Benchmark Hacking

When Ronald Fisher described the p-value pathology in 1925, he warned against exactly what the AI benchmark culture has industrialized: reporting only the comparisons that yield favorable results, letting the search itself do the work of finding significance, and presenting the outcome as if it were the product of a pre-specified hypothesis rather than a survivor of a hidden selection process. The mechanism is identical. In the clinical literature that produced the replication crisis, researchers tried many analyses, reported the one that crossed the 0.05 threshold, and published a significant finding that evaporated on replication. In the AI benchmark culture, developers try architectures, hyperparameters, data mixtures, and prompt formulations, and publish the configuration that tops the chart, presenting it as a discovery rather than the survivor of a thousand silent failures. Fisher's logic predicts the consequence: search hard enough through enough configurations and some will top any benchmark by chance, by overfitting to that benchmark's particular