Artificial intelligence is increasingly reshaping public administration and policymaking. Across the policy cycle, generative AI may support agenda setting, policy formulation, decision-making, implementation, and evaluation by processing large volumes of information, identifying emerging issues, comparing alternatives, generating policy arguments, and improving communication. These capabilities suggest that AI may make policymaking more evidence-based, adaptive, and efficient. However, the growing use of generative AI in public policy also raises significant normative, institutional, and practical concerns.

This article examines the problems, biases, and limitations associated with using generative AI in policymaking and considers their implications for policy change. It asks three main questions: how existing scholarship characterizes the risks of AI and generative AI in public administration; how these risks are expected to appear in a real-world policy simulation; and what safeguards are necessary to ensure that AI supports rather than weakens policymaking.

The study adopts a two-part research design. First, it conducts a PRISMA-informed systematic literature review of scholarship and policy-oriented literature on AI in public administration, AI governance, ethics, social equity, public service delivery, and AI-supported policy development. The review will synthesize recurring concerns, including data and algorithmic bias, hallucination, misinformation, opacity, weak explainability, accountability gaps, erosion of human discretion, social equity implications, and ethical governance mechanisms.

Second, the study conducts an empirical policy-simulation analysis using data from the HKUST Inter-University Case Analysis x AI Competition 2025. In this competition, student teams analyzed the case of “Pet-Friendly Transportation in Hong Kong,” assessing whether pet-friendly public transport policies are desirable and feasible. Teams produced digital posters and five-minute “behind-the-scenes” videos reflecting on their AI use. These materials offer an opportunity to observe how novice policy analysts use generative AI in practice, how they understand its limitations, and how they verify, question, or rely on AI-generated information.

As the project is ongoing, the article presents expected rather than final findings. It expects to find that generative AI offers clear benefits in early-stage policy analysis by helping policymakers summarize information, structure arguments, identify stakeholders, compare jurisdictions, and improve written presentation. At the same time, AI-generated outputs are expected to create risks of hallucination, unreliable evidence, fabricated examples, and overgeneralized policy comparisons. The study also expects that policymakers may be more attentive to obvious factual errors than to deeper problems such as embedded normative assumptions, algorithmic bias, shallow reasoning, and policy ideas that appear comprehensive but lack causal logic, institutional realism, or ethical justification.

The study further identifies algorithmic monoculture, or systemic policy homogenization, as an emerging risk. When governments, agencies, or analysts rely on similar AI systems, they may produce increasingly similar policy ideas and recommendations, stifling innovation and erasing local or minority perspectives. The article argues that responsible AI-assisted policymaking requires AI literacy, source verification, disclosure of AI use, documentation of prompts and outputs, human-in-the-loop review, deliberative testing, oral defense, and mechanisms for questioning assumptions. It concludes that generative AI can support policymaking only when it augments, rather than replaces, human judgment, contextual expertise, ethical reflection, and democratic accountability.