AI in Government: Landmark Review Maps Three Decades of Policy Research

Artificial intelligence has moved from the margins of public administration to the centre of how governments draft, deliver, and evaluate policy, yet the research landscape that documents this shift has never been mapped in full. A new systematic review published in AI & Society by Wajahat Mazahar Khan, Areiba Arif, and Naresh Singh of O. P. Jindal Global University offers the most comprehensive picture to date. Using the PRISMA framework, the team retrieved 257 articles published between 1995 and 2025 from the Scopus database and subjected them to rigorous thematic synthesis, combined with a bibliometric analysis that tracks how the field has grown, clustered, and matured over three decades. The result is both a technical audit of scholarship and a warning about the direction of AI-driven governance.

The methodological architecture of the study matters as much as its findings. PRISMA, the Preferred Reporting Items for Systematic Reviews and Meta-Analyses, imposes a transparent, replicable screening pipeline: records are identified, duplicates removed, titles and abstracts screened against inclusion criteria, and full texts assessed before final synthesis. By pairing this with bibliometrics, the quantitative study of publication patterns, citation networks, and keyword co-occurrence, the authors could distinguish genuine research trajectories from passing fashions. Thematic analysis, following the widely used Braun and Clarke approach, then allowed patterns to emerge from the corpus rather than being imposed on it. The six themes that crystallised from this process form the backbone of the review.

The first theme concerns AI-enabled governance across macro-, meso-, and micro-level policy processes. At the macro level, AI shapes national strategy, regulatory design, and the geopolitics of technological competition; at the meso level, it reorganises agencies, procurement, and inter-departmental coordination; at the micro level, it transforms the daily work of street-level bureaucrats who now interpret algorithmic recommendations rather than raw case files. The review draws on scholarship examining predictive policing in Berlin, fraud detection systems modelled through interpretive structural approaches, and the emerging ideal type of the algocracy, in which government organisations delegate consequential decisions to computational pipelines. The multi-level framing is significant because it shows that AI is not a single intervention but a restructuring force that operates simultaneously across scales of the state.

The second theme addresses regulatory and governance frameworks for managing AI systems themselves. Here the literature reveals a striking asymmetry: governments deploy AI faster than they can govern it. The review catalogues a proliferation of ethical guidelines, from the European Commission’s Ethics Guidelines for Trustworthy AI to UNESCO’s 2021 Recommendation on the Ethics of Artificial Intelligence, alongside framework proposals for AI governance that span risk tiers, oversight bodies, and lifecycle accountability. Yet the synthesis also documents deep uncertainty. Work on AI under great uncertainty argues that policymakers must choose decision strategies, precautionary, adaptive, or robust, when the technology’s trajectory is fundamentally unknowable. The review’s verdict is that context-sensitive regulation, tuned to sectoral risk and national capacity, outperforms one-size-fits-all codes that quickly become obsolete.

Ethics forms the third and arguably most charged theme. The corpus includes landmark studies of algorithmic bias: the ProPublica investigation into recidivism prediction software, the Gender Shades study documenting intersectional accuracy disparities in commercial gender classification, and analyses of gender bias embedded in word embeddings and machine translation. It also incorporates foundational critiques, from Algorithms of Oppression to Automating Inequality, showing how high-tech tools can profile, police, and punish the poor. The review highlights that bias is not merely a technical defect to be debiased after the fact; it reflects historical data, design choices, and institutional incentives. Accountability emerges as the central unsolved problem: transparency alone, as the literature on algorithmic accountability cautions, does not guarantee that affected citizens can contest or remedy automated decisions.

The fourth theme examines socioeconomic impacts on welfare systems and labour markets. Studies synthesised in the review cover automation’s differentiated effects across countries, the data-driven workplace and the case for worker technology rights, the future of labour unions at the dawn of AI, and national strategies such as India’s AI for inclusive growth agenda. The evidence points to distributional consequences that vary sharply by region and skill profile, with scholarship on algorithmic colonisation and decolonial AI warning that the benefits and risks of AI are unevenly distributed between the Global North and South. For welfare policy specifically, the review suggests that AI can expand access and targeting precision, but only if safeguards against exclusion errors and digital surveillance are built in from the start.

Public trust, perception, and social acceptance constitute the fifth theme, and the review treats them as empirical questions rather than rhetorical ones. Experimental work shows that explainability and perceived causability measurably affect trust in automated decision-making, and that algorithmic transparency shapes how trustworthy citizens judge a system to be, even when the outcome is unfavourable. Entertainment media, meanwhile, play an underappreciated role in forming public understanding of AI before citizens ever encounter a government algorithm. Studies of electronic monitoring in public and private sectors, and of context-dependent responses to AI on the street, indicate that acceptance is situational: the same technology may be welcomed in one domain and resisted in another. Trust, the synthesis concludes, is a governance resource that must be earned through demonstrated fairness, not assumed through deployment.

The sixth theme, AI in crisis management and real-time policy evaluation, has gained urgency since the COVID-19 pandemic. The corpus includes comparative analyses of digital tracing technologies in Germany, Norway, and the United Kingdom, examinations of how governments legitimised emergency surveillance, and warnings that excessive digital surveillance and privacy invasion now constitute a creeping crisis in their own right. Research on Japan’s punctuated politics of digital transformation and on AI-assisted expert advisory during the pandemic in Morocco shows how crises accelerate technology adoption while compressing deliberation. The review also notes emerging work using artificial neural networks as public policy evaluation methods, hinting at a future where policy feedback loops run continuously rather than at electoral intervals, with all the accountability questions that implies.

What unites the six themes, and what the authors present as the review’s central finding, is that the challenges of AI in public policy extend well beyond efficiency and technical design. The deepest questions concern legitimacy, accountability, and the relationship between citizens and the state. When an algorithm allocates benefits, predicts risk, or evaluates policy performance, it does not merely process data; it redistributes discretion, redefines evidence, and alters who must answer to whom. The bibliometric record shows scholarship racing to keep pace, with publication volume accelerating sharply in recent years, but the thematic synthesis reveals persistent gaps: limited empirical evaluation of deployed systems, thin coverage of Global South governance capacity, and few studies that follow algorithmic decisions through to their effects on real citizens.

The review closes with a three-part agenda that reads as a roadmap for the next decade of research and practice. First, regulatory approaches must be context-sensitive, calibrated to the specific risks, capacities, and institutional cultures of individual sectors and states rather than imported wholesale. Second, ethical safeguards need strengthening from voluntary principles toward enforceable mechanisms, with particular attention to bias, accountability, and inclusivity at every stage of the AI lifecycle. Third, governance structures must become more inclusive, bringing affected communities, workers, and marginalised groups into the design and oversight of the systems that judge them. For a field that has spent thirty years asking what AI can do for government, the review’s most provocative contribution may be redirecting the question toward what AI-governed states owe the people they serve.

Subject of Research: A systematic review and bibliometric analysis of artificial intelligence applications in public policy

Article Title: Artificial intelligence and public policy: a systematic review with bibliometric insights

Article References: Mazahar Khan, W., Arif, A., & Singh, N. (2026). Artificial intelligence and public policy: a systematic review with bibliometric insights. AI & SOCIETY. https://doi.org/10.1007/s00146-026-03363-5

Image Credits: AI Generated

DOI: 10.1007/s00146-026-03363-5

Keywords: artificial intelligence, public policy, governance, systematic review, bibliometrics, algorithmic accountability, AI ethics, algorithmic bias, public trust, crisis management, labour markets, AI regulation