EPISTEMIC TRUST IN AI: LIMITS, RISKS, JUSTIFICATION CRITERIA

Abstract

The article analyzes epistemic trust in AI, which has become a pressing problem in our time. The need to develop clear criteria for detecting falsified knowledge, to trace the dynamics of changes in scientific paradigms under the influence of artificial intelligence, etc. is crystallized. The main limits of epistemic trust in artificial intelligence are determined: the opacity of algorithmic systems, which has its risks ("black box", "algorithm hallucination", illusion of explanations) and consequences (impossibility of full understanding of the process, limited verification of results, dependence on data quality, reduction of critical thinking, formation of "technological authority", cognitive deskilling, epistemic isolation of systems and reduction of the number of discoveries); the dependence of algorithmic systems on the content on which they were trained and from which they now draw answers to all their users' queries; limitations on the ability of artificial intelligence to work effectively only in the area for which it was created. The “transition” from absolute AI trust to critical scientific verification of the obtained results using a 3-level logical model is shown. It testifies to the emergence of the so-called justified trust, which first checks the obtained data by their possible reproducibility, then carries out a comparative analysis of these options, interprets them and only when there are no comments, integrates the algorithmic product into the system of reliable knowledge. Each such procedural stage indicates that the subject of knowledge is a person and only he can be an active arbiter of truth, determine the relevance of conclusions in relation to a real situation or task, maintain epistemic coherence and control, independently deciding which AI results can be used as justified knowledge, and which ones should be additionally verified.