The distinction the authors draw between AI and earlier waves of digital government is technical as well as institutional. Previous generations of e-government technology primarily improved the efficiency, connectivity, and accessibility of administrative processes, digitising forms and speeding up workflows while leaving the underlying logic of human decision-making intact. AI, by contrast, increasingly performs the knowledge work of administration: it predicts, classifies, optimises, and generates scenario analyses that feed directly into judgements about rights, resources, and sanctions. The central analytical question, the review argues, is therefore no longer how governments digitalise administrative processes, but how administrative intelligence is governed and distributed across human and computational actors. In the authors’ framing, AI is best understood not as a discrete application but as a systems phenomenon whose consequences ripple across power structures, public values, institutions, decision-making, organisations, and frontline bureaucracy.

That systems perspective, drawing on the tradition of systems thinking associated with Barry Richmond and Donella Meadows, rejects the idea that AI’s effects can be assessed by examining any single agency or decision point in isolation. An algorithm deployed in one part of government depends on professional norms, institutional rules, data infrastructures, and frontline judgements that translate algorithmic outputs into real-world action. To capture these interdependencies, the editorial proposes a six-dimension framework covering the full administrative life cycle of AI. The framework moves from the strategic conditions under which governments acquire AI capacity, through the public values used to justify adoption, the institutions that authorise its use, the decisions officials make with it, and the organisations that implement it, to the street-level encounters where algorithms meet citizens. Each dimension, the authors insist, is a location where AI enters, circulates, and becomes institutionalised in government.

The first dimension, power structures, may be the most politically charged. The review argues that public administration scholarship has too often examined AI only after it arrives inside government, ignoring the political economy of the systems being procured. Governments frequently do not own the capabilities their AI depends on: digital infrastructure, specialist talent, computational resources, and research pipelines are concentrated in large technology firms. Drawing on scholarship on platform capitalism and technological sovereignty, the authors describe a fundamental asymmetry in technical capacity, epistemic authority, and regulatory knowledge between states and major tech companies. This dependence shapes what AI systems are available to governments, what officials can know about risk, bias, and explainability, and how much bargaining power agencies have when seeking access to proprietary source code or audit evidence. Supporting research in the special issue, using OpenAlex bibliometric data, finds that Big Tech’s presence in critical technology research is growing, particularly in AI, raising questions about who controls the knowledge on which public-sector AI increasingly rests.

Public value forms the second and normative core of the framework. The authors stress that in government, unlike the commercial sector, AI adoption cannot be justified by productivity or cost reduction alone; it must serve public purposes such as fairness, transparency, responsiveness, and trust. An algorithm that improves technical accuracy does not settle questions of proportionality or legitimacy, because public agencies must justify how decisions are reached and whose interests are served. A service that reduces processing time but increases exclusion or distrust cannot simply be counted as administrative improvement. Contributing papers in the collection develop this argument in detail: one proposes a value-chain approach for tracking where public value is created or lost across the stages of an AI-enabled service, while another maps 3,268 studies using BERTopic modelling and finds that after 2022, work on algorithmic fairness, ethics, and state-society relations has overtaken earlier emphases on efficiency and service delivery.

Institutions constitute the third dimension, translating public purposes into enforceable rules about who may use AI, under what conditions, and subject to which review. The editorial highlights a structural puzzle: most regulatory sandboxes were designed for private-sector innovation, where regulators supervise firms testing products, but in public-sector AI the state is simultaneously regulator, user, implementer, and accountable authority. One contributing paper adapts the sandbox model into an experimentalist governance framework for government AI, emphasising supervised testing, stakeholder participation, iterative learning, and institutional feedback. The authors connect this to Elinor Ostrom’s nested levels of institutional rules, showing how operational rules govern day-to-day AI use, collective-choice rules govern how those rules are revised, and constitutional-choice rules keep experimentation consistent with administrative law, procurement requirements, and rights protection. Institutions, in this account, do not merely permit or prohibit AI; they define the terms under which algorithmic systems can be tested, trusted, and treated as legitimate parts of public administration.

The framework’s fourth dimension turns to administrative decisions, where the editorial examines experimental evidence on whether public officials actually accept AI-assisted judgement, particularly in budgeting, a domain long understood as political rather than purely technical. The review contrasts two views of human-AI collaboration: augmentation, in which algorithms provide forecasts and rankings without displacing human responsibility, and de facto automation, in which officials defer to algorithmic outputs they lack the expertise or organisational backing to challenge. Extending Nathan Caplan’s two-communities theory, the authors suggest that technical experts, vendors, and data scientists may produce algorithmic knowledge that bureaucrats do not regard as usable or congruent with their problems, producing a decoupling between formal AI adoption and practical utilisation. Bureaucratic acceptance, they argue, is an overlooked condition for successful AI reform: officials must come to see algorithmic inputs as credible, legitimate, and compatible with the responsibilities attached to public office.

Organisations and street-level bureaucrats complete the framework. Implementation, the editorial stresses, is rarely smooth: public agencies inherit legacy systems, fragmented data arrangements, risk-averse cultures, and budget constraints, so AI reforms may look successful at launch yet lose momentum as maintenance costs rise and staff revert to familiar practices. Contributing research on AI-powered platform government shows that organisational inertia is not mere resistance but a mechanism through which prior investments and routines govern the pace and direction of AI use. At the frontline, drawing on Michael Lipsky’s theory of street-level bureaucracy, the authors argue that AI does not eliminate discretion but redistributes its materials, changing what officials see, how cases are categorised, and how departures from system recommendations must be justified. Ethnographic work in Finnish public organisations introduces the concept of the moral crumple zone: bureaucrats become accountability buffers who absorb blame for algorithmic failures while lacking meaningful control over the systems that shape their decisions. Experimental evidence from policing further finds that AI-supported enforcement can shift officers toward more formal, coercive styles, depending on the accountability mechanisms in place.

The editorial’s conclusion is stark in its ambition: AI represents a genuine paradigm shift rather than the latest phase of digital government, because it begins to change the nature of administration itself. If digital government changed the machinery of administration, the authors write, AI is displacing the human monopoly over administrative judgement. The review closes with a forward research agenda calling for scholarship to move from national AI policy to AI ecosystems, from AI ethics to AI legitimacy, from adoption to institutional capability, and from street-level discretion to behavioural adaptation. The decisive question for democratic governance, the authors argue, is no longer how governments digitalise administrative processes, but how they govern administrative systems in which computational actors increasingly participate alongside human ones, and whether citizens can still recognise, understand, and challenge the decisions that shape their lives.

Subject of Research: A systems framework for understanding how artificial intelligence enters, circulates, and becomes institutionalised across the public administration system.

Article Title: Artificial intelligence across the public administration system: public values, institutions, and bureaucrats in the algorithmic age

Article References: Cho, W., Fan, Z., & Zheng, Y. (2026). Artificial intelligence across the public administration system: public values, institutions, and bureaucrats in the algorithmic age. Global Public Policy and Governance, 6(2), 103-123. https://doi.org/10.1007/s43508-026-00151-1

Image Credits: AI Generated

DOI: 10.1007/s43508-026-00151-1

Keywords: artificial intelligence, public administration, public value, bureaucracy, street-level bureaucracy, algorithmic governance, technological sovereignty, regulatory sandboxes, digital government, administrative discretion, AI accountability, public sector innovation

Screening tools such as the federal Climate and Economic Justice Screening Tool and various state-level equivalents are designed to synthesize large volumes of environmental, demographic, and socioeconomic data into a single judgment about whether a community qualifies as disadvantaged. On the surface, this seems like a straightforward task of measurement. In practice, the researchers show, it involves a cascade of decisions, each of which can shift outcomes substantially. Analysts must decide which indicators to include, how to combine them, whether to compare communities at the national or state level, which geographic units to use as the basis of analysis, and what thresholds separate the designated from the undesignated. None of these decisions is dictated by data alone; each reflects a policy judgment about what disadvantage means and how it should be recognized.

The study systematically varies these methodological choices and measures how sensitive the resulting maps of disadvantage are to each one. The results are striking. Depending on the combination of choices made, the same underlying data can produce markedly different sets of designated communities, with some methodological configurations flagging far larger populations than others. The geographic unit of analysis emerges as one of the most powerful levers. Tools that operate on census tracts can identify pockets of disadvantage that are invisible when counties serve as the unit, while county-level aggregation can dilute concentrated hardship within larger, more heterogeneous areas. Conversely, very fine-grained units can be sensitive to boundary artifacts and small-sample noise in survey data.

Threshold selection proves equally consequential. Many screening tools designate a community as disadvantaged if it exceeds a percentile cutoff on one or more indicators, for example falling within the top quarter of all communities for pollution burden or the bottom quarter for income. Moving a cutoff even modestly can add or remove thousands of communities from the designated list. The researchers demonstrate that these threshold effects are not uniform across the country: in densely populated urban regions, small changes in cutoffs translate into large swings in the number of affected residents, while in rural areas the same changes may alter designations for vast land areas but relatively few people. The practical stakes of a seemingly technical parameter therefore differ enormously depending on where a community sits.

The choice of indicators themselves introduces another layer of variability. Some tools emphasize environmental exposure measures such as air toxics concentrations, particulate matter levels, and proximity to hazardous facilities. Others weight socioeconomic vulnerability more heavily, incorporating poverty rates, educational attainment, housing costs, linguistic isolation, and health prevalence data. Because these dimensions of disadvantage only partially overlap, a community that scores as severely burdened on an exposure-centered tool may fail to qualify under a vulnerability-centered one, and vice versa. The study shows that the correlation between designations produced by different indicator sets is far from perfect, meaning that communities can be treated inconsistently across programs even within the same jurisdiction.

Aggregation methods add further complexity. When multiple indicators must be combined into a composite score, analysts must choose between approaches such as averaging, summing binary flags, or requiring that a community exceed thresholds on multiple categories simultaneously. These choices encode different assumptions about whether disadvantages are interchangeable, whether they compound, or whether certain burdens are non-negotiable. A tool that designates a community when any single indicator crosses a threshold will cast a far wider net than one that requires burdens across several categories at once. The researchers find that this single design decision can be as influential as the choice of data sources, reshaping the designated population by substantial margins.

These methodological variations matter because designation carries real consequences. Disadvantaged community status increasingly serves as a gatekeeper for funding under major climate and infrastructure legislation, influencing where investments in clean energy, transit, water systems, and resilience projects flow. Communities that fall just outside a designation may be excluded from programs despite facing conditions nearly identical to those just inside the line. The study highlights how such boundary effects can produce sharp discontinuities in eligibility between neighboring areas, raising questions about fairness and administrative coherence. When two adjacent neighborhoods with similar pollution burdens and incomes receive different treatment because of where a percentile cutoff happens to fall, the legitimacy of the screening exercise is called into question.

The authors argue that the solution is not to search for a single objectively correct methodology, since every design choice involves legitimate value judgments, but to make those judgments transparent, deliberate, and accountable. They recommend that agencies document the rationale behind indicator selection, threshold placement, and geographic choices, and that they test the sensitivity of their designations to reasonable alternative configurations. Publishing designation maps alongside uncertainty or sensitivity analyses would allow policymakers, advocates, and residents to understand how robust a given designation is, and would help identify communities that hover near eligibility boundaries and may warrant case-by-case review. The study also suggests that tools could be designed with explicit equity objectives in mind, choosing methodological configurations that align with the distributive goals of the programs they serve rather than treating technical defaults as neutral.

For the growing community of researchers and practitioners working at the intersection of data science and environmental policy, the study offers a caution about algorithmic governance more broadly. Screening tools compress complex, multidimensional social and environmental conditions into binary categories, and that compression inevitably involves choices that shape outcomes. The lesson is not that such tools should be abandoned, since they bring consistency, scale, and defensibility to decisions that would otherwise be made ad hoc, but that their architecture deserves the same scrutiny as the policies they implement. As more governments at every level adopt screening tools to operationalize justice commitments, the methodological details examined in this research will increasingly determine whether those commitments reach the communities they were intended to serve, or whether they dissolve into the fine print of percentile cutoffs and geographic units chosen without deliberation.

Subject of Research: How methodological variations in environmental justice screening tools affect disadvantaged community designations

Article Title: Methodological Variations in Environmental Justice Screening Tools and Their Impact on Disadvantaged Community Designations

Article References: Robbins, T., Li, Q., Bird, S., & Powers, S. E. (2026). Methodological Variations in Environmental Justice Screening Tools and Their Impact on Disadvantaged Community Designations. Nature Communications. https://doi.org/10.1038/s41467-026-77824-2

Image Credits: AI Generated

DOI: 10.1038/s41467-026-77824-2

Keywords: environmental justice, screening tools, disadvantaged communities, Climate and Economic Justice Screening Tool, policy designations, census tracts, threshold sensitivity, indicator selection, algorithmic governance, environmental policy, funding allocation, geographic units

The study arrives at a moment when AI-based literature has gained unprecedented traction, particularly since 2022, and when automated decision-making has spread across business functions from human resource management to customer service platforms. Vulnerable and minority employees frequently face systemic barriers that include discrimination, underrepresentation in leadership positions, limited access to career advancement, and insufficient organisational support. AI-based recruitment, performance evaluation, and career development systems promise to reduce these inequities by emphasising standardised, data-driven decisions that minimise the influence of unconscious human prejudice. The technology can also support inclusive job matching, skill development, and accessibility initiatives, empowering workers who have traditionally experienced disadvantage. Yet poorly designed or unmonitored systems risk creating feedback loops that systematically disadvantage the very groups they could help, as high-profile cases of recruitment algorithms reproducing gender and racial bias from biased historical data have demonstrated.

To map this contested terrain, the review employed an unusually rigorous multi-phase methodology. The author conducted a keyword-based literature search in the Scopus database on August 10, 2025, using a Boolean query that combined terms for artificial intelligence, machine learning, and algorithmic decision-making with terms covering vulnerable and minority employee populations, human resource management, and ethics. The initial search identified 554 articles; after excluding 308 records that were not the required publication types and 9 articles not in English, 237 articles entered bibliometric mapping and principal component analysis. A final manual thematic synthesis drew on 30 articles published in A*, A, and B-ranked journals according to the Australian Business Deans Council index. The full PRISMA-guided screening process was documented to ensure transparency, and all 237 articles were included in the statistical analyses to capture the interdisciplinary breadth of the field, spanning information systems, AI ethics, disability studies, and organisational behaviour.

The quantitative core of the review combined VOSviewer bibliometric mapping with principal component analysis. Keyword co-occurrence analysis identified 93 recurring keywords, and a Kaiser-Meyer-Olkin test of sampling adequacy, with a threshold of 0.50, filtered these down to 16 keywords for the final PCA. Applying Kaiser’s criterion of eigenvalues greater than 1, the analysis extracted six principal components, which the author labelled as social equity and representation, ethical governance and AI technology, inclusion and organisational adaptation, employment and knowledge transformation, structural challenges in workforce diversity, and goals and workplace realities. Scree plots illustrated the effect of the dimensionality reduction. These statistically derived components were then consolidated, through qualitative interpretation, into five overarching themes that structure the review’s synthesis: AI adoption and workforce transformation; bias, equity, and fairness in AI systems; employee well-being, inclusion, and empowerment; ethical, legal, and governance considerations; and methodological approaches and tools.

Publication trends reveal how young the field is. Minimal contributions appeared between 2006 and 2019, but a sharp upward trajectory began in 2020, with publications rising to six that year and eight by 2023. Knowledge Management Research and Practice leads the top ten journals by total citations with 305, followed by Informing Science with 183 and the Journal of Information, Communication and Ethics in Society with 167. The theoretical landscape underpinning the literature is rich but fragmented, and the review groups it into four domains: ethics, justice, and fairness; organisational and human resource management; technology and digital transformation; and disability, diversity, and inclusion frameworks. The first domain includes procedural justice, algorithmic fairness, and responsible AI innovation models. The second draws on the Resource-Based View, the Dynamic Capability Framework, and strategic human resource management perspectives. The third relies on technology adoption and algorithmic management theories, while the fourth deploys frameworks such as Disability Justice, Feminist Design Thinking, neurodiversity models, and identity-consciousness versus identity-blindness approaches that position marginalised employees not merely as subjects of algorithmic governance but as knowledge holders and co-designers of inclusive AI systems.

Among the review’s most striking empirical findings is evidence that AI-generated communications can provoke stronger negative reactions toward employees with disabilities or women than traditional human-based bias, exceeding it in some contexts. Research on construction and engineering leadership documents a likeability versus competency dilemma, in which women with comparable qualifications and experience are perceived as less likeable than male peers. In digital skills and STEM training programmes, recruiters using AI-based candidate screening were found to favour male candidates during initial outreach, especially under high workloads, demonstrating how such systems can reinforce gender-based disparities even before applications are submitted. Studies of AI-assisted disability assessments show that biases may emerge from design choices, data selection, or operational deployment, underscoring the importance of participatory, disability-led design practices. Meanwhile, profiling models in public employment services can inadvertently misclassify and discriminate against minority or foreign-origin jobseekers, exposing a sharp trade-off between accuracy and equity.

The review also documents AI’s genuine potential for empowerment. AI-driven innovation can enhance transparency, strengthen internal controls, and shape workplace culture and performance evaluation systems, while applications in policy evaluation, exemplified by China’s Low-Carbon City Pilot program, show how algorithmic tools can improve job quality, entrepreneurship opportunities, and urban labour inclusivity. Healthcare studies demonstrate that AI systems designed with clear reasoning, adaptive triage, and data transparency can reduce cognitive burdens while promoting equitable outcomes for diverse employees. Research on workforce diversity, equity, and inclusion in healthcare further indicates that improvements across demographic and experiential dimensions correlate with better patient safety outcomes, particularly in regions with diverse patient populations. Studies of corporate diversity statements show that companies emphasising identity-conscious topics receive more favourable employee evaluations of DEI, and a tri-balance framework for AI in personnel selection illustrates how efficiency, fairness, and stakeholder voice might be reconciled.

From these threads the review advances an integrative framework built on three interacting dimensions: the technological mechanisms of AI systems, organisational and governance mediators, and employee outcomes. The framework’s central claim is that empowerment outcomes are not determined solely by the technologies themselves but by the interaction between algorithmic design, organisational implementation practices, and governance structures. When transparency, fairness auditing, and inclusive organisational policies are implemented appropriately, AI systems may support more equitable and empowering workplace environments. When they are not, the same systems can entrench exclusion. This reframing positions AI not as a neutral technical artefact but as a sociotechnical governance issue shaped by organisational values and institutional structures, an account that aligns closely with emerging information systems scholarship on responsible AI and algorithmic governance. The author also notes, however, that the field’s empirical contributions remain fragmented and uneven, with equity research often disconnected from employee experiences and methodological innovations rarely integrated into discussions of ethics or well-being.

The review acknowledges its own limitations, including reliance on a single database, the imperfect reach of keyword-based searches, the possible influence of journal quality filters on corpus composition, and the inherent interpretive judgments of qualitative synthesis. Yet its research agenda is ambitious. The author calls for future work to connect social role and congruity theories with strategic human capital and corporate governance frameworks, to combine feminist design thinking with identity-consciousness debates, and to integrate neurodiversity and disability-led design perspectives with digital transformation research. On the empirical side, the review urges multi-country, longitudinal, and high-dimensional fixed-effects models to capture institutional, cultural, and regulatory heterogeneity, noting that labour laws, political climate, and social norms may shape how AI and DEI initiatives are implemented and received. With most current studies relying on cross-sectional or single-country designs, the consequences, rather than merely the determinants, of AI-augmented diversity management remain largely unexplored. What is already clear, the review concludes, is a fundamental tension at the heart of workplace AI: whether the technology becomes an instrument of algorithmic exclusion or a genuine engine of empowerment will depend on choices, about design, oversight, and governance, that organisations are making right now.

Subject of Research: A systematic review of how artificial intelligence affects the empowerment, equity, and inclusion of vulnerable and minority employees in the workplace

Article Title: Artificial Intelligence and the Empowerment of Vulnerable and Minority Employees: A Systematic Review

Article References: Pokharel, P. R. (2026). Artificial Intelligence and the Empowerment of Vulnerable and Minority Employees: A Systematic Review. Information Systems Frontiers. https://doi.org/10.1007/s10796-026-10823-2

Image Credits: AI Generated

DOI: 10.1007/s10796-026-10823-2

Keywords: artificial intelligence, vulnerable employees, minority employees, algorithmic bias, diversity equity and inclusion, systematic review, human resource management, algorithmic governance, employee empowerment, workplace transformation, responsible AI, information systems

The research was conducted by Ge Wang, Haixin Teng, and Zengyang Xu of the School of Public Administration at Central China Normal University in Wuhan, and it addresses one of the most persistent blind spots in the study of algorithmic governance. Scholars have long argued that automation transforms street-level bureaucracy, echoing Michael Lipsky’s classic insight that frontline workers are the ultimate policymakers because they exercise discretion in every individual encounter. Earlier work by Bovens and Zouridis described a drift from street-level to system-level bureaucracy, in which decision-making migrates from human judgment to automated systems. But the new study asks a sharper question: what happens to how bureaucrats enforce — not just what they decide — when AI enters the loop, and how does felt accountability reshape that relationship?

To answer it, the team designed a preregistered survey experiment embedded in the real-world context of honking violations, a common and familiar enforcement scenario for traffic police. Participants were 356 street-level police officers drawn from a representative sample, making this one of the more methodologically robust attempts to measure enforcement style directly among practitioners rather than students or hypothetical respondents. Invalid questionnaires were excluded before hypothesis testing, based on substantial missing data, incomplete experimental responses, or failure to pass embedded attention-check items, and the experimental design was preregistered to guard against selective reporting. Data supporting the findings are available from the corresponding author upon reasonable request, and the preregistration is publicly archived on the Open Science Framework.

The experiment crossed two manipulations: whether officers received AI support in the enforcement scenario, and which type of accountability they experienced. In public administration, accountability is conventionally divided into process accountability, where officials must justify the procedures and reasoning behind their decisions, and outcome accountability, where they are judged by the results those decisions produce. Decades of psychological research, from Tetlock onward, have shown that these two forms of felt accountability trigger distinct information-processing strategies, often in opposite directions. The Chinese enforcement context makes the setting particularly instructive, given its documented history of campaign-style enforcement and shifting regulatory styles among frontline officials.

The core findings are cleanly delineated. AI-supported enforcement, on its own, pushed officers toward a more formal and coercive style of enforcement — rule-invoking, sanction-first, procedural — while weakening the educational component, the practice of explaining, persuading, and teaching violators why compliance matters. This matters because enforcement style is not cosmetic. A substantial body of regulatory scholarship, including work by May and Wood on inspectors at the regulatory front lines and by Braithwaite and colleagues on enforcement pyramids, links style directly to compliance outcomes, citizen trust, and the legitimacy of the state. A coercive turn among officers algorithmically nudged into formalism could therefore ripple outward into how millions of daily encounters between citizens and the state actually feel.

Accountability, meanwhile, produced effects of its own that align closely with prior experimental literature. Officers who felt process accountability gravitated toward a more educational, prioritization-focused, and accommodative enforcement style — they reasoned more about how and why they enforced, reserved harsh measures for the cases that mattered most, and left more room for discretion and dialogue. Officers under outcome accountability moved in the opposite direction, reinforcing a more formal and coercive posture, consistent with the idea that being judged purely on results encourages defensive, box-ticking enforcement designed to be blame-proof.

The most consequential and arguably most surprising result is the interaction. Process accountability did not buffer officers against the algorithmic push toward formalism and coercion; it intensified it. Under process accountability, the presence of AI support further reinforced formal and coercive enforcement while further weakening the educational approach. The authors’ interpretation, grounded in the combined-effects framework they develop, is that when officers must justify their processes, the documented, auditable recommendation of an algorithm becomes an attractive anchor — a defensible procedural basis for action. Following the machine feels procedurally safe, and in doing so officers may shed the relational, educative practices that algorithmic systems cannot capture or document. The accountability mechanism designed to make officers more thoughtful may, in the presence of AI, make them more mechanical.

These findings carry immediate implications for governments racing to deploy AI in policing, from automated violation detection to algorithmic risk assessment and case triage. The study suggests that technology procurement and accountability reform cannot be designed in isolation. An agency that pairs algorithmic enforcement tools with process-oriented oversight — often considered the ethically preferable arrangement — may inadvertently deepen the very dehumanization of frontline encounters that critics of algorithmic governance fear. Conversely, outcome accountability’s independent drift toward coercion compounds the problem. If policymakers want AI to augment rather than hollow out street-level discretion, the results imply that accountability frameworks must be engineered to explicitly protect and incentivize the educational, judgment-rich dimensions of enforcement that machines neither practice nor document.

The study also speaks to a growing theoretical conversation about artificial discretion — the hybrid of human and machine judgment described by Young, Bullock, and Lecy — and about why bureaucrats trust AI recommendations, with recent experimental work suggesting that confirmation of professional judgment drives algorithmic acceptance. By demonstrating that enforcement style is a joint product of technology and institutional design, the Chinese policing experiment moves the field beyond the simple question of whether bureaucrats accept algorithmic advice toward the richer question of what they become when they do. As cities worldwide wire AI into everything from traffic cameras to welfare fraud detection, the transformation of the frontline worker may prove to be the most important, and least visible, policy outcome of the algorithmic state.

Subject of Research: Experimental study of how AI support and accountability mechanisms transform street-level police enforcement styles

Article Title: Artificial intelligence, accountability mechanisms, and the transformation of street-level enforcement style: experimental evidence from frontline policing

Article References: Wang, G., Teng, H., & Xu, Z. (2026). Artificial intelligence, accountability mechanisms, and the transformation of street-level enforcement style: experimental evidence from frontline policing. Global Public Policy and Governance, 6(2), 176-198. https://doi.org/10.1007/s43508-026-00146-y

Image Credits: AI Generated

DOI: 10.1007/s43508-026-00146-y

Keywords: artificial intelligence, street-level bureaucracy, policing, enforcement style, accountability, process accountability, outcome accountability, algorithmic governance, police discretion, survey experiment, public administration, coercion

The research, conducted by Wonhyuk Cho of Ewha Womans University in Seoul and Danuvas Sagarik of the National Institute of Development Administration in Bangkok, appears in the journal Global Public Policy and Governance. The authors set out to address a gap that has grown as governments around the world, from Thailand with its digital government development plan to agencies across the European Union, experiment with embedding machine learning in administrative decision-making. While a substantial literature has documented the efficiency gains that AI can deliver in public services, far less is known about how bureaucrats themselves respond when algorithms are proposed for budgetary choices, a domain that concerns not merely productivity but the fundamentally distributive question of who receives funding and who does not.

The institutional viability of any algorithm-enabling budgeting reform, the authors argue, hinges on whether the bureaucrats charged with carrying it out view it as preferable to the alternatives. History offers plenty of cautionary tales on this point. Public administration research has repeatedly shown that bureaucratic organizations resist reforms perceived as threatening, whether those reforms involve shared service centers or broader restructuring programs, and that the success or failure of administrative change often depends on securing cooperation from insiders. If civil servants quietly oppose algorithmic budgeting, even the most technically sophisticated systems could stall in implementation.

To measure these preferences, the researchers implemented a three-arm survey experiment involving a large sample of 3,820 public sector personnel. Respondents were randomly assigned to read vignette scenarios describing one of three reform pathways: the adoption of AI-assisted budgeting, an expansion of the government workforce, or the introduction of participatory budgeting, in which citizens help decide spending priorities. Random assignment ensures that any differences in reported support across the three groups can be attributed to the reform scenario itself rather than to pre-existing differences among respondents, the standard logic of experimental design in the social sciences.

The headline finding is striking. In analyses restricted to respondents who correctly recalled their assigned treatment and weighted to account for differential treatment recall, AI-assisted budgeting reforms attracted significantly higher bureaucratic support than workforce expansion. This suggests that, at least among the civil servants surveyed, the prospect of algorithmic assistance in allocating public funds is not met with the resistance that fears of automated job displacement might predict. Instead, bureaucrats appear to view AI as a more attractive reform than hiring additional personnel, perhaps because algorithmic tools promise to augment their capacity without the organizational disruptions, coordination costs, and budgetary competition that come with expanding the payroll.

When it came to participatory budgeting, however, the picture was more nuanced. The robustness analyses found no statistically significant differences in bureaucratic support between the AI-assisted budgeting treatments and the participatory budgeting treatments. In other words, civil servants were roughly equally comfortable with delegating budgetary insight to algorithms and with opening budgetary decisions to citizen participation. This equivalence is notable because the two reforms embody very different theories of legitimacy: one rests on technical optimization and data-driven objectivity, while the other rests on democratic inclusion and deliberation. The finding hints that bureaucrats may judge reforms less by their philosophical underpinnings than by more practical considerations of workload, discretion, and administrative feasibility.

The authors were careful to probe the robustness of their results. Beyond the manipulation-restricted analyses, they estimated intent-to-treat effects, which include all randomized respondents regardless of whether they remembered their assigned scenario, adjusting for covariates and incorporating organizational fixed effects to account for differences across the agencies and institutions in which respondents work. Under this more conservative specification, the estimates did not show statistically significant differences across the outcome dimensions. The divergence between the two analytical strategies underscores a familiar lesson in experimental social science: results can be sensitive to how treatment recall and analytic choices are handled, and conclusions about bureaucratic preferences should therefore be drawn with appropriate caution.

The study sits within a rapidly expanding research landscape on AI in government. Previous work has documented automation bias and selective adherence to algorithmic advice among public sector decision-makers, showing that street-level bureaucrats tend to trust AI recommendations when those recommendations confirm their existing professional judgment. Other studies have mapped the barriers to AI adoption in public organizations, examined how AI is reshaping the role of bureaucrats in different organizational contexts, and explored how public values such as efficiency and equity shape civil servants’ willingness to use AI to reduce administrative burdens. Citizen-facing research has also flourished, with survey experiments revealing when and why the public accepts the use of AI in services such as policing and local government. What distinguishes the new study is its focus on budgeting, the heart of distributive governance, and its head-to-head comparison of AI against rival reform pathways rather than against the status quo.

The implications for policymakers are significant. Governments contemplating algorithmic budgeting often worry about backlash from public employees, whose cooperation is essential for data collection, model validation, and the day-to-day operation of any decision-support system. The evidence suggests that such fears may be overblown, at least in comparative perspective: bureaucrats do not appear to view AI-assisted spending as uniquely threatening. Yet the absence of a significant advantage over participatory budgeting also suggests that algorithmic reform is not a slam dunk. Reformers cannot assume that AI carries inherent legitimacy among the administrative workforce; it competes on roughly equal footing with democratic alternatives. The practical lesson may be that the success of AI in fiscal governance will depend less on winning bureaucratic hearts and minds than on careful system design, transparent safeguards, and clear communication about how algorithmic recommendations relate to human discretion.

As governments worldwide continue to draft national AI strategies and embed machine learning in everything from tax policy optimization to healthcare allocation, understanding the preferences of the people who run the administrative state becomes ever more important. This study provides some of the first experimental evidence that, when given a choice between algorithmic budgeting and simply hiring more staff, bureaucrats lean toward the machines. Whether that preference translates into successful implementation, and whether it holds across countries, sectors, and levels of government, remains an open question that future research will need to answer.

Subject of Research: Bureaucratic support for AI-assisted public budgeting compared with workforce expansion and participatory budgeting

Article Title: Do bureaucrats prefer algorithmic budgeting? Experimental evidence on bureaucratic support for AI-assisted public spending

Article References: Cho, W., & Sagarik, D. (2026). Do bureaucrats prefer algorithmic budgeting? Experimental evidence on bureaucratic support for AI-assisted public spending. Global Public Policy and Governance, 6(2), 157-175. https://doi.org/10.1007/s43508-026-00145-z

Image Credits: AI Generated

DOI: 10.1007/s43508-026-00145-z

Keywords: artificial intelligence, public budgeting, bureaucrats, participatory budgeting, public administration, survey experiment, algorithmic governance, fiscal policy, government reform, public finance, civil service, prefer

Public sector AI experimentation inverts that division of roles. When a tax authority pilots an algorithm to detect suspected fraud, or a welfare agency deploys machine learning to prioritize benefit claims, the state is simultaneously the innovator, the regulator, the evaluator, and often the sole affected counterparty for citizens. There is no external firm whose market entry must be facilitated, and no conventional consumer making a voluntary purchase decision. Instead, the people affected are frequently captive audiences: taxpayers, benefit claimants, asylum seekers, and patients who cannot opt out of interacting with the state. This means that the consumer protection rationale at the heart of the traditional sandbox translates only imperfectly. Protecting a consumer from a faulty financial product is meaningfully different from protecting a claimant from a biased eligibility algorithm, because the claimant’s interaction with the state implicates constitutional rights to due process, equal treatment, and administrative justice rather than market fairness alone.

The experimentalist governance tradition offers a useful lens for understanding what sandboxes are meant to accomplish, and also where they fall short. Experimentalist frameworks, as developed in the scholarship the article engages with, are characterized by a recursive loop: broad framework goals are set, local actors are given discretion to pursue those goals through experimentation, results are monitored and reported upward, and the framework goals are then revised in light of what was learned. Peer review, benchmarking, and iterative revision are central. A sandbox fits this logic naturally, because it generates evidence about an innovation under controlled conditions and feeds that evidence back into regulation. But experimentalism also presupposes a degree of independence between the experimenting unit and the monitoring unit. When the state experiments on its own administrative processes, the monitoring function risks becoming self-review, which is precisely the structural weakness that administrative law doctrines such as impartial decision-making and independent appeals were designed to counteract.

The pacing problem that motivates sandbox adoption is particularly acute for AI in government. Legal scholarship has long observed that innovation outpaces regulation, but AI compresses development cycles to a degree that reactive, statute-by-statute lawmaking cannot match. An algorithmic system can be retrained, redeployed, and materially altered in behavior within weeks, while legislative amendment takes years. Moreover, the technical properties of AI systems strain established legal categories. Opacity complicates the duty to give reasons for administrative decisions, a cornerstone of administrative law across many jurisdictions. Statistical bias complicates equality guarantees, because discrimination may emerge from training data rather than from any identifiable discriminatory intent. Distributed development pipelines complicate liability attribution, since a government agency, a commercial vendor, and an open-source model developer may each contribute to a harmful outcome. These are not merely compliance hurdles; they are challenges to the conceptual architecture of public law itself.

The EU AI Act adds an important institutional dimension to this landscape. By mandating that member states establish AI regulatory sandboxes, the Act embeds experimentalist governance into binding European law, and it explicitly contemplates testing before market entry or operational deployment. This is significant because it signals legislative recognition that supervised experimentation can serve both innovation policy and regulatory learning simultaneously. Yet the Act’s sandbox provisions, like the national sandboxes in the United Kingdom, Norway, and Finland that preceded them, were largely conceived with private developers in mind: companies seeking to bring AI products to European markets under conditions of legal uncertainty. The extension of sandbox logic to public sector deployment, where the state itself is the deployer, stretches a framework built around market entry toward institutional contexts it was never designed to govern.

One way to see the mismatch clearly is to compare the seven parameters along which the article distinguishes public sector sandboxes from their private sector predecessors. The primary purpose of a conventional sandbox is innovation promotion balanced against consumer protection; a public sector sandbox must instead balance administrative improvement against legality, fundamental rights, and democratic legitimacy. The legal basis differs because public authorities cannot simply be granted waivers from the constitutional and statutory obligations that bind them; a waiver that suspends due process in the name of experimentation would itself be unlawful in most legal systems. The risk model differs because the relevant harms are not market harms but rights harms, including wrongful denial of benefits, discriminatory enforcement, and erosion of procedural fairness. Each of these parameters requires deliberate redesign rather than straightforward transplantation.

Accountability structures illustrate the redesign problem concretely. In a private sector sandbox, remedies typically include compensation for affected consumers, withdrawal of the product, and enforcement action against the firm. In a public sector sandbox, the affected population may be an entire category of benefit recipients, and the remedy may require not merely withdrawing a tool but unwinding thousands of individual decisions made with its assistance. The Dutch childcare benefits scandal, frequently cited in the literature on algorithmic government, demonstrated how algorithmic decision-making at scale can generate mass injustice that ordinary complaint mechanisms were never equipped to remediate. A public sector sandbox must therefore build remediation capacity into the experimental design itself, including mechanisms for identifying affected individuals, reversing erroneous decisions, and providing redress, rather than treating remedies as an afterthought to be addressed at exit.

Transparency and participation raise parallel difficulties. Conventional sandboxes involve confidentiality arrangements that protect the commercial interests of participating firms, on the theory that firms will not disclose proprietary innovations to a regulator without assurance that trade secrets will be safeguarded. Public sector experimentation cannot rest on the same premise, because the public has a democratic interest in knowing how its government makes decisions about it. Meaningful participation requires more than publication of a final evaluation report; it requires engaging affected communities, civil society organizations, and independent experts at the design stage, when the objectives and risk tolerances of the experiment are being set. Without such participation, a public sector sandbox risks becoming a mechanism by which the state legitimizes decisions it has already made, rather than a genuine forum for democratic deliberation about the proper role of AI in governance.

Data governance is a further parameter where public sector sandboxes demand distinct treatment. Private sector sandboxes typically involve firms processing consumer data under relaxed regulatory supervision, with data protection law operating as one of the rule sets that may be flexibly interpreted. Public sector AI systems, by contrast, often depend on administrative datasets, such as tax records, social security files, and immigration data, that were collected for purposes unrelated to the proposed AI application. Linking such datasets for experimental purposes raises questions of purpose limitation, data minimization, and the legality of secondary use that go well beyond consumer privacy. A well-designed public sector sandbox must specify what data may be used, under what legal authority, with what safeguards against re-identification and function creep, and with what arrangements for deleting or archiving data once the experiment concludes.

Evaluation design and transfer pathways complete the picture. The point of a sandbox is not experimentation for its own sake but the generation of transferable knowledge: either the innovation is institutionalized into ordinary administration, or it is abandoned, and in either case the regulatory framework should be updated in light of what was learned. For private sector sandboxes, the transfer pathway is market entry followed by standard regulatory oversight. For public sector AI, the transfer pathway is institutionalization into administrative practice, which raises questions about whether the safeguards that applied during the experiment, such as human review of algorithmic outputs, enhanced documentation, and periodic audits, will persist after the sandbox closes. Experience with government algorithm deployments suggests that safeguards often erode after pilots end, as budget pressures and operational demands mount. A credible public sector sandbox framework must therefore specify binding conditions for institutionalization, not merely criteria for entry into testing.

Taken together, these considerations support the article’s central claim that adaptation, not adoption, is the appropriate posture toward sandboxes for public sector AI. The sandbox remains an attractive instrument because it preserves what experimentalist governance does best: structured, time-bound, evidence-generating experimentation under supervision, with feedback into the regulatory framework. But the normative foundations must be rebuilt around public law values rather than market values. This entails anchoring public sector sandboxes in explicit statutory authority, defining rights-protective risk thresholds that cannot be traded away for efficiency gains, establishing independent oversight that is institutionally separate from the experimenting agency, guaranteeing transparency and participation rights for affected populations, imposing strict data governance conditions, and designing evaluation and transfer mechanisms that carry safeguards forward into institutionalized deployment. Where these conditions are met, the sandbox can serve as a legitimate bridge between the pace of AI innovation and the stability that legality and democratic accountability require. Where they are not, the same instrument risks becoming a vehicle for normalizing practices that would not survive ordinary administrative law scrutiny.

Subject of Research: Adapting regulatory sandboxes as experimentalist governance frameworks for public sector artificial intelligence experimentation

Article Title: Adapting regulatory sandboxes as experimentalist governance frameworks for public sector artificial intelligence experimentation

Article References: Okonjo, J. (2026). Adapting regulatory sandboxes as experimentalist governance frameworks for public sector artificial intelligence experimentation. Global Public Policy and Governance, 6(2), 259-281. https://doi.org/10.1007/s43508-026-00147-x

Image Credits: AI Generated

DOI: 10.1007/s43508-026-00147-x

Keywords: artificial intelligence, regulatory sandboxes, public sector experimentation, experimentalist governance, publicness theory, administrative law, democratic accountability, rule of law, EU AI Act, algorithmic governance, transparency, innovation policy