The Genetic Fallacy at Scale

The rapid normalization of generative artificial intelligence has produced two distinct epistemic problems that are routinely conflated. The first is a problem of quality: language models can generate false, derivative, vacuous, or stylistically polished but epistemically poor material at enormous scale. The second is a problem of judgment: the causal provenance of a text is increasingly treated as a substitute for evaluating its propositional content. This paper argues that the latter practice, when the label “AI-generated” or “AI-assisted” is used to reject a claim, argument, or intellectual artifact without content-responsive evaluation, instantiates the genetic fallacy at institutional scale. I first distinguish truth, warranted credence, authorship, and policy compliance, showing that provenance may rationally affect prior confidence or trigger additional scrutiny without thereby refuting a proposition. I then analyze AI-text detection and automated content moderation as forms of probabilistic classification that become epistemically illegitimate when converted directly into adverse adjudication. Empirical work on machine-generated-text detectors demonstrates substantial vulnerability to false positives, distribution shift, linguistic bias, and adversarial evasion, while the literature on automation bias and algorithmic accountability shows why nominal “human oversight” is inadequate when humans merely ratify machine outputs. I argue for a principle of human answerability: tools may be causal contributors to error, but responsibility for consequential epistemic decisions remains with the persons and institutions that design, deploy, supervise, and rely upon them. Finally, I propose six norms for legitimate epistemic governance in the age of generative AI: provenance modesty, content-responsive refutation, proportionality, contestability, auditability, and human answerability. The central danger is not that machines can produce bad arguments. Bad arguments can be answered. The deeper danger is that automated systems can acquire authority over which arguments are permitted to receive an answer at all.