l5-antifragile_big-data-cherry-picking

As the number of variables in a dataset grows, the number of possible correlations grows combinatorially. With enough variables, some patterns will appear statistically significant purely by chance, so the false discovery rate increases rather than decreases with more data.

Researchers can mine enormous datasets, test thousands of hypotheses, and publish only the ones that appear to work. The failed hypotheses are never reported, so a large proportion of published findings are noise presented as signal.

Researchers collect the upside of publication such as citations and tenure while bearing no cost for false positives. Since failed hypotheses and replications are rarely published, the incentive structure guarantees a high proportion of false discoveries.

Be skeptical when a finding comes from mining a large dataset without disclosure of how many hypotheses were tested. More data requires a proportional increase in skepticism, not less, because the potential for self-deception grows with data volume.

This is a conceptual and epistemological argument, not a statistical tool. It does not compute false discovery rates or test specific datasets; it provides a reasoning framework for evaluating research claims and incentive structures.