Phantom Evidence: How and Why Generative AI Manufactures False Positives in Science

Abstract

Four centuries ago Francis Bacon warned against the anticipations of nature, hasty generalization that wins assent on a few facts, and set against it the table of absence: checking that a property fails to appear where it should not. The demand was that looking convincing should not, on its own, count as evidence. Science has professed that demand ever since, while in practice letting persuasiveness do the work of evidence. It could be let to do so because making something persuasive was itself hard. Generative AI removes that difficulty, and an old error returns on a scale and at a speed it never had before. We locate the problem not in evidence growing weaker but in how surprise is counted. An observer marvels at a convincing output as a single point hit among a vast range of possibilities, yet what a system can actually reach is a small part of that range. The gap between the breadth imagined and the narrowness actually reached is what we call phantom evidence, and we formalize it as one quantity that also absorbs the trial and error and the data leakage a research process adds. Three things follow. Higher resolution and greater fluency add no evidence. The evidence a single result can carry has a ceiling that neither polishing the output nor letting a generative system grade itself can exceed. And the fraction of published findings that are true falls back to what it was before anything was observed. The prescription lies in the same place: genuinely widen what a system can reach, and measure whether convincing outputs still appear when the target is absent -- Bacon's table of absence, restated in the language of probability. In a world where the persuasive has become cheap, the credibility of science rests not on more convincing outputs but on procedures that show they could not have arisen by chance.