduced. Not because researchers were dishonest. Because the methodological sins were invisible and looked like normal research.

The same journal editors, conference committees. Peer reviewers who failed to catch the reproducibility catastrophe in psychology are accepting AI benchmark results today. They are making the same mistake in a different domain with identical underlying logic. Researcher discretion is not rigor. It is indistinguishability from rigor. We do not enforce genuine standards in fields until public humiliation forces us to. It takes scandal and journalists and independent audits and reputational death spirals, not the quiet voice of one scientist saying things could be better. The variable this time is velocity. AI moves faster than psychology ever did. If benchmarking fails catastrophically and we deploy a system we thought was safe because testing was the statistical equivalent of theatre, the consequences will unfold not in classrooms or labs but in production, in hospitals, in decisions about you.