Artificial intelligence as an epistemological Swiss Army knife: when perceived authority replaces verification
Why do people stop verifying when a claim arrives with signals of authority, and what changes when the source is an AI? This article argues that the hallucinations of language models and the crisis of authority of expert institutions are the same problem: in both cases, perceived authority replaces verification.
The argument starts from the replication crisis and from the theories of motivated and argumentative reasoning (Kunda, 1990; Mercier & Sperber, 2011): we reason to convince, and correction comes from the interlocutor who evaluates. Generative AI removes that interlocutor and adds a bias of its own, sycophancy. The cases of the four large audit firms that, between 2025 and 2026, published reports with fabricated citations, and the dispute over how to audit their auditors, illustrate the mechanism at institutional scale.
The article proposes a decision framework for the value of verifying, based on the value of information, that distinguishes three levels of verification (existence, fidelity and validity) and places the effect of authority on perceived probability and the effect of sycophancy on the diagnosticity of the source. It also shows that there are two routes to not verifying, the desire for a conclusion and the cost of checking it, and that both fall under the same inequality. This is a position paper: it reports no data of its own. It closes with a two-part experiment whose predictions, primary measure and refutation criteria are set before any data are collected.
Ravello Joo, C. E. (2026). Artificial intelligence as an epistemological Swiss Army knife: when perceived authority replaces verification [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.24116983
@misc{ravellojoo2026navaja,
title = {Artificial intelligence as an epistemological Swiss Army knife: when perceived authority replaces verification},
author = {Ravello Joo, Carlos Eduardo},
year = {2026},
publisher = {Zenodo},
doi = {10.5281/zenodo.24116983},
url = {https://doi.org/10.5281/zenodo.24116983},
note = {Preprint. ORCID: 0009-0007-5631-7436}
}