Abstract
The use of artificial intelligence (AI) applications in government is receiving increasing attention from global research and practice communities. This article, introducing a Special Issue on Artificial Intelligence in Government published in the Social Science Computer Review, presents an overview of some of the main policy initiatives across the world in relation to AI in government and discusses the state of the art of existing research. Based on an analysis of current trends in research and practice, we highlight four areas to be the focus of future research on AI in government: governance of AI, trustworthy AI, impact assessment methodologies, and data governance.
Despite the recent boom of references to artificial intelligence (AI) in the media, in the corporate sector, and in public discourse, the notion of AI is not new. The term was coined in the 1950s in an academic context to indicate an emerging research field studying (1) the ability of machines to carry out tasks by displaying intelligent, humanlike behavior and (2) the ability of machines to behave as intelligent agents by perceiving the environment and taking actions to achieve some goals (Russell & Norvig, 2016; Tzafestas, 2016).
The often-highlighted value creation potential of AI applications includes the ability to augment labor and increase productivity, to more effectively allocate resources, and to foster innovation. AI instances include techniques such as supervised and unsupervised machine learning (Smola & Vishwanathan, 2008), artificial neural networks (Priddy & Keller, 2005), case-based reasoning (Kolodner, 1992), natural language processing (Chowdhury, 2003), multiagent systems (Wooldridge, 2009), and machine reasoning (Bottou, 2014). AI applications also include cyber-physical systems (Radanliev et al., 2020), such as robotics and Internet-of-Things (IoT) applications, image and facial recognition, speech recognition, virtual assistants, and autonomous machines and vehicles.
AI represents an ideal technology for use in the public-sector context, where environmental settings are constantly changing and preprogramming cannot account for all possible cases (Dwivedi et al., 2019; Sun & Medaglia, 2019). AI differs from traditional automation technologies because it does not make decisions on preprogrammed if-then logic, in which the same input instructions produce the exact same results. Instead, AI features autonomy in function and learning, an assumption of appropriate and available data, and the incorporation of physical and virtual spaces (Ahn & Chen, 2020). However, if the assumptions are not met, AI might not be the best option for automation and decision making.
Opportunities coming from the innovative disruptive power of AI in the public sector are primarily found in three areas: (1) improving the internal efficiency of public administration, (2) improving public administration decision making, and (3) improving citizen–government interaction, including the provision of better and more inclusive services and the enhancement of citizen participation in the activities of the public sector (Samoili et al., 2020). As a potentially disruptive sociotechnical phenomenon, AI is relevant to the full range of government’s roles: as a regulator and as a catalyst for research and development (governance of AI) and as a user (governance with AI or AI in government). Such potential could be realized if governments foster an environment characterized by a skilled workforce, an appropriate regulatory framework, resources that can be promptly mobilized, and incentives to innovate. Risks of AI, on the other hand, include, for example, widening societal divides, infringing citizens’ privacy rights, and clouding the accountability of public decision makers. Such risks require thoughtful strategies and regulation in order to be avoided or mitigated.
As AI-related policy initiatives and research on AI in government accumulate, this article aims to map key trends in policy and research and identify directions for future research. This article accompanies a Special Issue on Artificial Intelligence in Government edited by the authors for the Social Science Computer Review, building on a call for cutting-edge studies on this topic area from different social science perspectives, including public administration, information systems, sociology, information science, and management.
The rest of this article is structured as follows. In AI in Government: Policy and Research section, we provide an overview of the main policy initiatives concerning AI in government in Europe, the United States, and China and map the state of the art of research on AI in government drawing on a review of existing research. In Contributing Perspectives section, we present the six studies included in the special issue and map them against the state of the art of existing research. Finally, in the concluding section, we summarize our points and propose a set of key questions for future research on AI in government.
AI in Government: Policy and Research
Policies Across the World
Governments across the world are providing increasing attention to the potential of AI. This potential is seen not only in relation to economic growth, since national AI strategies also focus on common themes such as trust and ethics, security, and enhancing the talent pipeline (Berryhill et al., 2019). Increasing attention has also started to be devoted to the role of AI in achieving the sustainable development goals of the United Nations’ Agenda 2030 (Medaglia et al., 2021; Vinuesa et al., 2020). It is not surprising that different countries and different areas of the world are starting to develop approaches to AI in government that have diverse features. These features, some argue, are indicators of different value drivers (Viscusi, Rusu, & Florin, 2020) and governance orientations (Viscusi, Collins, & Florin, 2020).
For instance, Europe has been expressing the will to position itself in the global landscape with AI strategies that rely more heavily on enterprise and government data. This positioning has been attributed to its relative lag, within global competition, in competitiveness in the consumer data space (Groth & Straube, 2021). The European Commission reviewed its “Coordinated Plan on the Development and Use of Artificial Intelligence Made in Europe” (European Commission, 2021) to foster cooperation between all European Union Member States, plus Norway and Switzerland. The plan focuses on four key areas: increasing investment, making more data available, fostering talent, and ensuring trust and highlights the public sector as a “trailblazer for using AI” (European Commission, 2021, p. 46).
The European Commission devotes a specific section in a recently published white paper on AI to promote the adoption of AI by the public sector, considering it “essential that public administrations, hospitals, utility and transport services, financial supervisors, and other areas of public interest rapidly begin to deploy products and services that rely on AI in their activities” (European Commission, 2020, p. 8). The framework provided by the European Commission provides for a deployment of AI in the public sector that is aimed at developing human-centric AI. More specifically, health care, rural administrations, and the support of public procurement of and with AI are considered priority areas for consideration.
As of April 2021, 19 European countries have dedicated official strategies to AI. Despite the growing focus of these countries on the potential of AI in the public sector, the reality is that government is often seen merely as a regulator or as a facilitator of AI, that is mostly just providing guidance or legal and regulatory frameworks to minimize the potential risks of AI, while enabling the maximum opportunities from its application. Actual adoption of AI by governments to improve public services, policymaking, and internal operations do not gain the same amount of interest and related investment (Misuraca & van Noordt, 2020). A systematic analysis of these national strategy documents (van Noordt et al., 2020) reveals how the majority of policy initiatives put in place by European Union countries belong to the category of “sermons” (Bemelmans-Videc et al., 1998), that is, soft instruments, such as communication campaigns, private–public partnerships, and voluntary codes of conduct. Less frequent are “carrots” initiatives, that is, the allocation of economic resources and economic incentives, such as direct cash transfers, tax incentives, competitive research funding schemes, and so on or regulatory instruments (“sticks”), including binding laws and regulations, the establishment of intellectual property rights, competition regulation, or ethical regulations.
The European approach to AI is often mentioned in the context of the “global AI race” that also involves the United States and China (Craglia et al., 2018). Comparisons between these three entities involve levels of investments, research, training, and education. Successful deployment of AI is seen as key to efforts to dominate an arena characterized by strong network effects (Makridakis, 2017).
While the United States has by far the highest private-sector investment in AI (McKinsey, 2020), the take up of AI in the U.S. public sector is less clear cut, partly due to the fact that much less research and official information is available about the current and planned use of AI in public administration (Cath et al., 2018; Craglia et al., 2018).
The U.S. Federal Government has been putting forward initiatives to develop trustworthy AI for government services, aligned with constitutional values. Several federal agencies have already been using AI for various purposes, including processing grant applications, compliance checking, and predictive maintenance. A recent mapping of the use of AI technologies showed 142 uses of AI in relevant federal administrations (Engstrom et al., 2020). A noteworthy challenge facing the federal government, vis-à-vis the booming of AI in the business sector, is the lack of available talent in the public workforce. U.S. federal agencies are often forced to find ways to leverage expertise from the private sector to develop strategies to adopt AI, including new approaches to hire and train existing workers with new skills to innovate with AI technologies (The White House Office of Science and Technology Policy & Policy, 2020).
As in Europe, subnational levels of government in the United States seem to lag behind the federal government. A survey conducted among IT executives of 45 state agencies showed that only 1% have adopted AI across their state and that most of the state of AI adoption is merely in a proof-of-concept phase (Center for Digital Government et al., 2019). From a citizen perspective, U.S. agencies face a lack of confidence by citizenry in governmental organizations to successfully manage the development and use of AI technologies. Rush in development and opaque management are cited by citizens as potential causes of mistrust in government AI (Zhang & Dafoe, 2019).
The deployment of government AI initiatives in the People’s Republic of China is often referred to as the most relevant counterpart of those in Europe and the United States. The setting up of AI-powered surveillance systems, in general, and examples like the social credit system, in particular (Creemers, 2018), are frequently cited as dystopian initiatives curbing citizens’ rights and endangering global power balances. However, the prevalence of such fears, while not unfounded, often hinders a proper understanding of the complexity of the reality of government AI initiatives in China.
The “New Generation Artificial Intelligence Development Plan,” released by the Chinese State Council in 2017, was intended to be a unifying document for China’s various AI objectives. However, the document presents as more of a wish list than a central command directive, where the lower levels of government are expected to play a vital role in the actual transformation of society through AI (Roberts et al., 2020). While the national strategy focuses on technological enablement, pilots and experimentation are expected from lower levels of government, often following a principle of “experiment first and regulate later” (Elliott, 2020). Sometimes such initiatives go far beyond the national AI policy, sparkling relevant domestic controversies. Moreover, the actual delivery of AI-enabled public services appears to be mostly from large technology companies that are favored by the central government. Applications used in the Chinese public services are not necessarily cutting edge, but their strength lies in the integration of various systems and in the ability to scale up quickly (Elliott, 2020). China’s AI policies prioritize the speeding up of technology development, data collection, and implementing pilots, while risk management, data privacy, and accountability appear to be secondary to these imperatives (Elliott, 2020).
Despite popular conception, China has also been working on the adoption of AI ethical principles and frameworks (Ministry of Industry and Information Technology of the P.R.C., 2019). Such principles, while partly overlapping with the ones of Europe and the United States, do put a stronger emphasis on social responsibility (e.g., “harmony”) rather than individual rights. Notwithstanding the tensions and contradictions that characterize the Chinese government AI initiatives, citizens in China indicate relatively high support for government applications of AI (Carrasco et al., 2019). At the global level, there is emerging evidence that less developed economies and countries that have higher reported or perceived levels of corruption also tend to be more supportive of AI in government (Carrasco et al., 2019). This could be an indication of the neutrality that people expect from AI.
State-of-the-Art of Research
As a new field of policy intervention, the use of AI in government is attracting an increasing amount of attention from the research community, raising a range of new research questions. In order to map the body of knowledge on AI in government, we first analyze published literature reviews on AI in the public sector (A Brief Review of Literature Review Papers section), and then we provide a review of research papers published in the period 2020–2021.
A brief review of literature review papers
In order to have a solid understanding of the state of the art in terms of research on AI in government, this section summarizes the results of our review of existing literature review papers that explore AI in government. Using key words such as “artificial intelligence” and “public sector” or “government,” we looked for reviews in the last 5 years and found nine, all of them published between 2019 and 2020. For maximum coverage, we used Scopus and Google Scholar as our main databases and did not limit the search to certain fields or disciplines. Most studies were published in public administration, information systems, or digital government conferences and journals. The literature review papers were selected based on the topic but also looking for a broad focus on AI in government. Therefore, literature reviews focused on specific subtopics such as AI and smart cities, AI and health care, big data and AI, or AI and cross-sectoral collaboration were not included. Overall, we found that existing literature reviews on AI in government highlight different aspects, which could be classified in the following categories (see Table 1): (1) definition and attributes; (2) techniques and technologies; (3) uses and applications; (4) results, impacts, and benefits; (5) challenges and determinants; (6) strategies, best practices, and guidelines; and (7) ethical considerations. The next few paragraphs summarize the findings of these review efforts.
Categories of Topics Found in Existing Literature Review Papers.
Note. AI = artificial intelligence.
AI definition and attributes
Research efforts are far from producing a universally accepted definition of AI, partly due to the fact that any definition of the term has to include the very elusive notion of intelligence. The paradox of the so-called “AI effect” (McCorduck, 2004) is a case in point: As soon as technological breakthroughs realize some of the capabilities forecasted for AI—such as recognizing human speech or reading written script—many observers discount these capabilities as not “real” intelligence. As a result, the target of AI is constantly moving, since “intelligence is whatever machines haven’t done yet,” a claim referred to as “Tesler’s theorem” (Hofstadter, 1999). From the nine literature review papers found, four of them specifically highlight definitions of AI and its main attributes. For example, Wirtz et al. (2019) present a table with six different definitions of AI. Based on those definitions, they propose what they call an integrative definition, which is “AI refers to the capability of a computer system to show humanlike intelligent behavior characterized by certain core competencies, including perception, understanding, action, and learning” (Wirtz et al., 2019, p. 599). In contrast, Reis et al. (2019b) mostly adopt one definition of AI and devote several paragraphs to clarifying differences and similarities among several related terms, many of which are actually techniques that could be thought of as part of AI. Finally, Valle-Cruz et al. (2019) focus on the concepts related to AI in the public sector, including knowledge, learning, and intelligence. Overall, the literature review papers acknowledge that there is no single definition of AI in government, but there are a few attributes or characteristics that are present in many of the definitions. Russell and Norvig (2016) capture the essence of these definitions with their view that AI could be conceptualized as “systems with the ability to think and learn.”
AI techniques and technologies
Four of the nine literature review papers discuss different AI techniques and technologies as found in the literature. For example, Sousa et al. (2019) provide a list of AI techniques, including CBR, cognitive mapping, fuzzy logic, ML, ANNs, genetic algorithms, MAS, and NLP. However, they do not directly elaborate on any of them and the focus of their review article is on applications. Similarly, Reis et al. (2019b) also present a list of AI techniques and briefly describe each of them. The techniques included in their review paper are neural networks, deep learning, and ML. They also mention some important applications of these techniques such as computer vision, NLP, and speech recognition (Reis et al., 2019b). In contrast, as part of the technology infrastructure layer of the AI frameworks they reviewed, Wirtz and Müller (2019) identify many AI techniques, including ML, intelligent control, knowledge representation, pattern classification, data lake analytics, cognitive services, and neural networks. They argue that it is very important to consider the actual function needed in order to select a technique or combination of techniques that could help to sense, comprehend, or act (Wirtz & Müller, 2019).
AI uses and applications
From the nine papers identified, six include topics related to the different uses and applications of AI in government. For example, Ahn and Chen (2020) identify nine different uses of AI in government settings. These include smart allocation of public service resources, digital assistance with chatbots, pattern identification and predictive analytics models, automation and “Regu-Tech,” smarter public utilities, smart energy, the IoT and robotics sensors, autonomous driving, and sensor-based detection and prevention. Similarly, Wirtz and colleagues (2019) also highlight specific AI applications and how they could be used in the public sector. These include AI-based knowledge management software, AI process automation systems, virtual agents, predictive analytics and data visualization, identity analytics, cognitive robotics and autonomous systems, recommendation systems, intelligent digital assistants, speech analytics, and cognitive security analytics and threat intelligence. They also provide a list of public-sector use cases. In contrast, Sousa et al. (2019) propose a very different approach in which they refer to use as the main function of government, including general public service, public order and safety, defense, economic affairs, environmental protection, housing and community amenities, health, recreation, culture and religion, education, and social protection. In their review, excluding general public service, the functions of government in which there are more papers about AI are (1) economic affairs, (2) environmental protection, and (3) public order and safety. Finally, Reis and colleagues (2019a) provide a long list of areas in which AI may potentially change government but highlight that the impact to-date has been mainly on the delivery of government services, sometimes through partnerships with private companies.
AI results, impacts, and benefits
Four of the nine literature review papers highlight the results from AI as either general results or potential impacts or as benefits from its use in government settings. For example, Ahn and Chen (2020) present what they called AI-augmented bureaucracy and mention that the main outcomes are (1) accurate/detailed understandings of citizens’ needs and solutions, customizable service, enhanced simulation and planning capability via augmented reality, and predictive governance, which together could be conceptualized as smart government. In contrast, Reis and colleagues (2019b) identify impacts that could be positive or negative, such as job transformation, changes in government decision making, and citizen quality of life (including health and safety). Taking a more comprehensive view, Valle-Cruz and colleagues (2019) provide both potential benefits, but also some negative consequences. Some of the benefits mentioned are accuracy, efficiency, accountability, trust, cost saving, productivity gains, reduced fraud, better service provision, and improved policy analysis. In contrast, some of the potential negative implications are exclusion of certain actors, increased analysis complexity, new legislative requirements, dehumanization of daily activities, displacement of people by machines in their jobs, and a high dependence on intelligent technologies. Similarly, Wirtz and Müller (2019) also identify both prospective benefits and risks associated with the use of AI in the public sector. Among the benefits are improved information processing, accelerated processing of cases, improved case assignment, workforce substitution, and cutting red tape; some of the risks identified are technology obedience and loss of control, AI dominion, AI paternalism, and violation of privacy.
AI challenges and determinants
Four of the nine literature review papers discuss challenges of AI in government, and in some cases, they talk more generally about determinants. For example, Wirtz et al. (2019) propose a framework with 14 challenges to AI use in government settings. They include a few challenges that have been identified by previous research, such as data quality, privacy, and workforce substitution, but they also propose a way to group them in four categories related to technology implementation, society, ethics, and law and regulations. Focusing on the policy cycle, Valle-Cruz and colleagues (2019) have a very different view and propose specific challenges related to each of the stages of the policy cycle. Some of the challenges they mention are the digital divide, the cumbersome nature of the democratic process, goal displacement, and data obsolescence and homogeneity. In contrast, Desouza et al. (2020) categorize the challenges in design, development, and deployment issues related to cognitive computing systems. Some examples of the challenges they identify are data availability, current asset identification, disclosure of information, inadvertent bias in data and algorithms, lack of trained staff, and availability of tools to audit for bias. Taking a different approach to the topic of challenges, Ahn and Chen (2020) propose a series of questions that reflect some of the challenges related to the use of AI in government. Their main questions are: (1) How far are we going to allow AI to make decisions? (2) What would be the process of reconciliation when there is a conflict between AI augmented decisions and human-based decisions? (3) How will AI determine the “good” (or desirable) and the “bad” (undesirable) and, more importantly, for whose sake? (4) job displacement, (5) AI transparency and accountability, and (6) availability of relevant data.
AI strategies, best practices, and guidelines
From the nine literature review papers identified, only two include strategies, best practices, or guidelines to implement and use AI in government. For example, Desouza et al. (2020) present strategies to overcome challenges in the design, development, and deployment of cognitive computing systems. They also divide these strategies according to the public or private sectors. Some examples of strategies in the public sector are: assessing data availability, accessibility, and analyzability; focusing on the risk dimension, engaging outside experts, leveraging inherent government transparency, having an agile acquisition strategy, auditing algorithms to ensure accuracy, and taking advantage of tools. In contrast, Wirtz and Müller (2019) propose the following general guidelines: codifying ethical AI standards and regulations and monitoring their enforcement, setting an AI agenda defining targets, field of application, and a road map for employment, establishing limits and boundaries for AI usage and avoiding autonomous decision making, enhancing computer knowledge and AI-specific skills within the organization, providing insights to data acquisition and processing and creating verifiable AI algorithms, detecting options to automate administrative routine processes by means of AI, and enlarging the working capabilities of staff by AI usage.
Ethical considerations about AI
Three of the nine literature review papers specifically discuss ethical considerations to be considered when using AI in government settings. For example, Sousa et al. (2019) argue that “the policies and ethical implications of AI permeate all layers of the application” (Sousa et al., 2019, p. 6), including AI techniques, AI solutions to support public services, and the actual functions of government. They briefly reference several articles that support the importance of actively avoiding biases and discrimination and creating AI solutions for the good of society. As part of their general recommendations, Wirtz and Müller (2019) mention the importance of ethical AI standards and regulations. They propose the development of “a public AI code of ethics and to check and monitor its implementation by a public AI ethics committee” (Wirtz & Müller, 2019, p. 1087).
It is important to highlight that, while each of the nine papers touches on a few of the topics identified, they do so in significantly different ways, making it challenging to perform a more systematic comparison. However, they are still a good indication of what has been studied to date about AI in government and what opportunities exist for future research. AI has been studied from very different perspectives and, in general, there is no consensus on the most important aspects to consider. However, it is clear that all eight topic categories identified have been the focus of papers about AI in government within the last few years. It is also important to emphasize that there is still a dominant focus on the specific technologies and techniques, although more research is now highlighting organizational challenges and ethical issues. Only two review papers provide specific guidance or strategies, which could be an indication of the complexity in generating this type of guidance from a general point of view. Finally, we also want to emphasize that there are many potential applications of AI in the public sector, but the specific context of the program or agency must be considered in order to avoid some of the negative unintended consequences, such as biases and exclusion, and also to generate the expected benefits. Each use of AI in the public sector must be considered in context and in terms of both potential positive and negative consequences.
An evolving focus
As a complement to the analysis of the selected set of literature reviews provided above, here we draw on a set of recent empirical and conceptual papers to present some additional insight into the evolving literature on AI in the public sector. The goal was to identify a small number of nonliterature review papers published in 2020–2021. As above, these papers were selected using key words such as “artificial intelligence” and “public sector” or “government,” and we used Scopus and Google Scholar as our main databases and did not limit the search to certain fields or disciplines. The selected papers highlight what might be considered a logical progression from the generally descriptive consideration of the nature, potential, and challenges of AI to papers that systematically and empirically consider specific questions and challenges. For example, few publications in both the gray literature and the academic literature go beyond consideration of the public value potential to specific empirical evidence of value creation nor do they call, in general, attention to specific positive and negative consequences of use in specific contexts. The papers highlighted below collectively foreshadow movement beyond consideration of AI as a technology that is newly relevant to public administrations, to work that rigorously, and in some cases empirically, examines questions about AI use in the public sector, about specific uses of AI in public programs and services and begins to fill gaps in our available frameworks and models to assess, govern, and guide uses.
Bullock et al. (2020) and Wirtz et al. (2020) highlight gaps in knowledge about the relationship between AI, discretion, and bureaucratic form in public organizations. Despite the great potential of AI, Wirtz et al. (2020) note that “many challenges and risks are associated with implementing AI in public administration, constituting a darker side of AI” (Wirtz et al., 2020, p. 818). Drawing on insights from predictive policing and anti-fraud and improper payments reduction efforts, Bullock et al. (2020) seek to begin to close this gap by answering specific questions about how use of AI is both changing and changed by the bureaucratic form of public organizations and what is the consequence of use on discretion. Wirtz et al. (2020) look to regulation theory and former AI regulation approaches as models and put forward an integrated AI governance framework to guide the development and use of relevant regulations.
Van Noordt and Misuraca (2020a) and Makasi et al. (2021) draw attention to the lack of specific frameworks to assess the public value potential of AI as a consequence of specific uses in the public sector. Van Noordt and Misuraca (2020a) present what they identify as a first discussion on conceptual frameworks designed to support rigorous assessments of the effects of AI in government. Makasi et al. (2021) also call attention to this gap and the consequent lack of empirical evidence of value creation through the use of AI in public service management and delivery. Van Noordt and Misuraca (2020a) put forward a framework that takes into consideration the previous challenges of implementing information and communication technologies in government and is designed “to validate and truly assess the impact of Artificial Intelligence in government” (van Noordt & Misuraca, 2020a, p. 9).
Janssen, Brous, et al. (2020) put forward a new framework designed to reduce the risk associated with the use of big data algorithmic systems and increase accountability. Through its use, governments can “promote stewardship of data, processes and algorithms, the controlled opening of data and algorithms to enable external scrutiny, trusted information sharing within and between organizations, risk-based governance, system-level controls, and data control through shared ownership and self-sovereign identities” (Janssen, Brous, et al., 2020, p. 1).
Providing a specific case example of the argument presented by Janssen, Brous, et al. (2020) that without data governance to ensure quality and compliance with regulated uses, entrusting consequential decisions to AI introduces risk, Vogl (2020) draws attention to the need for research on data governance within specific contexts, such as the use of predictive analytics for the initial screening of cases in child protection services. Such consideration, he argues, calls first for attention to fundamental and well-known system and data problems. Vogl (2020) draws attention to the long-standing and intransigent problem of redundant records in human services systems and the potential consequences of the use of such data in this context. In the context of child protection, using data with redundant records exacerbates the complexity of decision making about program eligibility and service delivery decision making. Without attention to fundamental data and system problems, “valuable information about past abusive or neglectful behavior could be left out of analyses or predictions, or it could associate information about abusive or neglectful behavior with the wrong individual” (Vogl, 2020, p. 229). Vogl’s (2020) research challenges long-standing assumptions that problems with data can be fixed through the use of advanced technologies and techniques.
Liu et al. (2020) and Valle-Cruz et al. (2020) both offer systematic examinations of AI in specific use contexts, focusing, respectively, on AI and crowdsourcing and AI in public-sector budgeting. Both draw attention to the lack of understanding about the public value creation potential of AI in each of these specific use contexts. Liu et al. (2020) focus on the use of AI in crowdsourcing, drawing attention to the limits in our understanding of the connections between these “two types of intelligence and adoption conditions to properly utilize them for the public sector” (Liu et al., 2020, p. 224). Valle-Cruz et al. (2020) call for explorations of AI techniques in public budgeting but urge caution that the value creation might not be found where expected. In fact, they posit, the value might be in “supporting creative ways to analyze and understand the data used for specific government programs and policies” (Valle-Cruz et al., 2020, p. 241) rather than in automated decision making. They call for continuing consideration of AI in the use context of public budgeting highlighting the cost of some uses.
These selected papers illustrate a progression from the generally descriptive work characteristic of the early adoption phase of a new technology to work that examines more specific questions about value assessment, determinants of success and use in context, and among others. They illustrate the point that researchers are beginning to systematically examine AI in specific use contexts. Two of the selected papers address specific questions about the relationships between AI and discretion and the utility of regulation theory to efforts to adapt past models of AI regulation for use with future AI technologies. Two speak to the lack of specific frameworks to assess the public value potential of AI. Two speak to the specific role of data. The first addressing the general consequence of a lack of systematic governance of data; the second, addressing the consequences of a lack of data governance, or more specifically, the lack of attention to long-standing and intractable data problems, such as the specific context of child protection. Two final papers highlight the transition to work that examines the practical realities and potential of specific uses of AI in public programs and services, in particular, crowdsourcing and public-sector budgeting.
Taken together, these papers contribute to the development of questions about the extent to which previous work on the implementation of emerging technologies in the public sector is relevant with respect to AI, which requires particular adaptations, and in which cases we must start anew.
Contributing Perspectives
The papers in this special issue provide a mixture of cautious optimism and critical reflection about the potential of AI in the public sector. Some articles explore specific cases in depth, while others engage in big questions, such as the role of trust and discretion. The accepted articles include several of the topics identified in A Brief Review of Literature Review Papers section, particularly AI techniques and technologies and AI challenges and determinants (see Table 1). They explore these topics either conceptually or in the context of specific national realities. In fact, this special issue includes articles that examine AI in government settings in a variety of countries, including Belgium, China, Estonia, Sweden, the Netherlands, and the United Kingdom, providing evidence that considerations of AI in the public sector by both researchers and practitioners has become a global phenomenon. Some of the specific topics include the accuracy of different AI algorithms, interorganizational collaboration and AI adoption, trustworthiness in AI, factors that affect AI adoption and performance, the importance of human and technological agency, and adoption of AI in public administrations in the European Union. In terms of methods, the articles included in this special issue represent a diversity in research design and methods, including quasi-experiments, interviews, case studies, literature reviews, statistical analysis, desk research, and documents analysis. Table 2 provides an overview of the articles published in this special issue.
Summary of the Special Issue Articles: Issues, Countries, Topics, and Methods.
Note. AI = artificial intelligence.
In the article entitled “Will Algorithms Blind People? The Effect of Explainable AI and Decision-Makers’ Experience on AI-supported Decision-Making in Government,” Janssen, Hartog, et al. (2020) draw attention to the increasing use of computational AI algorithms to support decision making by governments. They note that regardless of this trend, AI often remains opaque to decision makers and lacks clear explanation for how decisions were made. Janssen and colleagues used an experimental approach to compare decision making in three situations: humans making decisions (1) without any support of algorithms, (2) supported by business rules, and (3) supported by ML. Their experiment shows that algorithms help decision makers to make more correct decisions. However, they found that even experienced persons were not able to identify all mistakes. Their findings imply that algorithms should be adopted with care and that selecting the appropriate algorithms for supporting decisions and training of decision makers are key factors in increasing accountability and transparency. This article shows that understanding the limitations of AI in government is as important as highlighting its potential benefits.
Campion et al. (2020), in their article entitled “Overcoming the Challenges of Collaboratively Adopting Artificial Intelligence in the Public Sector,” use a case study to examine the challenges that interorganizational collaborations face in adopting AI tools and implementing organizational routines to address them. The case study, involving a large research university in England and two different county councils in a multiyear collaborative project around AI, shows that the most important challenges facing such collaborations are a resistance to sharing data due to privacy and security concerns, insufficient understanding of the required and available data, a lack of alignment between project interests and expectations around data sharing, and a lack of engagement across organizational hierarchy. This article shows that implementing AI in the public sector faces important challenges, particularly related to interorganizational collaboration. Findings are consistent with previous research in proposing that the most important challenges are organizational or managerial in nature rather than technical.
The article entitled “Cultivating Trustworthy Artificial Intelligence in Digital Government” by Harrison and Luna-Reyes (2020) draws attention to a “growing consensus” about the potential of analytical and cognitive tools of AI to transform government in positive ways, but also notes that “AI challenges traditional government decision-making processes and threatens the democratic values within which they are framed” (Harrison & Luna-Reyes, 2020, p. 1). These conditions call for conservative approaches to AI that focus on cultivating and sustaining public trust. The authors use the extended Brunswik lens model as a framework to illustrate the distinctions between policy analysis and decision making as traditionally understood and practiced and how they are evolving in the current AI context. Through their recommendations for practices, processes, and governance structures to provide for trust in AI and for research that support them, the authors seek to provide a balanced view on the potential of AI in government, acknowledging its transformative potential, but also highlighting important challenges that may affect not only decision-making processes but also our democratic values. The results have important practical implications related to how to design processes and structures in government to build trustworthy AI applications.
Wang et al. (2020), in their article entitled “Understanding the Determinants in the Different Government AI Adoption Stages: Evidence of Local Government Chatbots in China,” investigate factors that influence local governments to adopt AI-powered chatbots and factors that influence the performance of chatbots postadoption. Drawing on a quantitative study of Chinese local authorities, the authors find that vertical administrative pressure, horizontal competition pressure, and environment readiness play different roles in different adoption stages. Although pressure can encourage local governments to implement chatbots, these governments’ readiness determined how well the chatbots perform after their initial adoption. Similar to more traditional technologies, decisions to adopt AI are affected by many factors and the benefits those applications will generate depend on their performance. This article contributes to a more nuanced understanding of some of the determinants of success by showing that the factors that affect adoption decisions are not the same as the factors that have an impact on performance.
In the article entitled “Digital Discretion: Unpacking Human and Technological Agency in Automated Decision Making in Sweden’s Social Services,” Ranerup and Henriksen (2020) present a case study of automated decision making driven by robotic process automation in social services in Sweden. The authors find that digitalization in social services has a positive effect on civil servants’ discretionary practices mainly in terms of their ethical, democratic, and professional values. The long-term effects and the influence on fair and uniform decision making also merit future research. In addition, this article finds that a human–technology hybrid actor redefines social assistance practices. Simplifications are needed to unpack the automated decision-making process because of the technological and theoretical complexities. The effect of AI on discretion in the public sector has been characterized by cautious optimism, while some authors strongly believe that the overall effect will be negative. This article shows that, at least in the short term, AI technologies can have a positive impact on civil servants’ discretionary practices. It contributes to a more detailed understanding of the potential consequences of AI in the public sector.
In their article entitled “Exploratory Insights on Artificial Intelligence for Government in Europe,” van Noordt and Misuraca (2020b) present findings from three cases of AI adoption in public-sector organizations. Their study finds strong similarities between the antecedents identified in previous academic literature and the factors contributing to the use of AI in government. The adoption of AI in government, they note, does not solely rely on having high-quality data but is facilitated by numerous environmental, organizational, and other factors that are strictly intertwined among each other. To address the specific nature of AI in government and the complexity of its adoption in the public sector, van Noordt and Misuraca propose a framework to provide a comprehensive overview of the key factors contributing to successful adoption of AI systems. Their framework goes beyond what they consider a narrow focus on data, processing power, and algorithm development often highlighted in the mainstream AI literature and policy discourse. This article highlights the intertwined nature of challenges related to data, organizational, and environmental aspects. It also proposes a framework to think about AI adoption success.
Conclusion
Research on AI in government is transitioning toward what promises to be a very important stage. After an initial stage characterized by a focus on mapping the risks and benefits of AI, with relatively little in terms of theorizing and unboxing processes and mechanisms, we are now witnessing a move toward systematic analysis of the benefits and challenges of design, management, adoption, and implementation of AI in government. The contributions included in this special issue well capture this transition and open up a series of avenues for future research that will become more essential as AI assumes an increasingly central role in government, including administrative processes, but also citizen service provision, and agency decision making.
Given the state of the art of both policies and research on AI in government, we highlight a number of areas worthy of increased focus. We are not aiming to be comprehensive with our list, but rather to prompt the research community to pay attention to research worthy issues that are gaining in relevance, and at the same time have not been adequately singled out in existing research agendas.
–
–
–
–
Research and policy initiatives focused on AI in government are developing rapidly and acquiring increasing relevance across the world. Given the dynamic nature of this complex phenomenon, it is necessary to take stock of the existing body of knowledge and monitor the evolving literature to ensure that we move forward in a way that will maximize the benefits and mitigate the risks of AI in a government setting.
Footnotes
Declaration of Conflicting Interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The authors received no financial support for the research, authorship, and/or publication of this article.