Abstract: As Large Language Models (LLMs) become deeply embedded in public governance, judicial assistance, and welfare allocation, social science is compelled to reexamine its own epistemological foundations. This article identifies three critical pitfalls in current social science research on LLMs:misinterpreting transient model deficiencies as structural social regularities, uncritically accepting manipulated benchmarks, and generalizing local optimizations of small-scale models into universal theories. To address these issues, the article establishes an epistemological boundary between engineering challenges and social propositions along two dimensions:whether a phenomenon can be resolved by scaling compute, and whether commercial logic provides sufficient incentive for its repair. This framework is crystallized into an operable four-quadrant matrix. Building on this foundation, the article identifies three research coordinates for social science:the linguistic erosion of administrative discretion, the selective disappearance of public values under optimization pressure, and the fracturing of accountability chains introduced by LLMs as a "fourth hand". The article further argues that the premise of social science standing "outside" AI is increasingly untenable.