The study, conducted by Zander Weisman Mintz of the LBJ School of Public Affairs at the University of Texas at Austin and Ji Ma, now of Indiana University Bloomington and the University of Oxford, combines a large-scale systematic literature review with cutting-edge computational text analysis. Drawing on OpenAlex, an open scholarly database, the researchers built a corpus through citation snowballing that began with four foundational review articles and expanded forward through more than 14,900 candidate works. They then screened this mountain of material using Meta’s open-source Llama3 large language model, employing role prompting and chain-of-thought prompting to classify each candidate against six qualifying conditions. Validation against a manually reviewed sample of 200 works achieved 91 percent agreement. The final corpus comprised 3,268 works, published between the mid-1980s and May 2026.

To extract the latent thematic structure of this literature, the team applied BERTopic, a neural topic-modeling technique that leverages transformer-based document embeddings and clusters abstracts into coherent topics represented through class-based term frequency–inverse document frequency. The method outperformed conventional bag-of-words alternatives such as latent Dirichlet allocation and non-negative matrix factorization, particularly on the short texts typical of academic abstracts. Hyperparameters were optimized through random grid search to maximize topic coherence, and outlier reclassification plus manual relabeling refined interpretability. The result was a map of twenty distinct topics spanning the entire field, from healthcare AI and municipal waste optimization to algorithmic fairness and citizen participation in datafied democracies.

But topic modeling alone would simply produce another taxonomy. The study’s central contribution is a functional framework that organizes the literature around four practical functions of public governance: creating public value, delivering public services, responsiveness to the public, and protecting state–society relations. The authors deliberately rejected more familiar organizing axes. Sorting by application domain scatters shared governance problems like bias and accountability across every sector. Sorting by technology tracks capability but ignores the public stakes that make government deployment distinctive. Sorting by governance risk collapses the field into its pathologies. And sorting by public value alone absorbs distinctions a public manager most needs to see. The four governance functions, the authors argue, are the questions a public manager actually asks before deploying a system: does it create value, improve a service, change how the organization hears from citizens, and what does it do to the state’s relationship with the governed?

The framework’s dimensions are enriched by four theoretical lenses that recur throughout the scholarship: technological operationalization, which asks how AI fits specific public problems; function of government, which maps AI techniques onto sectoral activities; the policy-cycle framework, which locates AI within agenda-setting, formulation, implementation, and evaluation; and public governance, which foregrounds who is affected. Notably, the authors show that technological operationalization is not confined to the application dimensions but runs vertically through all four: a model’s opacity, its optimization target, and its hardcoded thresholds are design choices that propagate into accountability deficits and entrenched inequality downstream. Governance failures, they suggest, are often most cheaply addressed upstream at the point of technical design rather than through oversight bolted on after deployment.

The empirical findings reveal a field in dramatic flux. Output grew rapidly through the late 2010s, peaked at 627 works in 2023, and remained elevated through 2025. More striking is the compositional shift. The public value dimension, dominated by domain applications such as energy-efficiency forecasting and air-pollution monitoring, accounted for 50 to 77 percent of works through 2017 but declined to roughly 16 to 22 percent by 2024 and 2025. The state–society relations dimension, encompassing algorithmic fairness, ethics, and regulation, more than doubled its share, from around 12 to 23 percent in the mid-2010s to 52 to 54 percent in 2024 and 2025. The transition concentrates around the public release of ChatGPT in November 2022, which the authors use as a heuristic marker rather than a causal explanation, noting that generative AI diffusion, regulatory momentum such as the EU AI Act, and an ongoing accountability debate all converged in the same period.

The post-2022 pivot is stark in the numbers. Classical domain-application topics shrank from about 19 percent of the pre-ChatGPT corpus to under 5 percent afterward, while several topics essentially became new literatures: ethics and equity in educational AI grew from 2 works to 66, and AI ethics, accountability, and transparency in public service delivery surged from 9 to 291 works, a thirty-one-fold expansion. Semantic analysis of the twenty topics shows that the six state–society topics form a tight, mutually similar cluster, reading as a single coherent research conversation despite spanning multiple domains, while the public value topics spread widely across incompatible technical vocabularies. The framework thus partitions the field asymmetrically, and the authors argue this asymmetry is itself informative for anyone navigating it.

Equally important is where the literature is thin. The responsiveness dimension, concerning where AI can support the policy cycle and democratic feedback, carries only about 462 works across just two topics, making it the smallest dimension despite AI’s obvious relevance to agenda-setting, consultation, and evaluation. Intriguingly, public administration journals are already over-indexing here: responsiveness accounts for about a quarter of works in top-ranked public administration outlets, nearly double its corpus-wide share. The authors sketch four research directions for this space, including how large language models reshaping the information environment of agenda-setting, whether AI-mediated consultation strengthens or hollows the link between citizen voice and government decision, what happens when optimizing tools are embedded in satisficing policymaking processes, and whether AI-driven evaluation genuinely closes the feedback loop or merely produces dashboards.

The framework also doubles as a practical diagnostic for public managers weighing a specific AI system. Each dimension supplies a question and a concrete check: whether the system creates value the public can recognize, whether it improves a service without eroding recourse to a human, whether it widens or narrows the channel through which citizen preferences enter decisions, and whether affected citizens can learn of, understand, and contest its output while responsibility remains attributable. Applied retrospectively to the City of Chicago’s 2016 predictive policing heat list, a system later judged less effective than a simple most-wanted list, the diagnostic surfaces failures on three of four questions, illustrating how design-stage deficits that later evaluations documented could have been identified before deployment.

The study acknowledges its boundaries, including reliance on abstracts rather than full texts, the undercounting of grey literature and government reports, restriction to English-language scholarship, and the possibility of calibration drift in large-language-model screening. Yet within those limits, the contribution is a translation layer: a compact, function-based vocabulary that makes an interdisciplinary literature, previously scattered across engineering, information systems, and public administration, discoverable and comparable across the communities that study it in parallel. As governments worldwide grapple with deploying AI under democratic constraints, the researchers offer not a new theory of public administration but something arguably more useful right now: a shared map of what is known, what is contested, and where the next decade of scholarship and practice most urgently needs to go.

Subject of Research: A computational mapping of artificial intelligence research in the public sector using a functional governance framework and BERTopic modeling

Article Title: Towards artificial intelligence for the public sector: framing and bridging academia and practice

Article References: Mintz, Z. W., & Ma, J. (2026). Towards artificial intelligence for the public sector: framing and bridging academia and practice. Global Public Policy and Governance, 6(2), 124-156. https://doi.org/10.1007/s43508-026-00148-w

Image Credits: AI Generated

DOI: 10.1007/s43508-026-00148-w

Keywords: artificial intelligence, public sector, public governance, algorithmic fairness, BERTopic, systematic literature review, policy cycle, public value, ChatGPT, accountability, e-government, computational social science