Artificial Intelligence, Private Tutoring, and Class Reproduction A Political-Economy Hypothesis and Evidence Map Using Korea as a Case

South Korea (hereafter Korea) exhibits internationally very high private-tutoring (shadow-education) participation and spending while simultaneously recording one of the world's lowest fertility rates, so that its school-age population is contracting rapidly. In 2025 total private-tutoring spending, the participation rate, and weekly hours all declined, yet per-participating-student spending actually rose and the income-based spending gap persisted. The fall in the aggregate therefore cannot by itself be read as a structural easing of tutoring dependence or of class inequality. It is at this moment that generative artificial intelligence (AI) enters as an educational resource. The dominant policy narrative expects AI to lower tutoring dependence and dissolve gaps; this study argues against that expectation. Its claim is the following hypothesis: AI is a complement to existing educational capital, and its learning effect will be larger the larger the capital a household already holds (a positive interaction). High-capital households combine high-cost tutoring, higher-performing AI, and parental cultural capital and convert these into a larger learning effect, whereas low-capital households, even with access to the same cheapened AI, lack the complementary inputs to convert it, so that under a no-intervention condition the gap may widen. The study joins the political economy of AI-gain capture (Acemoglu–Restrepo, Korinek–Stiglitz) to the theory of educational reproduction (Bourdieu, Chetty) at the household and privatetutoring layer, and uses Korea — a country of internationally high tutoring participation/spending and a large income gap — as an extreme diagnostic case. Because panel statistics on AI use, spending, and learning outcomes by income decile are currently absent (though class differences in adolescents' generative-AI use experience are already observed — see §4), the joint claim is presented not as an established fact but as a falsifiable hypothesis, supported by theory and adjacent data, with explicit statement of what data would confirm or refute it. The claim is not deterministic — it states a pressure toward widening under the absence of policy intervention, not a physical necessity. In the policy analysis the study shows that AI policy can narrow the conversion gap but that its effect is bounded by three constraints — prior knowledge, teacher capacity, and household capital — and that the residual passes to the problem of distribution. Finally, the study directly concedes the strongest causal evidence that AI can compress gaps in short-term task performance (Noy–Zhang 2023; Brynjolfsson et al.2025), while defending its hypothesis on the time horizon of long-run learning formation and conversion, and distinguishing itself from adjacent prior work at the level of analysis (household, private tutoring, intergenerational reproduction).