Bureaucrat or Artificial Intelligence: People’s Preferences and Perceptions of Government Service

Abstract

The increasing use of artificial intelligence (AI) in public service delivery presents important yet unanswered questions about citizens’ views of AI. Especially, are citizens’ perceptions of decisions made by AI different from those made by bureaucrats? We answer this question by conducting a conjoint experiment. Our results show that individuals prefer minority bureaucrats over AI to make decisions. This is particularly true for racially minoritized citizens. However, when passive representation within the bureaucracy is unavailable, racially minoritized individuals do not have a clear-cut preference between AI and out-group bureaucrats. Our findings provide insight into the interaction between automation, representation, and equity.

Full text

Bureaucrat or Artificial Intelligence: People’s Preferences and
Perceptions of Government Service∗
Dongfang Gaozhaoa, James E. Wright IIb, and Mylah K. Gaineyc
aDepartment of Political Science, University of Dayton
bAskew School of Public Administration and Policy, Florida State University
cIndependent Researcher
Abstract
The increasing use of artificial intelligence (AI) in public service delivery presents impor-
tant yet unanswered questions about citizens’ views of AI. Are citizens’ perceptions of decisions
made by AI different from those made by bureaucrats? We answer this question by conducting
a conjoint experiment. Our results show that individuals prefer minority bureaucrats over AI
to make decisions. This is particularly true for racially minoritized citizens. However, when
passive representation within the bureaucracy is unavailable, racially minoritized individuals do
not have a clear-cut preference between AI and out-group bureaucrats. Our findings provide
insight into the interaction between automation, representation, and equity.
Keywords: artificial intelligence, government service, administrative decision-making, discre-
tion, representation, representative bureaucracy, equity
Published in Public Management Review
January 3, 2023
∗Earlier versions of this manuscript were presented at a symposium at American University, 2021 Public Man-
agement Research Conference, and 2022 APPAM Annual Fall Research Conference. The authors thank participants
at the symposium and the conferences for helpful feedback. Address correspondence to Dongfang Gaozhao at
dongfang.gaozhao@gmail.com.
Dongfang Gaozhao is an Assistant Professor in the Department of Political Science at the University of Day-
ton. His primary research interests are in citizen-state interaction from the perspectives of information, technology,
and institutions, with a focus on social equity. James E. Wright II is an Assistant Professor in the Askew School
of Public Administration and Policy at Florida State University. His research interests center around public
management, public policy and social equity. Mylah K. Gainey is an independent researcher. Her research focuses
on deterrent, rehabilitative, and reintegration methods in the criminal justice system.
1

1 Introduction
Traditional public administration decisions have consisted of a public servant interacting with a
citizen1to determine the correct legal course of action for the citizen. However, there has been a re-
cent push to use artificial intelligence (AI) and algorithms2to help facilitate public decision-making
and service delivery. This is an initiative that began in the early 1990s and 2000s with public
organizations using advanced information technology to help them make better decisions in public
service delivery, such as predictive traffic congestion and COMPSTAT (Tong & Wong, 2000; Walsh,
2001). As part of this push for increased automation, scholars argue that AI-powered automation
can further increase organizational performance and efficiency in administrative decision-making
(Zekić-Sušac et al., 2021). Moreover, scholars argue that the use of digital technologies will help
administrators provide “better public services” while continuing to professionalize the public service
(Lindgren et al., 2019). One such example is the use of machine learning, a branch of AI that focuses
on data and algorithms to emulate human learning, to provide image or handwriting recognition
(IBM Cloud Education, 2020), which is then used to create chains of automation for straightforward
tasks (Veale & Brass, 2019). A second example that has become more prevalent in public service
is the use of automated call centers where AI is used to search documents and help agents to solve
customer inquiries (Mehr, 2017).
There is also a growing movement associated with trustworthiness in AI that has partly at-
tributed to the uptick of AI in public service. The concept dubbed “trustworthy AI (TAI)” is
predicated on the notion that buy-in and trust from individuals and organizations to use AI is pred-
icated on AI being designed with transparency, tangibility, and reliability as its hallmarks (Glikson
& Woolley, 2020; Medaglia et al., 2021; Thiebes et al., 2021). Similarly, European Commission
proposes that trust in AI is achieved by four principles, i.e., having respect for human decision-
making, fairness, preventing harm to others, and solvable solutions (Braun et al., 2021). When AI
is trustworthy, individuals will follow the system’s solutions or recommendations. This indicates
the need for AI to be trustworthy so that there can be path dependency between the AI setting
forth the recommendation and the individual or group following that recommendation (Aoki, 2020).
With scholars paying particular attention to trustworthiness in AI, there has been an increase
in AI usage in public administration. For example, applications for government programs (such
as unemployment benefits), government services (such as trash pick-up and pothole repair), and
requiring permits have moved to e-platforms that adopt AI (Gil et al., 2019). In the case of social
welfare, when individuals fill out application forms electronically, there could be two systems behind
a computer screen to review applicants and applications: bureaucrats and algorithms. In some cases,
algorithms start to work before a human decision-maker steps in—to search multiple databases and
screen out unqualified applicants to facilitate the decision-making process. As processes like this
become more salient within public administration, this drastically changes the operation and scope
of administrative decision-making (Bullock, 2019), which is the heart of public administration (Si-
mon, 1947). In this regard, understanding how it shapes equity in outcomes and representation has
1We use the term citizen to identify individuals affected by public employees’ decisions, rather than a specific legal
status of citizenship (Roberts, 2021).
2Technically, artificial intelligence refers to “the capability of a machine to imitate intelligent human behavior”
(Merriam-Webster, 2021b) whereas algorithms are a set of rules that a machine follows to achieve a particular
goal, such as imitating human intelligence (Merriam-Webster, 2021a). In this paper, for the sake of readability,
we use artificial intelligence (AI) and algorithms interchangeably. Similar usage is commonly seen, e.g., algorithm
appreciation and aversion (Dietvorst et al., 2015; Logg et al., 2019).
2

also become an important aspect, which has not been extensively studied.
Recent literature has started studying the relationship between automated decisions without
AI and the role of representation (representative bureaucracy) and found that traditional under-
resourced communities prefer automation using basic logic models but only in a traditionally an-
tagonistic public profession (policing), between communities of color and public service (Miller &
Keiser, 2021). However, automation is such an expansive field that even less is known about the
sub-disciplines it has created. To add to what we know about representative bureaucracy, in partic-
ular symbolic representation, there is a growing need for research in different policy areas outside
of policing. Given that automation with AI, as another form of automation that does not follow
the preprogrammed if-then logic to make decisions, would be far more intelligent (Gaynor, 2020;
Medaglia et al., 2021), understanding who benefits the most from AI-driven automation and if effi-
ciency and equity can co-exist in multiple public services areas is equally, if not more, important.
With this in mind, we conducted a conjoint experiment in which participants weighed two options
to deal with the quality of social welfare applications, namely assigning bureaucrats or AI reviewers
to examine applicant eligibility and application completion. We randomized a reviewer’s attributes
and asked participants to indicate their preference and predict each reviewer’s performance in terms
of efficiency, consistency, and ability to apply equity. The results show that all respondents regardless
of their race generally prefer a public employee who is an African American female, with 5-6 years of
training, to serve as the quality control reviewer. For African American participants, if they cannot
choose an African American bureaucrat to be their reviewer, they may regard other bureaucrats
and AI the same. In addition, we find that people believe that AI is more efficient than bureaucrats,
but less capable of applying equity. These findings add to the symbolic representation literature,
with its exploration comparing representation with automation. Our paper contributes in two ways.
First, it introduces a new policy area with important salience for citizen-state interactions (welfare).
Second, these results are the first to indicate that for Black individuals’ representation is most salient
for issues of equity and efficiency only when the government agent shares the same racial identity.
2 Literature Review
AI technologies are widely employed in today’s world, ranging from chatbots using natural lan-
guage processing, algorithm-driven marketing, medical image analysis, to financial modeling, and
supervising services involving facial recognition (Bughin et al., 2017; Jordan & Mitchell, 2015). In
the public sector, AI meets the needs of smart government to interact with citizens who spend
more and more time living and working in the digital realm, allowing public organizations to better
understand their citizens and clients (Margetts & Dorobantu, 2019; Vogl et al., 2020). Therefore,
AI has been used in social welfare programs to determine eligibility (Martinho-Truswell, 2018), in
courtrooms to predict recidivism (Van Dam, 2019), and in university admissions to predict student
performance (Moody, 2020). Depending on contexts and tasks, AI is complementing, supplanting,
or cooperating with human capabilities to make government decisions.
While the use of AI is expanding, our theories in the field of public administration have not
grappled with this trend (Liu & Kim, 2018). Only in recent years, related studies have grown and
touched on the potential impacts that AI may bring to public administration (e.g., L. Andrews,
2019; Busuioc, 2021; Wirtz et al., 2019; Young et al., 2019). These impacts include both the bright
side, such as better cost efficiency and restricted individual prejudice (Wirtz et al., 2019; Young
3

et al., 2019), and the dark side, including opacity in decision making (Busuioc, 2021) and favoritism
or unfairness due to algorithm manipulation ((L. Andrews, 2019). Still, it is largely unclear how
citizens understand and respond to the escalating involvement of AI in the government (Sharma
et al., 2020). Particularly, do people perceive government services provided by AI differently from
those provided by bureaucrats? Do they prefer AI or bureaucrats to make decisions in government
service? These questions are critical to public managers because the answers may directly suggest
what areas public organizations can leverage AI and what areas they still need humans to foster
people’s trust in the government (Ariely, 2013; Chingos, 2012). Successful government service needs
these answers as citizens’ positive perceptions cultivate public trust and satisfaction, leading to a
higher willingness to engage in coproduction and compliance with future policies among citizens
(Hibbing & Theiss-Morse, 2001; Im et al., 2014).
Many factors influence individuals’ attitudes toward the government and its service provider.
Studies have found that people’s attitudes may be positively associated with better performance
outcomes (Aytaç, 2021; Porumbescu et al., 2019), which can be affected by training and personnel
management. Dermol and Čater (2013) argue that years of training make public servants more pro-
fessional, leading to better performance outcomes. For AI agents, more time invested in improving
data quality, adjusting algorithms, and testing models is a general way to improve their performance
(Glikson & Woolley, 2020; Zewe, 2022). Nevertheless, actual performance is not necessarily equal
to perceived performance due to people’s differing needs for public service and understanding of
service performance (Bansal et al., 2021; Van de Walle & Bouckaert, 2003). Studies on citizen-state
interaction and TAI suggest that people’s views are also associated with performance information
(James & Moseley, 2014; Porumbescu et al., 2021), the administrative process concerning trans-
parency and justice (Grimmelikhuijsen, 2010; Van Ryzin, 2015), and autonomy (Song et al., 2021;
Thiebes et al., 2021). Moreover, a feeling of representation and empathy can affect people’s at-
titudes toward the government and its service providers. For this reason, researchers have found
that representative bureaucrats and anthropomorphic AI are more likely to gain people’s trust and
facilitate positive citizen-government interactions (Glikson & Woolley, 2020; Pelau et al., 2021; Van
Ryzin et al., 2017).
2.1 Attitudes Toward Government through A Representation Lens
Representative bureaucracy theory argues that the citizen-state interaction may benefit from shared
demographic and social characteristics between bureaucracies and the public (Bishu & Kennedy,
2020). When the bureaucracy’s workforce has similar demographic and socioeconomic compositions
to the constituent populations it serves, this is known as passive representation (Mosher, 1968). In
the decision-making process, bureaucrats may use their administrative discretion on a basis of their
favored passive representation, a moment in which active representation manifests and benefits the
represented social group (Bradbury & Kellough, 2007; Meier, 1975). Recent studies further investi-
gate the connection between representation and citizen perception of government service, pointing
out that a symbolically representative bureaucracy per se may create a feeling of commonality and
influence people’s attitudes (Riccucci et al., 2016; Roch et al., 2018; Theobald & Haider-Markel,
2008).
According to symbolic representation, without any changes in policy or affirmative steps made
by bureaucracies, citizens may improve their view of fairness, trustworthiness, legitimacy, and per-
ceived performance in government if they see matched social origins from the bureaucrats they are
interacting with (Gade & Wilkins, 2013; Scherer & Curry, 2010). Some researchers find that sym-
4

bolic representation can enhance people’s perception of government even when the administrative
outcome is unfavorable (Roch et al., 2018). Positive examples like this have important implications
for policy implementation, encouraging individuals to cooperate, comply with government decisions,
and coproduce desired policy outcomes (Hurwitz & Peffley, 2005; Riccucci & Van Ryzin, 2017).
Two types of social origins have been primarily examined in the research of symbolic represen-
tation, namely race and gender, partially because they are the most salient characteristics in the
United States to influence administrative behaviors and individual perceptions (Kennedy 2014). For
race representation, race-matching teachers may improve both parents’ and students’ perceptions
of school discipline as well as the fairness and legitimacy of bureaucratic behavior (Roch et al.,
2018). In a policing scenario, citizens are more likely to see police activities as legitimate when they
see police officers from their racial group present (Theobald & Haider-Markel, 2008). Also looking
into people’s evaluations of police service, Riccucci et al. (2018) find that race representation can
significantly affect people’s evaluations of performance, trustworthiness, and fairness. For Black
citizens, an increase in Black officers in a police department will lead to an increase in their overall
assessment of the department, even if the department faces more complaints about police miscon-
duct and has worse performance.
Similarly, studies find positive impacts of gender representation and congruence on people’s
attitudes toward government service (Baniamin & Jamil, 2021; Meier & Nicholson-Crotty, 2006;
Riccucci et al., 2014). For example, female police officers are more willing to support and actively
represent women in terms of domestic violence (R. Andrews & Miller, 2013). Relatedly, female
victims are also more likely to report sexual assault when facing female police officers (Meier &
Nicholson-Crotty, 2006). A potential explanation for this positive association is that female victims
may perceive gender-incongruent officers as less receptive and empathetic, deterring them from re-
porting crimes. In a recent study situated in a scenario of sexual harassment complication, female
complainants are found to prefer female mediators and rate their performance at a higher level
(Hibbard et al., 2022). These empirical studies indicate that better gender representation improves
people’s perceived job performance, trustworthiness, and fairness, increasing citizens’ willingness to
coproduce and comply (Baniamin & Jamil, 2021; Riccucci et al., 2014).
Many pieces of research have supported that racial and gender congruence between citizens
and bureaucrats can lead to citizens having more favorable attitudes toward government service.
However, a large portion of them concentrates on the fields of policing and education where race
and gender are salient enough to affect individuals’ perceptions (Cochran & Warren, 2012; Hilliard
& Liben, 2010). On the other hand, the effects of symbolic representation are inconclusive as its
influences on people’s attitudes vary by service area (Lee & Nicholson-Crotty, 2022). To address
this issue, this study situates itself in a scenario in which bureaucrats make low-stakes, instead of
urgent or significant, decisions for citizens so that race and gender are less salient. We hypothesize
that citizens in this situation have a choice of whom they would like to interact with:
H1: People will have more positive attitudes toward bureaucrats who share the same racial or
gender identities with them.
2.2 Lack of Representation: Out-group Bureaucrats and AI
Existing literature further looks into reverse representation, which occurs when citizens and gov-
ernment agents do not share commonalities (McLaughlin et al., 2021). In a police search context,
5

Theobald and Haider-Markel (2008) find that Black citizens are less likely to regard White officers’
activities as legitimate, whereas White citizens are less likely to view Black officers’ actions as le-
gitimate. Their results suggest that people may have negative perceptions of government agents
due to a lack of symbolic representation. Only when individuals see a symbolic representation of
themselves, may they be more open to opportunities for positive engagements with government
service (Riccucci et al., 2018). Otherwise, White citizens will respond to greater Black represen-
tation in a police department with increased negative attitudes toward the agency’s performance
and trustworthiness. And Black citizens’ perceptions of job performance, trust, and fairness may
decrease in the case of greater White representation.
However, some scholars argue that privileged individuals in terms of social origins may be less
concerned about the lack of representation (McIntosh, 1998; Miller & Keiser, 2021). Some empiri-
cal evidence supports this claim. As mentioned above, Black citizens’ perception of police fairness
varies substantially with police representativeness. In contrast, White citizens do not worry about
police officers’ fairness even though there is a greater representation of Black officers (Riccucci et al.,
2018). Regarding gender representation, there is also an asymmetry that both women and men care
more about women being underrepresented compared to men being underrepresented (Block et al.,
2019). Different treatments that Whites and males receive in daily life, as compared to non-White
individuals and females, may contribute to the disparity in their concerns about representation. An
example is that Black civilians may be subject to more force when facing White officers, whereas a
racial mismatch between White civilians and Black officers does not increase the level of the police
force (Wright & Headley, 2020). For privileged individuals, the advantages of dominating social
construction offset the disadvantages of missing symbolic representation at a certain point of gov-
ernment service delivery. Presumably, minorities would care more about representation and have a
strong preference for those who can represent them over those who cannot.
Having said that, minority citizens regularly interact with government services in which they
have no passive representation and perceive no symbolic representation. Therefore, when replacing
bureaucrats with AI agents in government decision-making and service delivery, this replacement
does not necessarily worsen minority citizens given that AI is another decision-maker that lacks
passive representation (Miller & Keiser, 2021). To be specific, people tend to perceive technology as
cold and inhuman (McFarland, 2015; Pols & Moser, 2009). While AI technologies appear in the real
world in a variety of ways, including physical robots, virtual assistants, and embedded functions,
these manifestations many times do not involve visualization, anthropomorphism, or social identi-
ties that human beings could connect with (Glikson & Woolley, 2020). The absence of commonality
prevents AI from representing people and alienates their trust in AI.
However, as the rose-colored glasses of representation fade when it compares AI with out-group
bureaucrats, minority citizens’ preferences now would be closely associated with their understanding
of how AI and out-group bureaucrats exercise discretion and make decisions. Regarding this, we
hypothesize:
H2: Minorities will have more positive attitudes toward AI when bureaucrats do not share the
same racial or gender identities with them.
This hypothesis is based on two underlying assumptions (1) that bureaucratic discretion and
AI are different, and (2) that people can perceive those differences and do have a preference. We
will discuss the first assumption in the next section and focus on the second one here. Prior
research documents an algorithm appreciation phenomenon that laymen prefer algorithmic to human
6

decisions (Logg et al., 2019; Thurman et al., 2019). In other words, decisions made automatically
by AI are frequently regarded as equal to or even better than those made by human professionals.
However, the argument for algorithm appreciation has mixed results and is not yet conclusive.
For example, Miller and Keiser (2021) discover that, in the case of traffic violation ticketing, Black
citizens, but not White citizens, are more likely to perceive AI better than police officers. In addition,
algorithm appreciation may be limited to the scenarios in which crucial decisions are involved, such
as those concerning safety, healthcare, and justice. People do not seem to have a strong preference
when human beings and AI are making trivial decisions (Araujo et al., 2020). Furthermore, and
quite the opposite, participants in some studies may constantly display a pattern of being averse to
AI even if they know AI outperforms human (Burton et al., 2020; Dietvorst et al., 2015). To this
end, we develop H2 to investigate people’s preferences in certain conditions.
2.3 Bureaucrats vs. AI: A Comparison in Efficiency, Consistency, and Equity
Although bureaucrats and AI both work under “rules of law,” the ways them making decisions are
arguably different. AI systems can process massive amounts of information within seconds and
be programmed to preserve public interest (Barth & Arnold, 1999; Chatterjee et al., 2022). This
principle requires AI agents to uphold fundamental human rights and promote the welfare of people
and the environment (Barth & Arnold, 1999; Toreini et al., 2020). On the other hand, bureaucratic
discretion refers to bureaucrats’ decision-making flexibility with which government policies are im-
plemented in complex and ambiguous problem environments (Bullock, 2019). Since bureaucrats
are motivated by self-interests and are cognitively bounded, bureaucratic discretion can possibly
involve discretionary abuse, personal bias, and administrative errors (Battaglio et al., 2019; Pren-
dergast, 2007; Simon, 1947). Even if improving bureaucrats’ professionalism and representation
may reduce the likelihood of these events (Dermol & Čater, 2013; Hong, 2017), they cannot rule
out prejudice and discrimination at the individual level, nor can they ensure consistent quality of
discretion across members of the organization. From this point of view, the use of AI in govern-
ment service has some key assets in terms of efficiency, consistency, and equity that help address
problems in bureaucratic discretion and general administrative decision-making (Young et al., 2019).
For many public organizations, inefficiency is the key obstacle to performance improvement.
Bureaucratic inefficiencies include unnecessary paperwork, administrative delays, and human er-
rors, which the use of AI could instrumentally help alleviate (Ingrams et al., 2022). Pandey and
Bretschneider (1997) predict that if the communication process is streamlined, many of the nega-
tive effects of red tape would vanish as they believe information and communications technology
can be used to contain red tape. This argument is later supported by Welch and Pandey (2006),
whose study identifies that a higher level of intranet implementation by public organizations is
associated with lower perceived levels of red tape because of the instrumental advantages in terms
of speed and efficiency. When it comes to AI in particular, it has good scalability that can easily
outpace human capacity for processing information and managing workload (Alexopoulos et al.,
2019; Wilson & Daugherty, 2018). For example, the U.S. Citizenship and Immigration Services
(USCIS) faces a class lawsuit for its long delays in handling work permits and immigration applica-
tions that are mostly caused by the limited workforce before and during the pandemic (Wiessner,
2020; Winokoor, 2021). On the other hand, the Internal Revenue Service (IRS) and its counterparts
around the world are employing AI systems to quickly detect tax evasion and noncompliance, a job
that traditionally would take weeks or months by manual reviews (Rubin, 2020). Because of its
machine learning-based architecture, AI systems are scalable and exploit information much faster
than human capacity, increasing government operation efficiency by improving per-task speed and
7

completing more tasks with lower marginal costs (Bullock, 2019; Ojo et al., 2019; Young et al.,
2019). As a result, higher efficiency achieved by AI may improve citizens’ perceptions of govern-
ment service (Kuziemski & Misuraca, 2020; Welch & Pandey, 2006).
Relatedly, scalable AI can provide better decision reliability or consistency than bureaucratic
discretion, which is an important perspective in shaping citizens’ perceptions of trustworthiness
(Toreini et al., 2020). Particularly, the reliability and consistency of AI are positively associated
with people’s acceptance (Glikson & Woolley, 2020). Compared to bureaucrats’ decision criteria
that may change from one agent to another, vary across different times, and even depend on various
moods (Andersen & Guul, 2019; Eren & Mocan, 2018), using AI technologies to make decisions can
prevent bureaucrats’ personal factors from influencing the decision-making process and uses one set
of algorithms to make decisions consistently and predictably (Mcknight et al., 2011; Young et al.,
2019).
The removal of bureaucrats’ personal bias has a special implication for minority citizens. In
agencies where the potential for representation is limited, in policy areas in which socialization may
limit the positive effects of passive representation, or if there is a lack of a critical mass of minority
bureaucrats, minority citizens may be or may perceive themselves to be placed at a disadvantage in
government service and administrative decision-making. From the standpoint of bureaucrats, they
may also feel unfamiliar with minority citizens’ situations and cannot serve minorities’ best interests.
In these scenarios, automated decision-making may be beneficial (Miller & Keiser, 2021). AI can
mitigate bias and enhance diversity and inclusion by being trained with loads of minorities’ data
and then be applied to places where bureaucrats from minority groups are hard to be recruited or
where minority clients are discriminated against by local officials (Daugherty et al., 2018; H. Zhang
et al., 2019). If minority citizens believe that unrepresentative bureaucrats are biased against them,
then they are likely to turn to AI for decision-making and service delivery.
With AI being incorporated into government operations, it is important to understand its lim-
itations. For now, there is still a long way to go to realize many magical benefits beyond simple
functions outlined by AI advocates as people are still figuring out how to take advantage of AI and
when (Hendler, 2020). In comparison to the private sector, there is less understanding of AI’s neg-
ative implications pertaining to government service and public decision-making (Wirtz et al., 2020;
Zuiderwijk et al., 2021). In this regard, a recent publication alerts the public sector of challenges
brought by AI, including the exclusion of certain stakeholders, increased complexity, dehumaniza-
tion of government service, infringement of privacy, technology obedience, as well as new legislative
and supervision requirements (Medaglia et al., 2021). In addition, the initial application of AI may
have rebound effects on efficiency and consistency as government operations need to adjust to the
new technologies (Nishant et al., 2020; van Leeuwen et al., 2022). Meanwhile, the lack of deep tech-
nical understanding of AI on the part of policymakers may lead to poorly designed or ill-informed
regulatory, legislative, or other policy responses (Brundage et al., 2018), exacerbating AI’s negative
implications. While there is a growing awareness of AI’s weakness in the government and society,
AI is increasingly autonomous and invisible, creating a black box in the decision-making process
that is hard to be explained or audited (Janssen et al., 2020; Zuiderwijk et al., 2021). Hence, it
could be difficult for public organizations to manage and for citizens to understand AI systems in a
transparent and accountable manner (Toreini et al., 2020; B. Zhang & Dafoe, 2020).
One of the consequences is that people doubt whether automated decisions can fully respect
the values of equity and fairness (Wachter et al., 2021). Their reservations are reasonable when
8

considering AI’s data feeding and learning nature. In theory, by relying on statistically fair links
between algorithmic inputs and the decision outcome, AI can decrease discrimination and promote
equity (Yang & Dobbie, 2020). Yet, human involvement in the development of AI systems can end
up creating inequity. System-level bureaucrats may contribute to the architecture and training of
algorithms while street-level bureaucrats may produce data for AI to train (Bovens & Zouridis, 2002;
Glikson & Woolley, 2020). The choice of using certain characteristics, such as race, gender, and even
seemingly neutral trait like education, in AI systems can cause unfair treatment of individuals if
these characteristics are correlated with discrimination (Barocas & Selbst, 2016). Furthermore, the
data used for training AI have been well-documented skewed and containing biases and limitations,
i.e., too few minorities’ data points and historical inequity and discrimination that are implicit
in these data (Zou & Schiebinger, 2018). As such, this can be damaging to minority groups and
poor populations, exacerbating the equity concerns of (and about) administrative decision-making
(Young et al., 2019). Based on the discussion above, we hypothesize that:
H3: People are likely to believe, in comparison to bureaucrats, AI to have better performance
on efficiency and consistency but worse performance on equity.
3 Method
We implemented a conjoint experiment to test our hypotheses. Conjoint experiments have been used
as a powerful means to capture and estimate individuals’ multidimensional preferences (Hainmueller
et al., 2014; Jilke & Tummers, 2018). A typical conjoint experiment asks participants to choose a
preferred profile from a group of profiles multiple times. In our case, we asked respondents to imagine
that their applications for social welfare were under review for applicant eligibility and application
completion. We have two reasons to situate this study in a social welfare context. First, it is
traditionally one of the scenarios in which bureaucrats and citizens, especially marginalized citizens,
frequently interact with each other (Hasenfeld et al., 1987; Keiser, 1999). Lipsky (1980) named
welfare officers as an example of street-level bureaucrats whose discretion occupies an influential
position in government service delivery. Second, social welfare is one of the services most likely to
be automated by AI as we move forward. Countries like the Netherlands, Denmark, and Australia
have employed AI technologies to detect welfare fraud and citizens at risk of neglect (Gantchev,
2019). Such uses have stirred up trust issues and public concerns about social justice, privacy
protection, and algorithmic transparency (Mann, 2020). However, the quality review is a tiresome
and monotonous task for bureaucrats and a low-stakes decision for citizens as compared to high-
impact decisions like police searches.
3.1 Experimental Design
In our conjoint experiment, each respondent faced three pairs of quality control reviewers and had
to choose one preferred reviewer from each pair. We presented four types of attributes in these
reviewer profiles, including reviewers’ identity, race, gender, and year of training. Table 1lists the
attribute levels we used in the experiment. We randomized the sequence of these attributes across
subjects to control for order effects but fixed the sequence within subjects to lower their cognitive
burden (Hainmueller et al., 2014). While we also fully randomized reviewers’ identity and years of
training, it is noteworthy that we employed restrictions on reviewers’ race and gender information to
exclude unrealistic attribute combinations. To be specific, algorithmic reviewers cannot be African
American, Caucasian, or Hispanic and should not have a male or female identity. Thus, our design
bounded algorithmic reviewers with “not applicable” racial and gender identities. Correspondingly,
9

we ruled out the possibility for government agent reviewers to have “not applicable” attributes. Our
subsequent analysis has considered these restrictions.
Table 1: Conjoint Attributes and Levels
Attribute Level
Identity Government agent, Algorithm
Race African American, Caucasian, Hispanic, Not Applicable
Gender Male, Female, Not Applicable
Year of Training Less than 1-year, 1-4 years, 5-6 years
We first asked about participants’ existing attitudes toward the importance of eligibility for social
welfare programs and the importance of equity in government decisions. Following these questions,
we presented participants with two reviewers’ profiles side by side on the screen and asked them four
questions that measured our dependent variables. The first question was a choice-based question
in which respondents must choose one preferred reviewer from two. The question reads “[w]hich
reviewer do you prefer to conduct quality reviews for social welfare applications?” Participants’
answers to this question will be referred to as choice outcomes hereinafter. After that, participants
were instructed to predict three dimensions of reviewers’ performance. We measured them to
understand whether their predicted reviewers’ performance influences their preference decisions.
The first dimension focuses on work efficiency, “how likely would these reviewers work efficiently?”
Respondents evaluated the likelihood on a five-point Likert scale varying from extremely unlikely to
extremely likely. Using the same scale, participants further projected reviewers’ work consistency
and ability to apply equity in their work. The next question reads “how likely would these reviewers
apply rules consistently to different people?” and the last question asks, “how likely would these
reviewers apply equity when reviewing applications?” We call participants’ answers to these rating
questions rating outcomes. To control for order effects, the order of these three rating questions was
randomized. As mentioned before, each participant rated three pairs of reviewer profiles. Upon the
end of their evaluation, a few questions regarding respondents’ characteristics were asked. Appendix
Aillustrates the flow of our research design.
3.2 Subject Recruitment
We pre-registered our study on Open Science Framework (OSF)3and received approval from the
Human Subjects Research Office at Florida State University.4We targeted 1,000 U.S. adults ad-
ministered through Amazon Mechanical Turk (MTurk) in October 2020. 1,301 individuals turned
out to participate in our experiment. Since our focus is on the opinions of people living in the
United States, we employed a protocol to detect participants with a non-U.S. IP address or a VPN
connection (Winter et al., 2019) and removed them per our pre-registration. We dropped 46 sub-
jects for this reason, 102 for their choices of withdrawal, 139 for not completing the survey,5and 44
for missing information. After these procedures, 970 respondents became our final sample. Since
participants in our conjoint experiment are asked to evaluate three pairs of reviewer profiles, this
study obtains 6 observations from each participant and 970 ×6 = 5,820 observations in total. We
report the descriptive statistics of respondents’ characteristics and responses in Appendix B. A
typical respondent in our experiment would be a Caucasian male who has a similar chance to be
3The pre-registration can be accessed via https://osf.io/mb7u4.
4The experiment has been reviewed for any potential harm to human subjects and granted an institutional review
board (IRB) “exempt” status (IRB Protocol STUDY00000993).
5Most incomplete participations happened after the survey’s quota on MTurk was fulfilled.
10

Republican or Democrat and identifies himself as moderately liberal. He has completed at least
undergraduate education and has a household income between $50,000 to $74,999. With 970 re-
spondents and 0.05 expected effect size, the predicted statistical power for our design is 86%, which
is generally acceptable in social sciences (Stefanelli & Lukac, 2020).
4 Analysis and Results
4.1 Conjoint Analysis
Looking into respondents’ choice outcomes, we conduct conjoint analysis and estimate each at-
tribute’s average marginal component effect (AMCE).6An AMCE captures the causal effect of a
reviewer’s attribute on the probability that this reviewer will be the preferred one. Since each par-
ticipant saw three pairs of reviewer profiles, we cluster standard errors by participant to account
for the potential non-independence of their choice outcomes.
Figure 1: Pooled AMCE Results
Identity
Race
Gender
Training
−0.1 0.0 0.1 0.2 0.3
Government agent reviewer
Baseline: Algorithm reviewer
Hispanic
Caucasian
Baseline: African American
Male
Baseline: Female
5−6 years of training
1−4 years of training
Baseline: Less than 1−year of training
Estimated AMCE
Attributes
Note: Estimates are based on the regression estimators with clustered standard errors. Bars show
95% confidence intervals. Regression coefficients are in Appendix C.
Figure 1shows the AMCE estimates and 95% confidence intervals for each attribute level. There
are six AMCE estimates relative to their baseline attributes. One may interpret these estimates as
the marginal effect of each attribute level on citizens’ preference of a particular reviewer. A positive
AMCE indicates citizens’ favorable attitudes toward the given attribute levels whereas a negative
value means unfavorable attitudes toward such level. Overall, our participants gave preference to a
government agent reviewer who was an African American female, with 5-6 years of training, to con-
duct a quality control review. In particular, participants were on average 10.3% (SE = 0.026) more
likely to choose a government agent rather than an algorithmic agent to review their applications.
Regarding reviewers’ training background, substantial training provides a bonus. Reviewers who
possess 5-6 years of training would be about 29.8% (SE = 0.017) more likely to become respon-
dents’ preferred reviewer in comparison with baseline reviewers with less than 1 year of training.
6For instance, the AMCE of 5-6 year of training on the probability of a reviewer being chosen as a preferred one can
be derived by: (1) estimate the difference in likelihoods that two reviewers who have two different levels of training
background, one being the baseline level and the other being the level of interest, but otherwise identical reviewers,
are chosen to be desired; (2) compute the same difference between two reviewers with these levels of training, but
with other possible combinations of profile attributes other than the year of training, i.e., algorithmic reviewers vs.
government agent reviewer; and (3) calculate the weighted average of these probability differences over the joint
distribution of all attributes (for the detailed estimation, see Hainmueller et al., 2014).
11

Relatively, 1-4 years of training has a less significant advantage over the baseline, increasing the
likelihood by 15.9% (SE = 0.016).
Given that we have restrictions to prevent meaningless attribute combinations, we do not
compare the relative effects between algorithms’ non-applicable identities and government agents’
gender- and race-specific identities. Instead, we examine the differences between government agent
reviewers. The difference in the probability of being chosen between male and female reviewers is
−0.037 (SE = 0.016), suggesting that females are 3.7% more likely to be people’s desired reviewers
than male bureaucrats. For racial identities, Caucasian and African American reviewers do not have
a statistically significant difference while a Hispanic reviewer is 5.7% less likely to be preferred over
an African American reviewer (SE = 0.018).
4.2 Heterogeneous Effects
In addition to AMCEs, we examine potential heterogeneous treatment effects, which can be caused
by the respondents’ characteristics. In other words, the causal effect of an attribute level is likely
to be conditional on participants’ personal characteristics. To control for this, we condition the
average of the attribute’s marginal effect on participants’ race and gender. Figure 2visualizes our
results which are reported in Appendix D.
Figure 2: AMCEs by Respondents’ Race and Gender
By Race
By Gender
Identity
Race
Gender
Training
−0.25 0.00 0.25 −0.25 0.00 0.25
Government agent reviewer
Baseline: Algorithm reviewer
Hispanic
Caucasian
Baseline: African American
Male
Baseline: Female
5−6 years of training
1−4 years of training
Baseline: Less than 1−year of training
Estimated AMCE
Attributes
Respondent's Characteristics African American Caucasian Other Race
Female Male
Note: Estimates are based on the regression estimators with clustered standard errors. Bars show
95% confidence intervals. Regression coefficients are in Appendix D.
Starting with race, it is worth mentioning that some racial groups in our study have relatively
12

small sample sizes, so we have classified them into the “Other Race” category. The results show
that both African American and Caucasian participants have a strong preference for government
agents over algorithms. Bureaucrat reviewers would be 16.6% and 11.6%, respectively, more likely
to be their reviewers in the quality control process than AI reviewers. Furthermore, African Amer-
ican participants show a strong preference, over 10% more likely, for African American reviewers
over other human decision-makers while Caucasian and other race participants do not show similar
in-group preferences. As for the reviewer’s gender, Caucasian respondents demonstrate a preference
for females (5.4% more likely) whereas others do not. The fourth attribute, year of training, is the
most regularly used reviewer attribute. Every racial group is disinclined to have a reviewer with
little training and most racial groups favor a reviewer with 5-6 years of training. The effects of 5-6
years of training increase the likelihood of being a preferred reviewer by 11.6% to 37.3%.
Then, we investigate the effects of reviewer attributes across respondents’ gender. We find a
consistent preference for government agent reviewers in both male and female respondents. More-
over, male respondents report a slight reluctance (5.6% less likely) to have a Hispanic reviewer
when the alternative could be a Black reviewer. Female subjects take the reviewer’s gender into
account and prefer female reviewers over males by 7.6%. Like what we have found from different
racial groups, reviewers’ year of training remains a crucial factor for male and female participants’
preference decisions. They tend to choose reviewers with longer years of training over those with
fewer years of training. These findings of racial and gender identities partially support H1, which
expects that individuals will prefer reviewers who share the same racial or gender identities with
them. Based on our results, only minority citizens care about representation and have a preference
for race and gender congruence between reviewers and themselves.
Because of this preference, it is natural to ask what would happen when minority citizens do
not have their preferred human decision-maker and whether they would turn to prefer AI. Our H2
posits that minorities would prefer AI reviewers when the alternative is out-group human agents
whose race or gender identity is incongruent with theirs. We test H2 with a closer look at a subset
of observations in which minority respondents choose between a human agent and an AI agent. For
this purpose, we use African American or female participants’ observations because other racially
minoritized categories do not have sufficient observations.
Figure 3presents the results for minority participants’ preference choices broken down by
whether the government agent reviewer is passively representing the participants. When making
these choices, the participants are offered two options: a human decision-maker and an AI reviewer.
The left facet shows the results when the government agent reviewer has a matched racial or gender
identity with the respondents. Put differently, they are passively representing the respondents. In
contrast, the right facet shows the results when the passive representation is not available. As
made clear by the right facet, we do not find evidence supporting H2, which would need a statisti-
cally significant negative difference in minority respondents’ preference between incongruent human
decision-makers and AI reviewers. Nevertheless, Black participants in the upper panel perceive
little difference between out-group bureaucrats and AI taking charge of the decision-making process
while they perceive statistically significant differences between in-group bureaucrats and AI. For
female participants in the lower panel, they do not pay too much attention to bureaucrats’ social
origin and regard gender-congruent and -incongruent agents the same when the other option is AI
reviewers.
13

Figure 3: Minority Subjects Choosing between Bureaucrat and AI
Black vs. algorithm reviewers
Non−black vs. algorithm reviewers
Identity
Gender
Training
−0.4 −0.2 0.0 0.2 0.4 0.6 −0.4 −0.2 0.0 0.2 0.4 0.6
Government agent reviewer
Baseline: Algorithm reviewer
Male
Baseline: Female
5−6 years of training
1−4 years of training
Baseline: Less than 1−year of training
Attributes
Black Subjects
Female vs. algorithm reviewers
Male vs. algorithm reviewers
Identity
Race
Training
−0.6 −0.3 0.0 0.3 0.6 −0.6 −0.3 0.0 0.3 0.6
Government agent reviewer
Baseline: Algorithm reviewer
Hispanic
Caucasian
Baseline: African American
5−6 years of training
1−4 years of training
Baseline: Less than 1−year of training
Estimated AMCE
Attributes
Female Subjects
Note: Estimates are based on the regression estimators with clustered standard errors. Bars show
95% confidence intervals. Regression coefficients are in Appendix E.
4.3 Logistic Regression
Lastly, we shift gears to three rating outcomes. Recall that participants were asked to predict
human and AI reviewers’ work efficiency, consistency, and ability to apply equity on a scale from
1 (extremely unlikely) to 5 (extremely likely). We dichotomize these ratings into 0 (not likely)
and 1 (likely) and then regress this binary variable on reviewer attributes as well as respondents’
characteristics and priors in simple logistic regression models. For the models comparing government
agents and algorithm reviewers,
logit (Pr (Yi= 1)) = β0+β1Reviewer +β2Year of Training +γControl +εi,
where Yi= 1 denotes that participant i’s rating is “likely.” Control indexes a 1×pvector of
respondent i’s gender, race, age, party affiliation, education, household incomes, and their preexist-
ing attitudes toward social welfare eligibility and social equity. γrepresents a p×1vector of the
coefficients for Control. Similarly, for the models comparing within government agent reviewers,
logit (Pr (Yi= 1)) = β0+β1Reviewer Race+β2Reviewer Gender+β3Year of Training+γControl+εi.
Since our logistic regression models are a family test regarding respondents’ evaluations of effi-
ciency, consistency, and ability to apply equity, we adjust the results’ p value by using the Bonferroni
14

correction and establishing the significance level at (0.1/3) = 0.033. Table 2demonstrates the main
results. When comparing government agents with algorithm reviewers, human decision-makers are
hypothesized in H3 to perform poorer than AI reviewers in terms of efficiency and consistency but
better regarding applying equity. The results find evidence to partially support H3. Specifically,
AI is rated to be lower in equity than government agents to a statistically significant level. Put
differently, according to our respondents’ prediction, government agent reviewers have 38% more
odds of having the ability to apply equity than AI reviewers. And AI reviewers are predicted to have
18% more odds of working efficiently than government agent reviewers. However, the difference in
efficiency becomes no longer statistically significant after applying the Bonferroni correction. Other
than that, human abilities to apply rules consistently to different people are perceived to be the
same as AI, a result opposite to what we hypothesize. Novice agents who receive less than 1 year
of training are anticipated to be less capable than those agents who have more years of training in
all three aspects.
Table 2: Logistic Regression Results
Attributes Algorithm vs. Government Agent Within Government Agents
Efficiency Consistency Equity Efficiency Consistency Equity
Reviewer (Baseline = Algorithm)
Government agent −0.200*
(0.097)
0.029
(0.090)
0.322***
(0.087) NA NA NA
Reviewer Race (Baseline = African American)
Caucasian NA NA NA −0.010
(0.086)
−0.058
(0.084)
−0.174
(0.084)
Hispanic NA NA NA −0.068
(0.084)
−0.203**
(0.082)
−0.074
(0.085)
Reviewer Gender (Baseline = Female)
Male NA NA NA −0.091
(0.069)
−0.012
(0.067)
−0.084
(0.069)
Year of Training (Baseline = Less than 1-year of training)
1-4 years of training 0.598***
(0.077)
0.410***
(0.075)
0.231***
(0.077)
0.614***
(0.083)
0.446***
(0.081)
0.254***
(0.083)
5-6 years of training 0.726***
(0.079)
0.487***
(0.075)
0.317***
(0.077)
0.745***
(0.084)
0.491***
(0.081)
0.332***
(0.083)
Race (Baseline = Caucasian)
American Indian or
Alaska Native
−0.171
(0.197)
0.225
(0.202)
0.183
(0.204)
−0.155
(0.209)
0.230
(0.213)
0.087
(0.215)
Asian or Pacific Islander −0.593***
(0.120)
−0.593***
(0.116)
−0.640***
(0.117)
−0.670***
(0.128)
−0.603***
(0.125)
−0.741***
(0.126)
Black or
African American
0.309***
(0.088)
0.245**
(0.084)
0.158
(0.084)
0.333***
(0.095)
0.268**
(0.091)
0.147
(0.092)
Hispanic or Latino 0.461**
(0.175)
0.143
(0.157)
0.302
(0.165)
0.393*
(0.183)
0.114
(0.166)
0.173
(0.173)
Mixed racial background −0.395
(0.207)
−0.074
(0.214)
−0.376
(0.205)
−0.589**
(0.221)
−0.195
(0.229)
−0.587**
(0.221)
Gender (Baseline = Female)
Male −0.006
(0.068)
−0.038*
(0.066)
0.000
(0.066)
−0.053
(0.073)
−0.089**
(0.071)
−0.047
(0.072)
Age (Baseline =18 to 24)
25 to 34 −0.096
(0.141)
−0.041
(0.134)
−0.095
(0.137)
−0.044
(0.148)
0.045
(0.141)
−0.013
(0.146)
15

35 to 44 0.045
(0.147)
0.199
(0.140)
0.078
(0.143)
0.083
(0.155)
0.293*
(0.148)
0.105
(0.152)
45 to 54 0.002
(0.159)
0.030
(0.151)
0.014
(0.155)
0.022
(0.168)
0.107
(0.160)
0.036
(0.165)
55 and over −0.011
(0.165)
0.199
(0.160)
0.250
(0.164)
0.158
(0.176)
0.351
(0.171)
0.355
(0.177)
Party (Baseline = Democrat)
Independent −0.265**
(0.093)
−0.217**
(0.090)
−0.368***
(0.090)
−0.286*
(0.099)
−0.176
(0.096)
−0.339**
(0.098)
Republican 0.000
(0.085)
0.048
(0.081)
−0.017
(0.082)
0.035
(0.091)
0.165
(0.087)
−0.027
(0.090)
Ideology
(Greater = More liberal)
−0.041
(0.027)
−0.020
(0.026)
−0.002
(0.027)
−0.044
(0.029)
−0.004
(0.028)
−0.004
(0.029)
Education (Baseline = High school or lower)
Some college but no degree 0.029
(0.148)
0.348*
(0.140)
0.054
(0.145)
−0.023
(0.157)
0.324*
(0.149)
0.042
(0.156)
Associate degree 0.493**
(0.163)
0.522***
(0.150)
0.354*
(0.156)
0.634***
(0.178)
0.668***
(0.164)
0.472**
(0.172)
Bachelor’s degree or higher 0.190
(0.127)
0.529***
(0.119)
0.271*
(0.125)
1.171
(0.135)
0.511***
(0.127)
0.284
(0.134)
Incomes (Baseline = Less than $24,999)
$25,000 to 49,999 0.048
(0.116)
0.010
(0.112)
−0.196
(0.116)
0.066
(0.123)
−0.024
(0.120)
−0.219**
(0.126)
$50,000 to 74,999 −0.097
(0.115)
−0.075
(0.112)
−0.187*
(0.116)
−0.065
(0.122)
−0.066
(0.120)
−0.231**
(0.127)
$75,000 to 99,999 0.281
(0.129)
−0.065
(0.122)
−0.108
(0.127)
0.259
(0.137)
−0.093
(0.131)
−0.083
(0.139)
$100,000 and greater 0.210
(0.137)
0.143
(0.133)
−0.156
(0.135)
0.239
(0.145)
0.107
(0.142)
−0.135
(0.147)
Social welfare eligibility 0.158*** 0.193*** 0.202*** 0.131*** 0.174*** 0.187***
(Greater = More agree) (0.023) (0.023) (0.023) (0.025) (0.025) (0.025)
Social equity priority
(Greater = More agree)
0.182
(0.022)
0.143
(0.021)
0.141
(0.021)
0.179
(0.023)
0.147
(0.022)
0.143
(0.023)
Intercept −0.869***
(0.259)
−0.443***
(0.249)
−1.134***
(0.253)
−0.871***
(0.268)
−1.391***
(0.260)
−0.630*
(0.267)
N(observations) 5,820 5,820 5,820 5,023 5,023 5,023
Note: The value without parentheses is the coefficient estimate of the given level. Standard errors
are in parentheses. Significance levels are corrected by the Bonferroni correction and established at
0.033. ∗p < 0.033; ∗ ∗ p < 0.017; ∗∗∗p < 0.003.
For differences within government agent reviewers, government agents’ race and gender do not
affect people’s expectations of work efficiency. However, for the other two dimensions, our respon-
dents believe that, in comparison to an African American agent, a Hispanic agent may be 18.4%
less likely to consistently apply rules to a significant level whereas a Caucasian agent would be
less likely to apply equity in their daily work, which becomes not significant after the Bonferroni
correction. Similarly, people have reservations about novice bureaucrats for their capabilities to
achieve efficiency, consistency, and equity.
We also explore the impacts of participants’ racial and gender characteristics on their projection
of reviewers’ performance as these characteristics are closely related to participants’ understanding
16

of representation. When compared to Caucasian participants’ viewpoints, Asians or Pacific Is-
landers are less likely to have positive faith in reviewers’ work efficiency, consistency, and ability to
apply equity while Black or African Americans are more likely to have higher perceptions regarding
efficiency and consistency. In comparison with female participants, males are less likely to believe
that decision-makers in the government, either the automated or human ones, would apply rules to
different people fairly and impartially. This gender difference is statistically significant before the
Bonferroni correction and becomes not significant afterward.
5 Discussion and Conclusion
In this study, we conducted a conjoint experiment to compare human bureaucrats with AI decision-
makers in terms of citizens’ preferences and perceptions. One of the primary contributions of this
study is the examination of citizens’ preferences for bureaucrats versus AI in the context of making
trivial decisions, i.e., quality control, in which representation is not a salient issue. We find evidence
for people’s ability to perceive differences between bureaucratic discretion and AI. Furthermore, we
disentangle individuals’ preferences for the intricate attributes of human and algorithm decision-
makers to some degree. We find that citizens tend to choose bureaucrats over AI to make government
decisions. However, this needs to be explored further in different contexts with low/high trust gov-
ernment agencies and in different policy areas, such as education, health, and policing. That said,
in our study, citizens generally prefer a human agent who is an African American female with sub-
stantial training.
When connecting people’s preference choices with their ratings, we find that they rate algorithm
reviewers to have a higher level of work efficiency but a lower level of ability to apply equity than
bureaucrats. These findings suggest the reasons behind people’s preferences. It indicates that when
employing AI in the government decision-making process, incorporating equity and fairness into the
decision-making process in all phases and communicating with people how AI can be trustworthy
would be important to improve citizens’ acceptance of AI in government (Toreini et al., 2020; Zou
& Schiebinger, 2018). This can work in several ways related to TAI. First, when designing the AI
system, the development team may want to recruit diverse and multidisciplinary members, with
experts focusing on discrimination and bias in the profession. In this way, the system’s underlying
architecture is likely to incorporate the TAI principle of fairness and promote the interests of both
minorities and non-minorities (Toreini et al., 2020). Second, since most AI models must go through
a pre-testing phase, this phase can include equity testing. System-level bureaucrats and outsourcing
companies responsible for designing the AI system can establish equity metrics and goals that help
ensure all groups are sufficiently represented and accurately accounted for in the data-generating
processes (Zuiderwijk et al., 2021). To have meaningful tests, public organizations are recommended
to adopt feedback lopes and partnerships with a variety of stakeholders, particularly the community
of color. Third, public organizations should publish reports regularly to disclose the AI system’s
decision patterns. By increasing the AI system’s transparency and ability to be perceived, we can
supervise and minimize discrimination and bias in the data processing and decision-making phase
(Glikson & Woolley, 2020).
Looking into human decision-makers’ racial and gender identities specifically, we hypothesized
that individuals would have more positive attitudes toward in-group bureaucrats. We find partial
evidence for this relationship. Indeed, Black participants and female participants are more likely
to prefer Black bureaucrats and female bureaucrats, respectively. Nevertheless, White participants
17

and male participants do not necessarily have this in-group favoritism. In addition, previous liter-
ature suggests that people may prefer automated decision-making when passive representation is
not available (Miller & Keiser, 2021). While we find no evidence for this result from the minority
participants in our experimental setting, we do learn that, under the circumstance of unavailable rep-
resentation, African Americans no longer have a strong preference between out-group bureaucrats
and AI decision-makers. These results imply that racially minoritized citizens value government
representation and they give preference to those bureaucrats who share their demographic origins.
They may view those bureaucrats as more cognizant of their needs and more willing to use their
discretion in meeting citizens’ needs. This pattern is consistent with past studies of passive and
symbolic representation studies that articulate this relationship between racially minoritized citi-
zens and government agents (Miller & Keiser, 2021; Riccucci et al., 2016; Roch et al., 2018).
Together these results provide important theoretical implications for the study of representa-
tion and raise several new questions about the interaction between representation, automation, and
equity. First, when studying frontline worker decision-making, we frequently focus on the role of
representation in high-impact decision scenarios without understanding the underlying mechanism
of how representation is salient. If we do not consider the context, we can miss the extent to which
representation is salient and under what circumstances it matters more to government service de-
livery and policy implementation. To this end, our study finds that citizens care about passive
representation even in a low-impact decision scenario in social welfare, in which bureaucrats and
AI decision-makers are only performing technical and trivial duties. Relatedly, we begin to un-
pack under what situations AI can be viewed as optimal, e.g., when passive representation is not
available. This leaves important questions to discern how or if citizens would still view represen-
tation as the most important aspect despite perceived efficiency not being a top priority in their eyes.
Further, we investigated the impacts of decision-makers’ skill attributes and hypothesized that
people would prefer reviewers who had more years of training. Our results support this expec-
tation and show that individuals choose experienced human decision-makers or AI reviewers over
their novice counterparts in welfare applications. However, we caution that this result is context-
dependent and needs to be explored in other social services and citizen-state interactions. Other
factors about the reviewer may also impact these outcomes, such as previous decisions they have
made regarding citizens’ applications or if they have worked with other government agencies before.
Citizens’ appreciation of training has important practical implications for both human resource
management and the adoption of AI. For the former, it means that citizens are likely to support
maintaining continuous training of current employees and strengthening new employees’ training.
It also suggests that, when recruiting a diverse body of public servants to improve representation
is challenging, public organizations may choose to focus more on current personnel’s training. Al-
ternatively, if bureaucrats prefer individuals who look like them this implies that agencies need to
make a concerted effort to recruit and retain racially minoritized individuals within organizations.
Organizations typically focus on retention, but also need to devise a clear strategy to retain indi-
viduals of color to create a more inclusive and diverse environment in the positions individuals of
color occupy in the organization.
For the employment of algorithms in government decision-making, using the most advanced
technology does not always result in improved citizen evaluation. A more tuned AI system may
seem to be a safer choice in the citizens’ eyes, as the literature on trustworthy AI has suggested
that reliability over time is crucial to increasing technology trustworthiness (Glikson & Woolley,
2020). Relatedly, long-term interactions between citizens and AI also reduce people’s need for hu-
18

man contact and AI’s uncertainty as AI systems becomes more tangible, predictable, and reliable,
developing citizens’ trust in the algorithm decision-maker (Krämer et al., 2018). As AI systems
become more fine-tuned it may cause ripple effects with human employment, or fundamentally shift
the role of humans. If individuals can create an AI system that is not only efficient but equitable in
its decision-making front-line workers will need to have increased data skills and less interpersonal
skills. To account for the increased usage of AI to manage services, schools and human resource
training within organizations need to change how they train future employees. Finally, it is worth
noting that without further study of the tradeoff between AI and humans making decisions, it is dif-
ficult to tell if these results are consistent across policy types, but they can be explored. Additional
research can examine what decisions are easily realized by AI and what decisions are more difficult
because of the discretionary nature of the interaction between the bureaucrat and the citizen.
Finally, there are several limitations worth noting in our study. First, our experiment does not
oversample different groups of minoritized individuals, so some groups have relatively small sample
sizes. Accordingly, the experimental results are mainly based on Black and White participants who
are males and females. A future study oversampling other minority groups would be a good way
to verify and extend our findings. Second, our design involves a hypothetical situation that asks
participants to indicate their preferences. Although the conjoint experiment has unique properties
of reducing social desirability bias and simulating real-world decision-making by presenting various
attributes jointly to participants (Hainmueller et al., 2014), people’s evaluation and preference, in
reality, can be also contingent upon whether the reviewer’s decision is a positive or negative out-
come, how they interact with a human or AI reviewers, and various other factors. Third, our study
does not measure participants’ understanding of machine-learning algorithms. While this under-
standing does not necessarily affect our findings because of randomization, including it in the model
could have produced more nuances regarding people’s preferences between AI and bureaucrats. We
recommend future research to measure this and investigate its influence on individuals’ preferences.
Despite these limitations, our study provides important insights into how citizens view AI and
bureaucrats when reviewing their applications for services. Scholars and practitioners in public
administration have been working on improving organizational efficiency by, for example, reforming
service delivery and adopting new technologies. Meanwhile, social equity has become the third pillar
of public administration (Frederickson, 1990) and better representation in public organizations has
been found as a useful means of promoting improved service quality. Given that AI is expected to
serve as a complement to, if not a substitute for, human expertise in the long run, the investigation
into people’s preferences of decision-makers in government service delivery inspires discussions about
what tasks are appropriate for AI and what for bureaucrats in future government operations. To
boost people’s trust in government service delivery and decision-making, public organizations and
managers should develop and implement assimilation strategies that synergize AI with bureaucratic
discretion (Alshahrani et al., 2022). To be specific, the responsibility of making decisions and
interacting with citizens for different things should be taken up by different decision-makers. For
those services that citizens prioritize equity and fairness, bureaucrats may be more suitable for
delivering the services and gaining people’s trust. On the contrary, it would be appropriate for AI
agents to perform time-sensitive functions that people value efficiency. From this point of view,
this study points to a potential direction for future research on AI in government. Researchers
and practitioners should pay attention to the technological aspects of AI technologies and the
organizational aspects of AI applications in the public sector. By giving nuances to various decisions
and services, governments can maximize the benefits of AI and representative bureaucracy and
balance the values of efficiency and equity.
19

Appendix
A Experimental Setup
Respondents to study invitation ( 𝑛= 1,301)
Warm-up questions about respondents’ evaluation of the importance of eligibility to social welfare programs
and equity to government decision.
Many social welfare applications are incomplete or ineligible. The government is weighing two options to deal
with the quality of social welfare applications. Two approaches to conducting quality control are being
considered: (1) Assigning government agents to review; (2) Deploying machine-learning algorithms to review.
The cost difference between these two options is negligible and the government supports either one. They need
your feedback to implement the new protocol.
Choice outcome (reviewer attributes randomized): Assume that you are making a decision to choose a reviewer
you prefer. For each pair of reviewers, please indicate your attitudes toward the reviewers involved.
Rating outcomes (order randomized): On a scale from 1 to 5, how likely would these reviewers work
efficiently? How likely would these reviewers apply rules consistently to different people? How likely would
these reviewers apply equity when reviewing applications?
Evaluate 3 pairs of
scenarios in total
Questions related to respondents’ characteristics
20

B Descriptive Statistics (n= 970)
Characteristic Frequency Percentage Characteristic Frequency Percentage
Gender Party
Female 384 39.59 Democrat 407 40.14
Male 586 60.41 Independent 167 16.47
Republican 419 41.32
Race
American Indian or Alaska Native 27 2.78 Ideology
Asian or Pacific Islander 66 6.80 Strongly conservative 177 18.25
Black or African American 225 23.20 Moderately conservative 186 19.18
Caucasian 591 60.93 Neutral 187 19.28
Hispanic or Latino 41 4.23 Moderately liberal 213 21.96
Mixed racial background 20 2.06 Strongly liberal 207 21.34
Age
18 to 24 58 5.98 Education
25 to 34 395 40.72 High school or lower 71 7.32
35 to 44 272 28.04 Some college but no degree 107 11.03
45 to 54 140 14.43 Associate degree 93 9.59
55 and over 105 10.82 Bachelor’s degree or higher 699 72.06
Incomes
Less than $25,000 106 10.93 $75,000 to 99,999 163 16.80
$25,000 to 49,999 251 25.88 $100,000 and greater 125 12.89
$50,000 to 74,999 325 33.51
21

C Pooled AMCE Results
Attribute Result
Reviewer (Baseline: Algorithm)
Government agent 0.103***
(0.026)
Reviewer Race (Baseline: African American)
Caucasian −0.035
(0.019)
Hispanic −0.057**
(0.018)
Reviewer Gender (Baseline: Female)
Male −0.037**
(0.016)
Year of Training (Baseline: Less than 1-year of training)
1-4 years of training 0.159***
(0.016)
5-6 years of training 0.298***
(0.017)
N(observations) 5,820
N(individuals) 970
Note: Standard errors in parentheses. ∗p < 0.05; ∗ ∗ p < 0.01; ∗∗∗p < 0.001.
22

D Heterogeneous Treatment Effects
Attribute
Race of Respondent Gender of Respondent
African American Caucasian Other Race Male Female
Reviewer (Baseline: Algorithm)
Government agent 0.166** 0.116*** −0.053 0.087** 0.127**
(0.053) (0.034) (0.066) (0.033) (0.044)
Reviewer Race (Baseline: African American)
Caucasian −0.109** −0.016 0.012 −0.046 −0.021
(0.041) (0.024) (0.046) (0.025) (0.030)
Hispanic −0.150*** −0.038 0.001 −0.056*−0.058
(0.040) (0.023) (0.043) (0.023) (0.030)
Reviewer Gender (Baseline: Female)
Male −0.013 −0.054** 0.003 −0.010 −0.076***
(0.034) (0.020) (0.038) (0.020) (0.024)
Year of Training (Baseline: Less than 1-year of training)
1-4 years of training 0.123*** 0.163*** 0.185*** 0.146*** 0.179***
(0.033) (0.020) (0.038) (0.019) (0.026)
5-6 years of training 0.116** 0.345*** 0.373*** 0.278*** 0.328***
(0.035) (0.022) (0.040) (0.021) (0.028)
N(total observations) 1,350 3,546 924 3,516 2,304
N(total individuals) 225 591 154 586 384
Note: Participants who reported their racial and ethnicity identities as American Indian, Asian,
Hispanic, or mixed racial background are classified into the “Other Race” categories for their small
sample sizes. Standard errors in parentheses. ∗p < 0.05; ∗ ∗ p < 0.01; ∗∗∗p < 0.001.
23

E AMCE Results of Minority Subjects Choosing between Bureaucrat and AI
Attribute
Black Participants Female Participants
Black vs.
algorithm
reviewers
Non-Black vs.
algorithm
reviewers
Female vs.
algorithm
reviewers
Male vs.
algorithm
reviewers
Reviewer (Baseline: Algorithm)
Government agent 0.301* 0.105 0.160 0.118
(0.143) (0.106) (0.098) (0.113)
Reviewer Race (Baseline: African American)
Caucasian NA NA 0.064 −0.042
(0.094) (0.110)
Hispanic NA NA −0.114 −0.026
(0.103) (0.109)
Reviewer Gender (Baseline: Female)
Male −0.143 −0.029 NA NA
(0.138) (0.098)
Year of Training (Baseline: Less than 1-year of training)
1-4 years of training −0.016 0.246** 0.098 0.063
(0.113) (0.086) (0.074) (0.071)
5-6 years of training 0.118 0.179 0.186* 0.125
(0.126) (0.084) (0.073) (0.073)
N(observations) 124 206 294 274
N(individuals) 56 84 128 126
Note: Standard errors in parentheses. ∗p < 0.05; ∗ ∗ p < 0.01; ∗∗∗p < 0.001.
24

References
Alexopoulos, C., Lachana, Z., Androutsopoulou, A., Diamantopoulou, V., Charalabidis, Y., & Lout-
saris, M. A. (2019). How machine learning is changing e-government. Proceedings of the 12th
International Conference on Theory and Practice of Electronic Governance, 354–363. https:
//doi.org/10.1145/3326365.3326412
Alshahrani, A., Dennehy, D., & Mäntymäki, M. (2022). An attention-based view of AI assimilation
in public sector organizations: The case of Saudi Arabia. Government Information Quarterly,
39 (4), 101617. https://doi.org/10.1016/j.giq.2021.101617
Andersen, S. C., & Guul, T. S. (2019). Reducing minority discrimination at the front line—Combined
survey and field experimental evidence. Journal of Public Administration Research and The-
ory,29 (3), 429–444. https://doi.org/10.1093/jopart/muy083
Andrews, L. (2019). Public administration, public leadership and the construction of public value
in the age of the algorithm and ‘big data’. Public Administration,97 (2), 296–310. https:
//doi.org/10.1111/padm.12534
Andrews, R., & Miller, K. J. (2013). Representative bureaucracy, gender, and policing: The case
of domestic violence arrests in England. Public Administration,91 (4), 998–1014. https :
//doi.org/10.1111/padm.12002
Aoki, N. (2020). An experimental study of public trust in AI chatbots in the public sector. Govern-
ment Information Quarterly,37 (4), 101490. https://doi.org/10.1016/j.giq.2020.101490
Araujo, T., Helberger, N., Kruikemeier, S., & de Vreese, C. H. (2020). In AI we trust? Perceptions
about automated decision-making by artificial intelligence. AI & Society,35 (3), 611–623.
https://doi.org/10.1007/s00146-019-00931-w
Ariely, G. (2013). Public administration and citizen satisfaction with democracy: Cross-national
evidence. International Review of Administrative Sciences,79 (4), 747–766. https://doi.org/
10.1177/0020852313501432
Aytaç, S. E. (2021). Do voters respond to relative economic performance? Public Opinion Quarterly,
84 (2), 493–507. https://doi.org/10.1093/poq/nfaa023
Baniamin, H. M., & Jamil, I. (2021). Effects of representative bureaucracy on perceived performance
and fairness: Experimental evidence from South Asia. Public Administration.https://doi.
org/10.1111/padm.12758
Bansal, G., Wu, T., Zhou, J., Fok, R., Nushi, B., Kamar, E., Ribeiro, M. T., & Weld, D. (2021).
Does the whole exceed its parts? The effect of AI explanations on complementary team
performance. Proceedings of the 2021 CHI Conference on Human Factors in Computing
Systems, 1–16. https://doi.org/10.1145/3411764.3445717
Barocas, S., & Selbst, A. D. (2016). Big data’s disparate impact. California Law Review,104 (3),
671–732. https://doi.org/10.15779/Z38BG31
Barth, T. J., & Arnold, E. (1999). Artificial intelligence and administrative discretion: Implications
for public administration. The American Review of Public Administration,29 (4), 332–351.
https://doi.org/10.1177/02750749922064463
Battaglio, R. P., Belardinelli, P., Bellé, N., & Cantarelli, P. (2019). Behavioral public administration
ad fontes: A synthesis of research on bounded rationality, cognitive biases, and nudging in
public organizations. Public Administration Review,79 (3), 304–320. https://doi.org/10 .
1111/puar.12994
Bishu, S. G., & Kennedy, A. R. (2020). Trends and gaps: A meta-review of representative bureau-
cracy. Review of Public Personnel Administration,40 (4), 559–588. https://doi.org/10.1177/
0734371X19830154
25

Block, K., Croft, A., De Souza, L., & Schmader, T. (2019). Do people care if men don’t care about
caring? The asymmetry in support for changing gender roles. Journal of Experimental Social
Psychology,83 (July 2018), 112–131. https://doi.org/10.1016/j.jesp.2019.03.013
Bovens, M., & Zouridis, S. (2002). From street-level to system-level bureaucracies: How information
and communication technology is transforming administrative discretion and constitutional
control. Public Administration Review,62 (2), 174–184. https :/ /doi .org / 10. 1111/ 0033-
3352.00168
Bradbury, M. D., & Kellough, J. E. (2007). Representative bureaucracy: Exploring the potential for
active representation in local government. Journal of Public Administration Research and
Theory,18 (4), 697–714. https://doi.org/10.1093/jopart/mum033
Braun, M., Bleher, H., & Hummel, P. (2021). A leap of faith: Is there a formula for “trustworthy”
AI? Hastings Center Report,51 (3), 17–22. https://doi.org/10.1002/hast.1207
Brundage, M., Avin, S., Clark, J., Toner, H., Eckersley, P., Garfinkel, B., Dafoe, A., Scharre, P.,
Zeitzoff, T., Filar, B., Anderson, H., Roff, H., Allen, G. C., Steinhardt, J., Flynn, C.,
HÉigeartaigh, S. Ó., Beard, S., Belfield, H., Farquhar, S., . . . Amodei, D. (2018). The ma-
licious use of artificial intelligence: Forecasting, prevention, and mitigation. arXiv:preprint.
http://arxiv.org/abs/1802.07228
Bughin, J., Hazan, E., Ramaswamy, S., Chui, M., Allas, T., Dahlström, P., Henke, N., & Trench,
M. (2017). Artificial intelligence: The next digital frontier? (Tech. rep.). McKinsey Global
Institute. https://www.mckinsey.com/$%5Csim$/media/mckinsey/industries/advanced%
20electronics/our%20insights/how%20artificial%20intelligence%20can%20deliver%20real%
20value%20to%20companies/mgi-artificial-intelligence-discussion-paper.ashx
Bullock, J. B. (2019). Artificial intelligence, discretion, and bureaucracy. American Review of Public
Administration,49 (7), 751–761. https://doi.org/10.1177/0275074019856123
Burton, J. W., Stein, M.-K., & Jensen, T. B. (2020). A systematic review of algorithm aversion
in augmented decision making. Journal of Behavioral Decision Making,33 (2), 220–239.
https://doi.org/10.1002/bdm.2155
Busuioc, M. (2021). Accountable artificial intelligence: Holding algorithms to account. Public Ad-
ministration Review,81 (5), 825–836. https://doi.org/10.1111/puar.13293
Chatterjee, S., Khorana, S., & Kizgin, H. (2022). Harnessing the potential of artificial intelligence to
foster citizens’ satisfaction: An empirical study on India. Government Information Quarterly,
39 (4), 101621. https://doi.org/10.1016/j.giq.2021.101621
Chingos, M. M. (2012). Citizen perceptions of government service quality: Evidence from public
schools. Quarterly Journal of Political Science,7(4), 411–445. https://doi.org/10.1561/100.
00011071
Cochran, J. C., & Warren, P. Y. (2012). Racial, ethnic, and gender differences in perceptions of
the police: The salience of officer race within the context of racial profiling. Journal of
Contemporary Criminal Justice,28 (2), 206–227. https://doi.org/10.1177/1043986211425726
Daugherty, P. R., Wilson, J., & Chowdhury, R. (2018). Using artificial intelligence to promote
diversity. MIT Sloan Management Review,60 (2). https:/ /sloanreview .mit .edu /article /
using-artificial-intelligence-to-promote-diversity/
Dermol, V., & Čater, T. (2013). The influence of training and training transfer factors on organi-
sational learning and performance. Personnel Review,42 (3), 324–348. https://doi.org/10.
1108/00483481311320435
Dietvorst, B. J., Simmons, J. P., & Massey, C. (2015). Algorithm aversion: People erroneously avoid
algorithms after seeing them err. Journal of Experimental Psychology: General,144 (1), 114–
126. https://doi.org/10.1037/xge0000033
26

Eren, O., & Mocan, N. (2018). Emotional judges and unlucky juveniles. American Economic Journal:
Applied Economics,10 (3), 171–205. https://doi.org/10.1257/app.20160390
Frederickson, H. G. (1990). Public administration and social equity. Public Administration Review,
50 (2), 228. https://doi.org/10.2307/976870
Gade, D. M., & Wilkins, V. M. (2013). Where did you serve? Veteran identity, representative
bureaucracy, and vocational rehabilitation. Journal of Public Administration Research and
Theory,23 (2), 267–288. https://doi.org/10.1093/jopart/mus030
Gantchev, V. (2019). Data protection in the age of welfare conditionality: Respect for basic rights
or a race to the bottom? European Journal of Social Security,21 (1), 3–22. https://doi.org/
10.1177/1388262719838109
Gaynor, M. (2020). Automation and AI sound similar, but may have vastly different impacts on
the future of work. Retrieved March 1, 2022, from https://www.brookings.edu/blog/the-
avenue/2020/01/29/automation-and-artificial-intelligence-sound-similar- but- may- have-
vastly-different-impacts-on-the-future-of-work/
Gil, O., Cortés-Cediel, M. E., & Cantador, I. (2019). Citizen participation and the rise of digital
media platforms in smart governance and smart cities. International Journal of E-Planning
Research,8(1), 19–34. https://doi.org/10.4018/IJEPR.2019010102
Glikson, E., & Woolley, A. W. (2020). Human trust in artificial intelligence: Review of empirical
research. Academy of Management Annals,14 (2), 627–660. https://doi.org/10.5465/annals.
2018.0057
Grimmelikhuijsen, S. G. (2010). Transparency of public decision-making: Towards trust in local
government? Policy & Internet,2(1), 4–34. https://doi.org/10.2202/1944-2866.1024
Hainmueller, J., Hopkins, D. J., & Yamamoto, T. (2014). Causal inference in conjoint analysis: Un-
derstanding multidimensional choices via stated preference experiments. Political Analysis,
22 (1), 1–30. https://doi.org/10.1093/pan/mpt024
Hasenfeld, Y., Rafferty, J. A., & Zald, M. N. (1987). The welfare state, citizenship, and bureaucratic
encounters. Annual Review of Sociology,13 (1), 387–415. https://doi.org/10.1146/annurev.
so.13.080187.002131
Hendler, J. (2020). Virtual roundtable on making government AI ready. Retrieved March 1, 2022,
from https : / / napawash . org / grand - challenges - blog / virtual - roundtable - on - making -
government-ai-ready
Hibbard, P. F., Blomgren Amsler, L., & Jackman, M. S. (2022). Representative bureaucracy and
organizational justice in mediation. Journal of Public Administration Research and Theory,
32 (4), 717–735. https://doi.org/10.1093/jopart/muab044
Hibbing, J. R., & Theiss-Morse, E. (Eds.). (2001). What is it about government that Americans
dislike? (1st ed.). Cambridge University Press.
Hilliard, L. J., & Liben, L. S. (2010). Differing levels of gender salience in preschool classrooms:
Effects on children’s gender attitudes and intergroup bias. Child Development,81 (6), 1787–
1798. https://doi.org/10.1111/j.1467-8624.2010.01510.x
Hong, S. (2017). Black in blue: Racial profiling and representative bureaucracy in policing revisited.
Journal of Public Administration Research and Theory,27 (4), 547–561. https://doi.org/10.
1093/jopart/mux012
Hurwitz, J., & Peffley, M. (2005). Explaining the Great Racial Divide: Perceptions of Fairness in
the U.S. Criminal Justice System. The Journal of Politics,67 (3), 762–783. https://doi.org/
10.1111/j.1468-2508.2005.00338.x
IBM Cloud Education. (2020). What is machine learning? Retrieved March 1, 2022, from https:
//www.ibm.com/cloud/learn/machine-learning
27

Im, T., Cho, W., Porumbescu, G., & Park, J. (2014). Internet, trust in government, and citizen
compliance. Journal of Public Administration Research and Theory,24 (3), 741–763. https:
//doi.org/10.1093/jopart/mus037
Ingrams, A., Kaufmann, W., & Jacobs, D. (2022). In AI we trust? Citizen perceptions of AI in
government decision making. Policy & Internet,14 (2), 390–409. https://doi.org/10.1002/
poi3.276
James, O., & Moseley, A. (2014). Does performance information about public services affect citizens’
perceptions, satisfaction, and voice behaviour? Field experiments with absolute and relative
performance information. Public Administration,92 (2), 493–511. https://doi.org/10.1111/
padm.12066
Janssen, M., Brous, P., Estevez, E., Barbosa, L. S., & Janowski, T. (2020). Data governance: Or-
ganizing data for trustworthy Artificial Intelligence. Government Information Quarterly,
37 (3), 101493. https://doi.org/10.1016/j.giq.2020.101493
Jilke, S., & Tummers, L. (2018). Which clients are deserving of help? A theoretical model and
experimental test. Journal of Public Administration Research and Theory,28 (2), 226–238.
https://doi.org/10.1093/jopart/muy002
Jordan, M. I., & Mitchell, T. M. (2015). Machine learning: Trends, perspectives, and prospects.
Science,349 (6245), 255–260. https://doi.org/10.1126/science.aaa8415
Keiser, L. R. (1999). State bureaucratic discretion and the administration of social welfare programs:
The case of social security disability. Journal of Public Administration Research and Theory,
9(1), 87–106. https://doi.org/10.1093/oxfordjournals.jpart.a024407
Krämer, N. C., Lucas, G., Schmitt, L., & Gratch, J. (2018). Social snacking with a virtual agent –
On the interrelation of need to belong and effects of social responsiveness when interacting
with artificial entities. International Journal of Human-Computer Studies,109, 112–121.
https://doi.org/10.1016/j.ijhcs.2017.09.001
Kuziemski, M., & Misuraca, G. (2020). AI governance in the public sector: Three tales from the
frontiers of automated decision-making in democratic settings. Telecommunications Policy,
44 (6), 101976. https://doi.org/10.1016/j.telpol.2020.101976
Lee, E., & Nicholson-Crotty, S. (2022). Symbolic representation, expectancy disconfirmation, and
citizen complaints against police. The American Review of Public Administration,52 (1),
36–45. https://doi.org/10.1177/02750740211034427
Lindgren, I., Madsen, C. Ø., Hofmann, S., & Melin, U. (2019). Close encounters of the digital kind: A
research agenda for the digitalization of public services. Government Information Quarterly,
36 (3), 427–436. https://doi.org/10.1016/j.giq.2019.03.002
Lipsky, M. (1980). Street-level bureaucrats as policy makers. Street-level bureaucracy, dilemmas of
the individual in public services (pp. 13–25). Russell Sage Foundation.
Liu, S. M., & Kim, Y. (2018). Special issue on internet plus government: New opportunities to solve
public problems? Government Information Quarterly,35 (1), 88–97. https: //doi. org/10.
1016/j.giq.2018.01.004
Logg, J. M., Minson, J. A., & Moore, D. A. (2019). Algorithm appreciation: People prefer algorithmic
to human judgment. Organizational Behavior and Human Decision Processes,151 (April
2018), 90–103. https://doi.org/10.1016/j.obhdp.2018.12.005
Mann, M. (2020). Technological politics of automated welfare surveillance: Social (and data) justice
through critical qualitative inquiry. Global Perspectives,1(1). https://doi.org/10.1525/gp.
2020.12991
Margetts, H., & Dorobantu, C. (2019). Rethink government with AI. Nature,568 (7751), 163–165.
https://doi.org/10.1038/d41586-019-01099-5
28

Martinho-Truswell, E. (2018). How AI could help the public sector. Harvard Business Review.https:
//hbr.org/2018/01/how-ai-could-help-the-public-sector
McFarland, M. (2015). Technology stripped a human touch from our lives. Can innovation bring it
back? https://www.washingtonpost.com/news/innovations/wp/2015/04/22/technology-
stripped-a-human-touch-from-our-lives-can-innovation-bring-it-back/
McIntosh, P. (1998). White privilege, color, and crime: A personal account. In C. R. Mann &
M. S. Zatz (Eds.), Images of color, images of crime: Readings. Roxbury Publishing Com-
pany. https:/ /cpt.org/wp- content/ uploads/Undoing20Racism20- 20White20Privilege20-
20McIntosh-11.pdf
Mcknight, D. H., Carter, M., Thatcher, J. B., & Clay, P. F. (2011). Trust in a specific technology.
ACM Transactions on Management Information Systems,2(2), 1–25. https://doi.org/10.
1145/1985347.1985353
McLaughlin, D. M., Mewhirter, J. M., Wright, J. E., & Feiock, R. (2021). The perceived effectiveness
of collaborative approaches to address domestic violence: The role of representation, ‘reverse-
representation,’ embeddedness, and resources. Public Management Review,23 (12), 1808–
1832. https://doi.org/10.1080/14719037.2020.1774200
Medaglia, R., Gil-Garcia, J. R., & Pardo, T. A. (2021). Artificial intelligence in government: Taking
stock and moving forward. Social Science Computer Review, 089443932110340. https://doi.
org/10.1177/08944393211034087
Mehr, H. (2017). Artificial intelligence for citizen services and government (tech. rep. August).
Harvard University. Cambridge, MA. https://ash.harvard.edu/files/ash/files/artificial_
intelligence_for_citizen_services.pdf
Meier, K. J. (1975). Representative bureaucracy: An empirical analysis. American Political Science
Review,69 (2), 526–542. https://doi.org/10.2307/1959084
Meier, K. J., & Nicholson-Crotty, J. (2006). Gender, representative bureaucracy, and law enforce-
ment: The case of sexual assault. Public Administration Review,66 (6), 850–860. https :
//doi.org/10.1111/j.1540-6210.2006.00653.x
Merriam-Webster. (2021a). Algorithms | Definition of Algorithms by Merriam-Webster. Retrieved
June 11, 2021, from https://www.merriam-webster.com/dictionary/algorithms
Merriam-Webster. (2021b). Artificial Intelligence | Definition of Artificial Intelligence by Merriam-
Webster. Retrieved June 11, 2021, from https://www.merriam- webster.com/dictionary/
artificial%20intelligence
Miller, S. M., & Keiser, L. R. (2021). Representative bureaucracy and attitudes toward automated
decision making. Journal of Public Administration Research and Theory,31 (1), 150–165.
https://doi.org/10.1093/jopart/muaa019
Moody, J. (2020). Algorithms for college admissions: What to know. https://www.usnews.com/
education /b est- colleges / articles/ how - admissions- algorithms - could - affect- your- college -
acceptance
Mosher, F. C. (1968). Democracy and the public service (2nd). Oxford University Press.
Nishant, R., Kennedy, M., & Corbett, J. (2020). Artificial intelligence for sustainability: Challenges,
opportunities, and a research agenda. International Journal of Information Management,
53, 102104. https://doi.org/10.1016/j.ijinfomgt.2020.102104
Ojo, A., Mellouli, S., & Ahmadi Zeleti, F. (2019). A realist perspective on AI-era public management.
Proceedings of the 20th Annual International Conference on Digital Government Research,
159–170. https://doi.org/10.1145/3325112.3325261
Pandey, S. K., & Bretschneider, S. I. (1997). The impact of red tape’s administrative delay on public
organizations’ interest in new information technologies. Journal of Public Administration
Research and Theory,7(1), 113–130. https://doi.org/10.1093/oxfordjournals.jpart.a024335
29

Pelau, C., Dabija, D.-C., & Ene, I. (2021). What makes an AI device human-like? The role of
interaction quality, empathy and perceived psychological anthropomorphic characteristics
in the acceptance of artificial intelligence in the service industry. Computers in Human
Behavior,122, 106855. https://doi.org/10.1016/j.chb.2021.106855
Pols, J., & Moser, I. (2009). Cold technologies versus warm care? On affective and social relations
with and through care technologies. Alter,3(2), 159–178. https://doi.org/10.1016/j.alter.
2009.01.003
Porumbescu, G. A., Neshkova, M. I., & Huntoon, M. (2019). The effects of police performance on
agency trustworthiness and citizen participation. Public Management Review,21 (2), 212–
237. https://doi.org/10.1080/14719037.2018.1473473
Porumbescu, G. A., Piotrowski, S. J., & Mabillard, V. (2021). Performance information, racial
bias, and citizen evaluations of government: Evidence from two studies. Journal of Public
Administration Research and Theory,31 (3), 523–541. https:/ /doi .org /10. 1093/ jopart/
muaa049
Prendergast, C. (2007). The motivation and bias of bureaucrats. American Economic Review,97 (1),
180–196. https://doi.org/10.1257/aer.97.1.180
Riccucci, N. M., & Van Ryzin, G. G. (2017). Representative bureaucracy: A lever to enhance social
equity, coproduction, and democracy. Public Administration Review,77 (1), 21–30. https:
//doi.org/10.1111/puar.12649
Riccucci, N. M., Van Ryzin, G. G., & Jackson, K. (2018). Representative bureaucracy, race, and
policing: A survey experiment. Journal of Public Administration Research and Theory,28 (4),
506–518. https://doi.org/10.1093/jopart/muy023
Riccucci, N. M., Van Ryzin, G. G., & Lavena, C. F. (2014). Representative bureaucracy in polic-
ing: Does it increase perceived legitimacy? Journal of Public Administration Research and
Theory,24 (3), 537–551. https://doi.org/10.1093/jopart/muu006
Riccucci, N. M., Van Ryzin, G. G., & Li, H. (2016). Representative bureaucracy and the willingness
to coproduce: An experimental study. Public Administration Review,76 (1), 121–130. https:
//doi.org/10.1111/puar.12401
Roberts, A. (2021). Who should we count as citizens? Categorizing people in public administration
research. Public Administration Review,81 (2), 286–290. https://doi.org/10.1111/puar.13270
Roch, C. H., Elsayed, M. A. A., & Edwards, J. (2018). Students’ and parents’ perceptions of dis-
ciplinary policy: Does symbolic representation matter? The American Review of Public Ad-
ministration,48 (4), 329–345. https://doi.org/10.1177/0275074016686420
Rubin, R. (2020). AI comes to the tax code. https://www.wsj.com/articles/ai-comes-to-the-tax-
code-11582713000
Scherer, N., & Curry, B. (2010). Does descriptive race representation enhance institutional legiti-
macy? The case of the U.S. courts. The Journal of Politics,72 (1), 90–104. https://doi.org/
10.1017/S0022381609990491
Sharma, G. D., Yadav, A., & Chopra, R. (2020). Artificial intelligence and effective governance:
A review, critique and research agenda. Sustainable Futures,2(November 2019), 100004.
https://doi.org/10.1016/j.sftr.2019.100004
Simon, H. A. (1947). Administrative behavior: A study of decision-making processes in administrative
organizations. Palgrave Macmillan.
Song, M., An, S.-H., & Meier, K. J. (2021). Quality standards, implementation autonomy, and citizen
satisfaction with public services: cross-national evidence. Public Management Review,23 (6),
906–928. https://doi.org/10.1080/14719037.2020.1730939
Stefanelli, A., & Lukac, M. (2020). Subjects, trials, and levels: Statistical power in conjoint experi-
ments.https://doi.org/10.31235/osf.io/spkcy
30

Theobald, N. A., & Haider-Markel, D. P. (2008). Race, bureaucracy, and symbolic representation:
Interactions between citizens and police. Journal of Public Administration Research and
Theory,19 (2), 409–426. https://doi.org/10.1093/jopart/mun006
Thiebes, S., Lins, S., & Sunyaev, A. (2021). Trustworthy artificial intelligence. Electronic Markets,
31 (2), 447–464. https://doi.org/10.1007/s12525-020-00441-4
Thurman, N., Moeller, J., Helberger, N., & Trilling, D. (2019). My friends, editors, algorithms,
and I: Examining audience attitudes to news selection. Digital Journalism,7(4), 447–469.
https://doi.org/10.1080/21670811.2018.1493936
Tong, C., & Wong, S. (2000). A predictive dynamic traffic assignment model in congested capacity-
constrained road networks. Transportation Research Part B: Methodological,34 (8), 625–644.
https://doi.org/10.1016/S0191-2615(99)00045-4
Toreini, E., Aitken, M., Coopamootoo, K., Elliott, K., Zelaya, C. G., & van Moorsel, A. (2020). The
relationship between trust in AI and trustworthy machine learning technologies. Proceedings
of the 2020 Conference on Fairness, Accountability, and Transparency, 272–283. https://
doi.org/10.1145/3351095.3372834
Van Dam, A. (2019). Algorithms were supposed to make Virginia judges fairer. What happened was
far more complicated. https://www.washingtonpost.com/business/2019/11/19/algorithms-
were- supposed -make- virginia-judges-more- fair- what- actually- happened- was- far- more-
complicated/
Van de Walle, S., & Bouckaert, G. (2003). Public service performance and trust in government:
The problem of causality. International Journal of Public Administration,26 (8-9), 891–913.
https://doi.org/10.1081/PAD-120019352
Van Ryzin, G. G. (2015). Service quality, administrative process, and citizens’ evaluation of local
government in the US. Public Management Review,17 (3), 425–442. https://doi.org/ 10.
1080/14719037.2013.841456
Van Ryzin, G. G., Riccucci, N. M., & Li, H. (2017). Representative bureaucracy and its symbolic
effect on citizens: A conceptual replication. Public Management Review,19 (9), 1365–1379.
https://doi.org/10.1080/14719037.2016.1195009
van Leeuwen, K. G., de Rooij, M., Schalekamp, S., van Ginneken, B., & Rutten, M. J. C. M.
(2022). How does artificial intelligence in radiology improve efficiency and health outcomes?
Pediatric Radiology,52 (11), 2087–2093. https://doi.org/10.1007/s00247-021-05114-8
Veale, M., & Brass, I. (2019). Administration by algorithm? Public management meets public sector
machine learning. In K. Yeung & M. Lodge (Eds.), Algorithmic regulation (pp. 121–149).
Oxford University Press.
Vogl, T. M., Seidelin, C., Ganesh, B., & Bright, J. (2020). Smart technology and the emergence of al-
gorithmic bureaucracy: Artificial intelligence in UK local authorities. Public Administration
Review,80 (6), 946–961. https://doi.org/10.1111/puar.13286
Wachter, S., Mittelstadt, B., & Russell, C. (2021). Why fairness cannot be automated: Bridging
the gap between EU non-discrimination law and AI. Computer Law & Security Review,41,
105567. https://doi.org/10.1016/j.clsr.2021.105567
Walsh, W. F. (2001). Compstat: An analysis of an emerging police managerial paradigm. Policing:
An International Journal of Police Strategies & Management,24 (3), 347–362. https://doi.
org/10.1108/13639510110401717
Welch, E. W., & Pandey, S. K. (2006). E-government and bureaucracy: Toward a better understand-
ing of intranet implementation and its effect on red tape. Journal of Public Administration
Research and Theory,17 (3), 379–404. https://doi.org/10.1093/jopart/mul013
31

Wiessner, D. (2020). USCIS sued over delays in issuing work permits to visa holders. Retrieved
April 22, 2021, from https:/ /www.reuters.com/article/immigration -workpermits/uscis -
sued-over-delays-in-issuing-work-permits-to-visa-holders-idUSL2N2EV20T
Wilson, H. J., & Daugherty, P. R. (2018). Collaborative intelligence: Humans and AI are joining
forces. Harvard Business Review,2018 (July-August), 114–123. https://hbr.org/2018/07/
collaborative-intelligence-humans-and-ai-are-joining-forces
Winokoor, C. (2021). ‘We want to be involved in the country’: How COVID has delayed the path to
citizenship. Retrieved April 22, 2021, from https://www.heraldnews.com/story/news/2021/
04/09/pandemic-has-caused-delay-immigrants-eager-become-u-s-citizens/7146198002/
Winter, N., Burleigh, T., Kennedy, R., & Clifford, S. (2019). A simplified protocol to screen out
VPS and international respondents using Qualtrics. SSRN Electronic Journal.https://doi.
org/10.2139/ssrn.3327274
Wirtz, B. W., Weyerer, J. C., & Geyer, C. (2019). Artificial intelligence and the public sector—Applications
and challenges. International Journal of Public Administration,42 (7), 596–615. https://doi.
org/10.1080/01900692.2018.1498103
Wirtz, B. W., Weyerer, J. C., & Sturm, B. J. (2020). The dark sides of artificial intelligence:
An integrated AI governance framework for public administration. International Journal of
Public Administration,43 (9), 818–829. https://doi.org/10.1080/01900692.2020.1749851
Wright, J. E., & Headley, A. M. (2020). Police use of force interactions: Is race relevant or gender
germane? The American Review of Public Administration,50 (8), 851–864. https://doi.org/
10.1177/0275074020919908
Yang, C., & Dobbie, W. (2020). Equal protection under algorithms: A new statistical and legal
framework. Michigan Law Review, (119.2), 291. https://doi.org/10.36644/mlr.119.2.equal
Young, M. M., Bullock, J. B., & Lecy, J. D. (2019). Artificial discretion as a tool of governance: A
framework for understanding the impact of artificial intelligence on public administration.
Perspectives on Public Management and Governance,2(4), 301–313. https://doi.org/10.
1093/ppmgov/gvz014
Zekić-Sušac, M., Mitrović, S., & Has, A. (2021). Machine learning based system for managing
energy efficiency of public sector as an approach towards smart cities. International Journal
of Information Management,58, 102074. https://doi.org/10.1016/j.ijinfomgt.2020.102074
Zewe, A. (2022). A technique to improve both fairness and accuracy in artificial intelligence.
Zhang, B., & Dafoe, A. (2020). U.S. public opinion on the governance of artificial intelligence.
Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 187–193. https :
//doi.org/10.1145/3375627.3375827
Zhang, H., Feinzig, S., Raisbeck, L., & Mccombe, I. (2019). The role of AI in mitigating bias
to enhance diversity and inclusion (tech. rep.). IBM Smarter Workforce Institute. https:
//www.ibm.com/downloads/cas/2DZELQ4O
Zou, J., & Schiebinger, L. (2018). AI can be sexist and racist — it’s time to make it fair. Nature,
559 (7714), 324–326. https://doi.org/10.1038/d41586-018-05707-8
Zuiderwijk, A., Chen, Y.-C., & Salem, F. (2021). Implications of the use of artificial intelligence
in public governance: A systematic literature review and a research agenda. Government
Information Quarterly,38 (3), 101577. https://doi.org/10.1016/j.giq.2021.101577
32