This thesis investigates the extent to which state-of-the-art large language models (LLMs) can substantively improve the design of monetary, fiscal, ESG, educational, and intra-firm policies while preserving political legitimacy, economic efficiency, and ethical fairness. Part I undertakes a systematic review of 32 peer-reviewed publications (2000 – April 2025), contrasting pre-LLM and post-LLM periods to delineate the policy-making stages—problem diagnosis, option generation, and ex-ante impact assessment—in which artificial intelligence has demonstrably altered workflows. Part II analyzes the technical vulnerabilities of contemporary LLMs, including training-data bias, hallucination, and computational externalities, and evaluates mitigation strategies such as retrieval-augmented generation and guardrail fine-tuning. Part III synthesizes normative frameworks of distributive justice, welfare maximization, and democratic accountability to define a “balanced policy,” then applies this construct to multiple mixed-methods case studies. Empirical results indicate that LLM-assisted drafting accelerates preliminary policy formulation by approximately 50%, yet expert oversight remains indispensable for quantitative calibration and value-trade-off transparency. The thesis concludes by proposing evidence-based guidelines for responsible LLM deployment in public and corporate policy contexts and by identifying priority research avenues for next-generation models.
Zelin Zhao (Sat,) studied this question.