Two Paths of AI in Social Science Research: Super Agent and Human Mirror

Large language models (LLMs) are being applied across the social sciences along two qualitatively different trajectories. The *super-agent* path optimizes models to reason correctly and solve problems as an idealized rational agent; the *human-mirror* path repurposes models as simulators of human behavior, including the biases, misconceptions, and heuristic failures that real people exhibit. This review names and unifies the distinction, proposes a two-axis taxonomy that crosses development-time and usage-time choices, and traces the two paths across political science, educational measurement, and psychology. We then argue that the paths are in *training-objective tension*: convergent evidence suggests that the same alignment procedures that strengthen super-agent performance actively undermine human-mirror granularity by homogenizing outputs and suppressing the very variance that human-mirror research requires. To illustrate the contrast concretely, we submitted a single well-documented physics item to two large language models (Claude Sonnet 4.6 and ChatGPT o3) under a super-agent prompt and a human-mirror prompt. Both models converged on a correct, principle-based answer under the super-agent condition; both reproduced the documented pre-instruction misconception under the human-mirror condition, but diverged in the depth of the misconception chain invoked. We discuss implications for methodological discipline, dual-use ethics, and the case for purpose-built vertical-domain models in social science research.