Borrowed Certainty: The AI Risk That Feels Like Relief

Abstract

The prevailing picture of AI risk is a picture of error: the machine is wrong and we are misled. This paper argues that a more insidious risk lies in relief. A fluent, confident answer relieves the discomfort of not knowing, and that relief tends to switch off the vigilance that would have caught the machine's mistakes. The result is borrowed certainty: the psychological state of certainty without the epistemic achievement that would make it knowledge. Drawing on Sosa's virtue epistemology, Kant's account of self-incurred minority, Fromm's analysis of the flight from freedom and the philosophy of epistemic dependence and trust, the paper develops four concepts: borrowed certainty; the relief threshold, at which confidence cues end inquiry; the distinction between waking and sleeping trust; and a vigilance-preservation criterion, under which AI systems are judged by what they do to their users' vigilance and not only by their accuracy. A simple expected-harm model shows when a highly accurate system that extinguishes checking produces more undetected harm than a less accurate one that keeps users alert. Published work on automation bias and over-reliance on large language models is consistent with this picture. Taha Abdurrahman's trusteeship ethics and Ibn Rushd's account of demonstration ground the claim that responsibility cannot be transferred to a machine. Design and policy implications and an empirical agenda follow.