The increasing integration of artificial intelligence systems into scientific research, clinical decision-making, and everyday knowledge practices has generated a fundamental epistemological challenge: how can human agents be justified in believing outputs generated by systems whose inner workings are epistemically opaque? Deep Neural Networks, exemplified by systems such as AlphaFold2, are opaque in the sense that their internal functioning is often unintelligible to human investigators. This opacity appears to threaten traditional accounts of epistemic justification that require agents to have access to reasons or understanding. Recent scholarship has debated whether AI dependence poses genuinely novel problems for social epistemology or whether it parallels existing cases of reliance on human experts and instruments. This paper argues that while the core structure of epistemic dependence on AI may not be entirely novel, the scale, scope, and systematic character of this dependence generate distinctive challenges for theories of justification, epistemic authority, and epistemic responsibility. Drawing on recent work on Artificial Epistemic Authorities and the identification and deference problems they raise, the paper examines the conditions under which reliance on opaque AI systems can be epistemically justified. It contends that computational reliabilism and institutional approaches to epistemic warrant offer promising resources for navigating the tension between opacity and justification, while also highlighting the limits of transparency as an epistemic ideal. The paper then engages critically with recent challenges to computational reliabilism, including the problems of adversarial attacks, update opacity, error-related opacity, and warrant transmission, arguing that these challenges require the framework to incorporate adversarial resilience, context-specific reliability thresholds, and institutional mechanisms of accountability. It further argues that epistemological analysis of AI dependence has implications for global epistemic justice, as AI systems developed primarily in wealthy countries risk reproducing epistemic hierarchies that marginalise non-Western knowledge traditions. The paper concludes by outlining the implications of AI epistemic dependence for debates about epistemic autonomy, trust, and the division of cognitive labour in an era of increasingly autonomous algorithmic systems