Synthetic Confirmation: AI, Evidential Independence, and the Drift from Empirical Constraint

Abstract

Artificial intelligence is usually discussed as a threat to academic integrity, employment, or authorship. This paper argues that its more consequential effect is epistemic. AI sharply lowers the cost of generating hypotheses, synthesising literature, constructing arguments, analysing data, simulating outcomes, and drafting scientific prose, while the cost of obtaining independent evidence from the world falls far more slowly. This asymmetry alters the architecture of scientific evidence. I introduce synthetic confirmation: the apparent strengthening of a claim by evidence, analysis, simulation, or evaluation whose generative assumptions substantially overlap with those that produced the claim. It is not fabrication; its danger lies in the competence of every step. Science has historically kept hypothesis generation, evidence generation, analysis, interpretation, and evaluation partially independent; AI can perform all five from a shared informational base, so dependence masquerades as corroboration. I distinguish statistical from epistemic independence, describe assumption laundering, and locate the problem on a continuum of increasing dependence rather than in synthetic data as such. Building on Sher’s account of epistemic friction, I develop empirical friction: the resistance encountered when propositions meet phenomena the epistemic system does not control. Three cases — synthetic survey respondents, synthetic clinical data, and autonomous laboratories — show the framework discriminating between open and closed loops. I propose principles of provenance, evidential independence, reality holdouts, and adversarial testing. The defining question is not how much knowledge we can generate, but how much opportunity we preserve for the world to refuse it.