The distinction between traditional machine learning and generative AI is often treated as a matter of capability or scale. This paper argues it is architectural: ML systems close through reality; GenAI systems close through tokens. We develop this distinction through four stages: (1) conceptual analysis with ChatGPT establishing the theoretical difference through the robot arm and motion sensor as pedagogical anchors, (2) a two-phase empirical experiment with Perplexity AI that collapses GenAI into ML-like behavior through constraint, (3) synthesis with Claude connecting the results to the AI Dunning-Kruger (AIDK) framework, and (4) implications for AI development, deployment, and epistemics. Both experimental phases collapsed into repetition despite explicit prohibition, confirming the absence of symbolic bookkeeping as architectural rather than incidental. The methodology itself instantiates human-curated, AI-enabled (HCAE) collaboration: three AI systems performed derivation in distinct roles while origination, design, and interpretation remained human throughout.