Does artificial intelligence improve social equity in public services? A systematic literature review on opportunities and pitfalls

As governments increasingly rely on AI to improve the efficiency of public services, equity is a critical dimension (McDonald et al., 2022; Young et al., 2025). Despite extensive work on efficiency and effectiveness in digital government (Cordella and Paletti, 2018), social equity remains underexamined even as AI reallocates discretion and reshapes distributional outcomes (Frederickson, 2015; Valle-Cruz et al., 2024), with only a few exceptions (Lahat and Nathansohn, 2025). Social equity is a constitutive value guiding who receives public services, on what terms and with what dignity. In practice, equity conditions trust, legitimacy and the fair distribution of burdens and benefits across populations: precisely the domains AI is poised to reshape (Acemoglu and Johnson, 2023; Wong et al., 2024). AI is particularly salient for equity because it relocates discretion: from caseworkers to data, from procedures to models and from local knowledge to generalized predictions (Covilla, 2025; Kulal et al., 2024).

Most studies approach the AI–equity question from narrow, discipline-specific angles, such as application domains (Bonomi Savignon et al., 2024; Jones and McKelvey, 2024; Selten and Meijer, 2021; Jin and Ryu, 2025), specific geographical settings (Margetts et al., 2024; Baykurt, 2022; Plantinga, 2024), or technical prerequisites (Ratner and Thylstrup, 2024; Varona et al., 2021). In this paper, we argue that organizing evidence according to different but interconnected potential government roles clarifies distinct levers of distributive impact and reveals interdependencies often hidden in domain-specific analyses.

This systematic literature review spans across disciplines to develop a framework for social equity challenges and solutions in AI-mediated public services, considering the different roles of governments in AI adoption (Guenduez and Mettler, 2023). Building on Guenduez and Mettler's analysis of national AI strategies, governments tend to adopt four roles. As regulators, they establish rules and remedies to ensure AI aligns with rights and accountability. As users, they implement systems in service delivery and determine how human judgment, automation and avenues for appeal should work together. As enablers, they invest in skills, standards and shared infrastructure, allowing public and private actors to develop and adopt AI. As leaders, they set directions and foster collaborations both domestically and internationally. These roles serve as narrative frameworks influencing problem and solution definitions, with their importance varying over time and across countries (Guenduez and Mettler, 2023).

We opt for a multidisciplinary approach to understand how governments can shape more equitable public services through AI. Indeed, including literature from different fields in the review improves evidence on how equity should be created by considering interdependent government roles, not only through regulation or front-line use.

Therefore, our research question is: “Considering the different roles governments play in developing, implementing, and governing AI technologies, what are the main challenges and solutions for governments delivering AI-driven equitable public services?” The answer aims to uncover latent tensions and propose solutions for both policy and practice.

The article is structured as follows: Section 2 defines social equity and justifies the four-role framework; Section 3 details the method; Section 4 reports descriptive and qualitative findings; Section 5 discusses implications and a research agenda; Section 6 concludes with recommendations and limitations.

Social equity has been central to public administration since the 1968 Minnowbrook Conference (Fredrickson, 1971), though its concept continues to evolve (Nabatchi and Carboni, 2019). Modern public management research evolves from focusing on efficiency to prioritizing social equity (Cepiku and Mastrodascio, 2021; McDonald et al., 2022; Wiley et al., 2025). However, social equity has different and sometimes conflicting meanings. Definitions vary among references to inputs, processes, outputs and outcomes, reflecting the heterogeneous conceptualization given by public management scholars (Riccucci and Van Ryzin, 2017). A widespread and shared definition is provided by the U.S. National Academy of Public Administration (NAPA) in which social equity is considered as “the fair, just and equitable management of all institutions serving the public directly or by contract; and the fair and equitable distribution of public services, and implementation of public policy; and the commitment to promote fairness, justice, and social equity in the formation of public policy” (NAPA, 2005: p. 11).

The relationship between AI and social equity presents an interdisciplinary challenge that modern scholars of public administration and public management should confront (McDonald et al., 2022). Investigating the social equity implications of AI facilitates a deeper understanding of how disruptive technologies shape public sector values in contemporary contexts (Peters and Torfing, 2025). Understanding AI’s dual role in improving public services and potentially increasing inequalities is vital to ensure AI advances public sector social equity goals (Wiley et al., 2025).

Indeed, AI can be seen as a “double-edged sword,” with the potential to either mitigate or reinforce existing biases (Ravanera and Kaplan, 2021). On the one hand, AI can create positive outcomes for marginalized groups in situations where human decisions may be clouded by cognitive biases, by removing the negative impact of human error in decision-making (Kleinberg et al., 2018). AI in practice, however, does not necessarily act as a counterweight to discrimination, as technology often reflects and reproduces the social dynamics of the context in which it is adopted and used, and AI is not an exception. Societal power relations and inequality may shape the data, the algorithms and how these algorithms are used; therefore, AI support to humans will be biased as a consequence (Garcia, 2016).

In public services, model design, data provenance and human–AI interaction jointly determine accountability and legitimacy (Mittelstadt et al., 2016; Busuioc, 2021; Grimmelikhuijsen and Meijer, 2022). Potential benefits of algorithmic decisions are weakened by the tendency of both decision-makers and street-level bureaucrats to select AI recommendations that align with their stereotypes and experience (Alon-Barkat and Busuioc, 2023; Selten and Meijer, 2021).

Regarding AI, we adopt the broad and updated OEDC (2024) definition of “a machine-based system that generates outputs […] that can influence both physical and virtual environments by inferring from received inputs to meet explicit or implicit objectives”. We operationalize this definition by distinguishing four loci of potential equity claims: data collection and representativeness (inputs), modeling choices and explainability (processes), allocation and eligibility decisions (outputs) and rights protections (outcomes). The regulator–user–enabler–leader framework maps these four equity loci to concrete levers: rules and redress (regulator), operational practice and discretion (user), infrastructures, skills and standards (enabler) and strategic direction and countervailing power (leader) (Guenduez and Mettler, 2023). More in detail, as regulators, governments define constraints and remedies (e.g., audit mandates, due process and non-discrimination) that influence upstream design and downstream compensation (Dwivedi et al., 2021). As users, they implement AI in high-stakes decisions, where hybrid human-machine arrangements and participatory assessments determine actualized equity (Schiff et al., 2022; Taylor et al., 2024). As enablers, they invest in infrastructures, skills and partnerships that govern who can develop and benefit from AI innovations (Ulnicane et al., 2021). As leaders, they establish strategic directions and balanced powers (Margetts, 2022), aligning AI development with democratic values and global justice. Each role addresses a different equity challenge and, importantly, these roles influence one another in practice. This proposed framework complements equity loci (inputs, processes, outputs, outcomes) by mapping them to concrete governmental levers, enabling cumulative comparison across domains.

A systematic literature review was chosen to consolidate past and ongoing research findings (Briner and Denyer, 2012; Rousseau, 2012) across multiple disciplines, recognizing that AI’s impact on social equity cannot be fully understood from a single perspective.

We searched Web of Science, Scopus and EBSCO using AI, equity and public-sector terms, without imposing a start date (PRISMA in Figure 1). The earliest included article is from 2016; searches were updated to January 2025. The Figure shows our literature search strategy, which combines three key elements: AI-related terms (including “artificial intelligence,” “machine learning,” and “algorithms”), equity-focused language (covering fairness, inclusion and related concepts) and public sector terminology (spanning government, public administration and services). By linking these components with AND operators, we ensured our search captured only studies examining AI’s equity implications specifically within government contexts.

We associated equity implications with AI, different types of technologies (GenAI, LLM, Algorithms for decision-making, Facial recognition systems), when disclosed and specifically coded cases with an unspecified type.

After applying the search query to the three databases, an initial set of results was obtained, then a filter was applied to include only English-language publications and the outputs were limited to academic articles. We did not restrict the search to specific subject categories, as the review aims to combine different disciplinary perspectives.

This process yielded the following results: Web of Science (817), Scopus (1,033) and EBSCO (743). The next step involved removing duplicates across three databases, then two authors independently reviewed abstracts. Inclusion criteria focused on relevance to the public sector, considering government roles in AI and social equity.

We used a consensus process to resolve disagreements about study inclusion: when the two reviewers differed, they discussed their reasons and a third reviewer made the final decision if needed.

Studies were eligible if they directly examined the intersection of AI, social equity and the public sector; reported domain-specific implementations (e.g., healthcare, education, social services) with explicit equity analysis; proposed guidelines, standards, or best practices for equitable government AI; or evaluated strategies and policies that advance AI while mitigating unfair or discriminatory impacts.

We included n = 128 studies and conducted structured qualitative coding to extract role, modality, challenges and solutions. Each article was then assigned a government role, based on Guenduez and Mettler (2023).

The full research protocol is illustrated in the diagram in Figure 1.

We first outline sample characteristics, then synthesize qualitative insights by role.

Interest in AI and equity topics has surged over the past decade, highlighting their growing importance in academia and policy. The 2016 paper is a foundational work that discusses ethical challenges, including opacity, bias and accountability. It calls for interdisciplinary approaches to develop governance frameworks that ensure transparency and fairness (Mittelstadt et al., 2016).

In our final sample (n = 128), conceptual articles predominate, followed by a marked growth in empirical studies after 2021; systematic reviews emerged subsequently. Regarding technologies, decision-support algorithms are most common; 56 articles refer to “AI” without specifying a technology, while robotic process automation and facial recognition appear in only three articles each. Overall, the sample mainly treats AI as an overarching category or discusses “algorithms” without specifying subtypes. This pattern is most evident in empirical legal and social science studies, where institutional design, accountability and rights are the primary focus and the technical foundation is regarded as secondary background (Figure 2).

The current academic debate primarily focuses on government regulation of AI, with limited attention to government leadership or enabling roles, which have recently gained some interest. Our analysis (Figure 3) reveals the regulatory role in forty-nine papers, particularly from 2021 onward, with 19 papers published in 2024. The user role is featured in eleven papers, highlighting the practical use of AI in public services. The enabler role is featured in eleven papers, primarily in 2022, while the leader role appears in ten papers, with a peak in 2024. The distribution of publications along the four roles confirms the topic's interdisciplinary nature, with specific concentrations in particular subject areas. Government Information Quarterly is the most represented journal, with eleven articles. AI & Society contributed eight publications, offering a unique perspective from the humanities and computer science. Big Data & Society and Public Administration Review each published five articles, examining the topic from data science and public management angles, respectively. Figure 3 shows the distribution based on roles and research categories (source: Scimago Journal & Country Rank (SJR)). The Social Science domain shows diverse engagement, especially in the Regulatory category, with a strong presence across all roles. The discipline primarily examines how AI influences societal structures and institutions, with a focus on regulation. Business, management and accounting, often combined with other disciplines, have eight publications in the User category, emphasizing practical use. Computer science, alone or in collaboration with others, is represented in user and regulatory roles, emphasizing both technical and governance focus. Engineering and medicine make selective but significant contributions, mainly in enabling roles.

This disciplinary pattern indicates a compartmentalization of studies that may obscure equity externalities at the ecosystem level: legal and governance scholars analyze rules and tools, while organizational scholars observe implementation; very few studies integrate market-shaping policies, capacity building and geopolitical asymmetries into a comprehensive equity analysis.

The analysis of how AI and social equity interact in the context of each of the four roles of government inspires the definition of a comprehensive taxonomy of challenges and potential solutions that governments must navigate in the era of AI. Table 1 summarizes challenges and instruments by role, highlighting distinct decision logics and metrics. We elaborate on each quadrant below.

However, the literature demonstrates how roles are often deeply interconnected and overlap. Their balance may also shift over time as AI technologies and societal needs evolve. For this reason, in the discussion, we dedicate a specific paragraph to highlight the interconnection between some challenges and solutions and the potential synergy among the government's roles. The four-role taxonomy analytic labels clarify their conceptual distinctions and practical implications: (1) Ethics and normative oversight (Regulator): anticipatory, context-specific rules, audits and redress; (2) Equitable Integration (User): hybrid human-in-the-loop arrangements, participatory impact assessments and algorithmic literacy; (3) Inclusive Enablement (Enabler): market-shaping investments, skills and infrastructure for underserved communities and concrete standards translating ethics into enforceable requirements; (4) Participatory AI Leadership (Leader): long-range public value orientation, countervailing power to platform dominance and global justice commitments.

By acting as regulators, governments should manage the potential conflict between AI development and public values (Selten and Meijer, 2021; Nzobonimpa and Savard, 2023; Margetts, 2022; Robles and Mallinson, 2023). While AI can foster efficiency in public services, it can exacerbate issues of transparency and inclusivity of marginalized communities, affecting government accountability (Busuioc, 2021). Early engagement of disadvantaged groups in AI adoption and governance, along with specific regulations targeting bias and discrimination, could safeguard social equity in system and service design (Gaozhao et al., 2023). Overall, proactive measures to ensure that algorithmic systems are designed and implemented in ways that promote equitable outcomes–including mechanisms for civic participation, legal frameworks and closer monitoring of algorithms–are essential regulatory actions to ensure legitimacy and protect marginalized communities (Grimmelikhuijsen and Meijer, 2022; Kaur et al., 2023). Adaptive instruments, such as regulatory sandboxes with equity guardrails, tiered audit requirements proportional to decision stakes and due-process rights for algorithmic decisions, are potential solutions to balance fast-evolving technologies and asymmetric impacts.

AI adoption risks eroding human discretion and judgment, raising concerns about reliance on technology (Levy et al., 2021; Young et al., 2019; König and Wenzelburger, 2021). A hybrid human-machine model is preferred, enabling caseworkers to handle complex cases and appeals (Ranerup and Henriksen, 2022; Waldman and Martin, 2022). Insufficient citizen involvement in AI-based service co-design may harm human autonomy (Park and Humphry, 2019). Governments should engage communities, conduct participatory system design and assessments for social equity, create oversight bodies with community input and explore alternative visions that enhance citizen participation (Moon, 2023). To mitigate risks, governments must inform users they are interacting with an AI, continually evaluate AI systems, avoid gender role assumptions in AI design and apply the “Principled AI” framework in development (Varona et al., 2021). Promoting algorithmic literacy and involving communities in AI design and oversight are essential for trustworthy AI services (Wang et al., 2024). Equitable Integration can be strengthened by concrete practices, such as remit-specific “equity playbooks” that translate abstract principles into decision frameworks and appeals procedures for services (e.g., benefits eligibility, permitting) and ongoing, mixed-methods evaluation that triangulates model metrics with lived experience data from affected communities.

When enabling AI adoption across the public service ecosystem, pressing challenges mostly relate to digital divides in access and skills (Bonomi Savignon et al., 2024). Forward-thinking governments are investing in their workforce’s digital skills while creating oversight systems that promote innovation without compromising accountability (Gibbons, 2021). Indeed, trust and the opacity of AI systems present another challenge, particularly when these systems make decisions that could impact citizens’ lives. The lack of transparency makes it challenging for administrators to explain decisions and maintain public confidence while still advocating for technological progress (Margetts, 2022). However, rather than merely establishing rules, some governments are actively shaping AI integration, turning abstract ethical guidelines into practical standards that balance innovation with accountability (Fernandez-Aller et al., 2021). Market-shaping interventions, such as public data trusts with equity clauses, procurement criteria that reward fairness benchmarks and subsidized civic-tech incubators focused on underserved populations, could be implemented.

Steering AI calls for a new kind of public digital leadership. As Big Tech firms draw on public data, governments risk dependency for essential services and infrastructure (Valle-Cruz et al., 2024). States should decouple digitalization from marketisation and govern core digital infrastructures as public utilities with clear distributional stakes. Core welfare sectors, health, education and social protection, require statutory safeguards against overreach and excessive reliance on the private sector that can undermine public values (Sharon and Gellert, 2024). Data should be protected from extractive uses and benefits should also be distributed in the Global South (Taylor, 2024). National AI strategies should be presented as public-value charters, featuring transparent capability roadmaps, sovereign infrastructures for essential services and solidarity mechanisms such as compute credits and equitable licensing to reduce data extraction and distribute gains more fairly.

Our synthesis shows that equity challenges and remedies cluster by role but operate as a system. Regulatory and user roles are disproportionately represented because they fit existing institutional arrangements, data availability and near-term incentives; enablement and leadership require longer horizons, cross-sector coordination and political mandates, and thus remain underdeveloped. This imbalance risks portraying equity as remediation at the point of use rather than as a matter of design and proactive governance upstream.

Modality non-disclosure weakens external validity: equity claims built on one family (e.g., risk scoring) are tacitly generalized to others (e.g., biometrics, generative models), encouraging one-size-fits-all remedies. Two forces likely drive this ambiguity. First, empirical legal and social science work often scrutinizes institutions – rules, discretion, legitimacy – rather than artifacts, pulling analysis away from model classes and toward governance arrangements. Second, disclosure constraints and fast-changing toolchains complicate precise labeling in practice: commercial confidentiality to mitigate competition, the fast evolution of AI development tools, with constantly updated libraries and emerging frameworks integrating different components and versions, create challenges to maintain accurate, up-to-date system labels and proper definitions. The costs are nontrivial: weaker external validity, limited replication and difficulty aligning safeguards to concrete harms.

These limits do not undermine role-based synthesis; rather, they clarify where each role can close the modality gap. Ethics and normative oversight should be anticipatory and proportionate (regulator); equitable integration should institutionalize participation and appeals (user); inclusive enablement should rebalance who innovates and who benefits (enabler); and participatory leadership should pursue strategic autonomy and global justice (leader). Figure 4 presents a taxonomy that links role-specific challenges to instruments, emphasizing interdependencies across roles.

Cross-cutting principles remain pivotal. Transparency, accountability and legitimacy are necessary conditions but not sufficient on their own (Mittelstadt et al., 2016; Busuioc, 2021; Grimmelikhuijsen and Meijer, 2022). Meaningful human oversight requires capability, authority and information to override model outputs (user); enablement requires inclusive access and enforceable standards (enabler); leadership requires culturally sensitive strategies that counter platform dominance and dependency (leader) (Ranerup and Henriksen, 2022; Sharon and Gellert, 2024). Regulatory frameworks engage with questions of distributive justice, so they should incorporate adjustments to mitigate the historical distribution of inequities (Minow, 2023; Mittelstadt et al., 2016). Governments should therefore approach “adaptive oversight”. Namely, mechanisms that can grow alongside the technologies they govern, adopting regulatory sandboxes and including potential underrepresented communities (Margetts, 2022; Jobin et al., 2019).

As AI users, governments can achieve the most immediate and tangible benefits for society. Our findings suggest that the most successful AI implementations preserve, rather than diminish, human involvement in judgment (Alon-Barkat and Busuioc, 2023; Young et al., 2019). Hybrid models that enhance human capabilities while maintaining human accountability are adequate to support government functions (Ranerup and Henriksen, 2022; König and Wenzelburger, 2021). However, meaningful human oversight necessitates that human operators possess the capability, authority and relevant information to override AI recommendations when necessary (Levy et al., 2021; Waldman and Martin, 2022).

As enablers, Governments are challenged to develop more sophisticated approaches to digital inclusion, building AI literacy, supporting public investment in AI applications serving underserved communities and fostering public-private partnerships with explicit equity requirements (Bonomi Savignon et al., 2024; Arnaout et al., 2023; Gibbons, 2021).

Although still understudied, the government as an AI leader is the most ambitious role among others. It requires more than strategic planning to articulate visions for AI’s role in society while building the institutional capacity to realize those visions (Valle-Cruz et al., 2024; Hjaltalin and Sigurdarson, 2024). This is particularly challenging, especially in small national contexts where public resources are scarce, given the dominance of large technology companies (Sharon and Gellert, 2024) and necessitates building domestic capacity to develop AI applications that serve national priorities and values (Taylor, 2024; Plantinga, 2024).

Regulatory and user roles dominate due to tractability (existing institutions and data), salience (visible harms) and incentives (short-term measurability). Enabler and leader roles require broader perspectives and stronger cross-sector coordination, which hampers both implementation and academic research documentation. However, without Inclusive Enablement and participatory AI leadership, equity potentially becomes a remedial action after regulation or a local optimization during use. The upstream distribution of capabilities and downstream distribution of power remain under-theorized, weakening the transformative potential of many technical reforms.

We thus embed design-oriented implications in each quadrant. Adaptive oversight should be anticipatory and proportionate; equitable integration should institutionalize participation and appeals; inclusive enablement should re-balance who innovates and who benefits; participatory AI leadership should pursue strategic autonomy and global justice.

When we examine how governments engage with AI across their various roles, it emerges that those roles influence each other and fundamentally reshape how public institutions can pursue social equity. Overall, governments create rules and then must adhere to them when deploying AI systems (Ranerup and Henriksen, 2022).

Governments attempting to encourage AI adoption across society (the enabler role) face a trustworthiness challenge. Private actors are more likely to follow government enablement when the public sector demonstrates competent deployment and credible oversight. Scandals involving biased algorithms in their public systems, such as the health system (Fountain, 2022), can diminish the effectiveness of government enablement efforts for stakeholders in the ecosystems. Leadership in AI governance cannot simply be declared but is earned through competence in the other roles (Hjaltalin and Sigurdarson, 2024).

To support governments navigating these interconnections, public management scholars should investigate potential managerial solutions to provide institutional arrangements where regulatory, operational and strategic dimensions are oriented towards specific needs (Criado et al., 2025). The measurement challenge extends beyond traditional performance frameworks that treat regulatory compliance and operational effectiveness as distinct categories. What emerges from our review is the need for integrated assessment approaches that capture the dynamic relationships between different governmental functions (Grimmelikhuijsen and Meijer, 2022; Waldman and Martin, 2022). Governments have started to experiment with “relational indicators”, intended as metrics designed to track not isolated outcomes but the interactions between regulatory frameworks and operational realities. These approaches recognize that a regulation’s actual impact only becomes visible through its implementation, while operational experiences continually reshape regulatory understanding (Wang et al., 2024). Thus, AI governance must be led by public managers who can engage with technical teams, have sufficient regulatory knowledge to ensure compliance and strategic thinking to connect AI with broader public purposes (Madan and Ashok, 2023; Bernhard and Wihlborg, 2022). A solution could be the piloting of relational indicators that connect, for example, the enforcement of audit mandates (regulator) to appeal volumes and resolutions (user), to supplier diversity and fairness benchmarks in procurement (enabler) and to platform dependency measures in critical services (leader).

Managing the tension between innovation and precaution as a productive friction yields more robust and equitable outcomes (Margetts, 2022; Valle-Cruz et al., 2024; Busuioc, 2021). Considering the discussion, we propose a research agenda, acknowledging a fundamental challenge: the object of study evolves more rapidly than our analytical capacity. Yet this temporal mismatch makes rigorous investigation even more relevant.

Our review reveals that governments face multifaceted challenges across several domains. The solutions to these challenges offer fertile ground for scholarly inquiry. Table 2 maps potential research directions across the four quadrants outlined in Figure 4, providing a framework for future investigation.

We argue that equity cannot be bolted on at the point of use: it must be designed into infrastructures (enabler) and safeguarded through strategic capacity (leader), with regulation and use completing the system (Mittelstadt et al., 2016; Williams et al., 2022; Busuioc, 2021).

A recurring constraint is definitional blur: “AI” is too often a catch-all, yet different systems entail different risks and safeguards. Future work should name the technology and task (e.g., risk scoring, FRT, LLMs) and meet minimum reporting standards on data provenance, deployment context, human-in-the-loop and appeal routes. Therefore, we provide the following suggestions for future research.

First, to reframe equity as a system of interconnected roles: the taxonomy links different levers, rules (regulator), practices (user), infrastructures and skills (enabler) and strategy and countervailing power (leader). Devote more scholarly attention to the ecosystem dimension of AI through the leadership and enablement functions of governments. Secondly, in recognition of the multidisciplinary nature of AI, to specify the AI technology under analysis in every field of study: equity claims should be limited to the specific AI technology involved (e.g., risk scoring, facial recognition, generative systems), with implications for due process, audit design, explanation and redress. We suggest minimum reporting standards for cumulative science and cross-disciplinary analysis (such as the modality and task; describe data origin; outline deployment context and human-in-the-loop arrangements, etc.).

Actionable implications for practice include:

For Regulators: mandate risk audits tied to decision stakes; codify process rights (notice, reasons, appeal) for decisions supported by algorithm; require participatory impact assessments for high-risk deployments; use adaptive oversight (e.g., sandboxes) that includes underrepresented communities (Waldman and Martin, 2022; Grimmelikhuijsen and Meijer, 2022; Mittelstadt et al., 2016; Jobin et al., 2019; Margetts, 2022; Kaun, 2022; Busuioc, 2021; Taylor et al., 2024; Horvath et al., 2023; James et al., 2023).

Service managers (users): adopt remit-specific equity playbooks; install human-in-the-loop checkpoints with clear authority to override; combine model metrics with lived-experience monitoring; ensure accessible appeal routes and feedback loops into model maintenance (Alon-Barkat and Busuioc, 2023; Young et al., 2019; Ranerup and Henriksen, 2022; Levy et al., 2021; König and Wenzelburger, 2021; Waldman and Martin, 2022; Wang et al., 2024; Varona et al., 2021).

Enablers: fund shared data/compute with enforceable standards for access, quality and representativeness; invest in AI literacy for officials and affected communities; include equity clauses in public-private partnerships and procurement; support applications targeted at underserved groups (Williams et al., 2022; Fernandez-Aller et al., 2021; Ulnicane et al., 2021; Gibbons, 2021; Bonomi Savignon et al., 2024; Arnaout et al., 2023; Pah et al., 2022; Saldanha et al., 2022).

Leaders: reduce strategic dependence on dominant platforms for critical services; align national roadmaps with public-value charters; back cross-border cooperation that advances interoperability, rights and equitable access; support Global South capacity through compute credits, open tooling and fair licensing (Valle-Cruz et al., 2024; Sharon and Gellert, 2024; Hjaltalin and Sigurdarson, 2024; Plantinga, 2024; Taylor, 2024; Ulnicane et al., 2021).

Equity in AI is inseparable from democratic legitimacy. Meaningful oversight requires capability, authority and information for humans to contest and reverse automated recommendations; inclusion requires that communities help set problem definitions, not only react to harms; and fair distribution requires rebalancing who gets to innovate and who benefits. In a geopolitically concentrated AI economy, leadership also entails building domestic capacity and acting collectively to counter lock-in to proprietary infrastructures that may constrain public values.

To conclude, we acknowledge that our research has limitations. First, categorizing challenges and solutions within a government role framework helps identification and classification, but may miss some interconnections between the four quadrants. While the framework synthesizes findings, it simplifies a complex reality and requires further validation through empirical research. Future studies could include case studies, content analysis of documents and practitioner engagement. Longitudinal research could examine AI’s long-term social equity effects and solution effectiveness.