This paper introduces the AI Dunning-Kruger (AIDK) framework, a theoretical structure for understanding the inherent epistemic limitations of Large Language Models. Unlike human Dunning-Kruger effects, which are developmental and correctable through encounter with reality, AIDK is architectural and permanent—arising from the categorical separation between AI systems and the reality they purport to describe. The framework identifies the Interactive Dunning-Kruger Effect (IDKE), which occurs when AI epistemic limitations meet human epistemic limitations, producing confidence amplification untethered from warrant. The paper proposes the Human-Curated, AI-Enabled (HCAE) deployment framework and the Model Advanced Persistent Threat (MAPT) security posture as design responses. Developed under the ECAE (Expert-Curated, AI-Enabled) model described in this framework, with derivational contributions from Claude, Grok, ChatGPT, Perplexity, and Gemini.