Thinking Outside the Black Box: Justifying Beliefs in the Age of Opaque Autonomous AI Systems

Abstract

In March 2025, Manus AI emerged as “the world’s first fully autonomous AI agent” that’s capable of end-to-end decision-making without human intervention. Unlike previous AI systems, Manus represents a fundamental shift in AI-Human interactions as the system can independently identify problems, formulate approaches, execute solutions, and, most significantly, adapt its methodologies based on outcomes without human guidance. Yet this rush toward autonomous AI has neglected a fundamental philosophical question that should precede any deployment of self-governing AI systems: What is the epistemic status of beliefs formed based on outputs from AI systems whose operations remain fundamentally opaque to human understanding? When we deploy autonomous AI systems to make decisions, we are effectively authorizing them to form “beliefs” about the world and then act upon those beliefs.This question becomes particularly pressing in light of the “black box” nature of current AI systems. While these systems consistently produce accurate outputs, the processes through which they arrive at these conclusions remain fundamentally opaque. In this paper, I examine the epistemic opacity challenges of AI systems alongside an empirical analysis of current epistemic attitudes everyday AI users hold toward these systems, based on survey data collected on this subject. I will then argue that Goldman's reliabilist theory of justification offers a compelling solution by shifting focus to the reliability of belief-forming processes, which accommodates our counterintuitive acceptance of beliefs formed through processes whose operations remain opaque to us.