Allocating scarce resources between groups and individuals is a perennial challenge of social life (Hardin, 1968). Deciding who is worthy of university admission, a loan to start a new business, or even an organ transplant involves trade-offs among competing claims and values. Decision-makers charged with the task of allocating such scarce resources face a daunting challenge. Because resources are finite, allocation decisions will benefit some and disadvantage others, and outcomes will often be perceived as unfair by some affected parties (Camps et al., 2019). Historically, nearly all such decisions were made by humans. What reliably emerges from both research and the common experience is that human decision-makers are not always impartial and often show systematic biases in judgment (Simon, 1951; Tversky and Kahneman, 1973, 1974; Gilovich et al., 2002). Studies show that desirable attributes of decision processes, like consistency, integrity, and impartiality (Thibaut and Walker, 1975; Leventhal, 1980), can be easily derailed by ingroup biases (Brewer, 1979; Hughes, 2017), biases against the outgroup (Hebl et al., 2020), or simple preference for those who offer instrumental value to the decision-maker (Cornelis et al., 2013; Zhao et al., 2015; Hughes, 2017). Regardless of the domain (e.g., HR or marketing), decisions are perceived as fairer when they are made without biases or prejudices, and when they are based on accurate information (Colquitt et al., 2013, 2018; Matta et al., 2017). Emphasizing the importance of using fair procedures to make decisions has long been advocated as one of the most effective ways to counteract the many biases that can undermine the acceptance and legitimacy of allocation decisions (Cropanzano et al., 2001, 2007; Miller, 2001; Helberger et al., 2020).

Recent advances in Artificial Intelligence (AI) computer systems that can sense, reason, and respond to their environment in real-time, often with human-like intelligence (Robert et al., 2020), have made many optimistic that AI will soon eliminate human biases and overcome the limitations that often lead to injustice and suboptimal allocation decisions (Ghahramani, 2015; Silver et al., 2017; Singh et al., 2017; Mozer et al., 2019; Glikson and Woolley, 2020). Indeed, forecasts predict that decision-makers will increasingly turn to AI when allocating scarce resources in domains such as business, law, and healthcare (Brynjolfsson and Mitchell, 2017; Fountaine et al., 2019; Frank et al., 2019; Rawson et al., 2019). The use of algorithms that rely on big data holds the promise of debiasing decision-making procedures by removing human subjectivity typically inherent in judging and comparing individuals (Newman et al., 2020). For example, much work has focused on utilizing AI to detect bribery and other forms of corruption to eliminate impartiality violations in governmental and other organizational contexts (Köbis and Mossink, 2021).

Despite their apparent advantages as decision-making tools, people’s trust in AI often lags behind its rising capabilities, and many are averse to turning over allocation decisions to non-human entities (Glikson and Woolley, 2020). Their aversion has been traced to a belief that machines do not possess a complete mind and, therefore, cannot freely choose actions, nor can they adequately reflect on their consequences (Johnson, 2015; Bigman and Gray, 2018; Bigman et al., 2019). Furthermore, Newman et al. (2020) found that people perceive algorithmically-driven decisions as less fair because of AI’s inability to consider and contextualize qualitative information. Castelo et al. (2019) also investigated people’s aversion to relying on algorithms to perform tasks previously done by humans and found that algorithms are trusted and relied on more for tasks that require cognitive abilities and rationality vs. tasks that depend more on emotional intelligence or intuition.

However, whether laypeople perceive AI as more impartial than human deciders has not been explicitly addressed. The aim of the present study is to bridge research on procedural justice (Leventhal, 1980) with algorithm aversion to explain how laypeople’s impartiality perceptions between AI and human decision-makers differ (Study 1), and whether impartiality violations shift people’s preferences for AI in allocation decisions (Study 2 and 3).

Philosophers and scholars have offered several perspectives on why just procedures matter (Rawls, 1971/1999; Leventhal, 1976, 1980; Lind and Lissak, 1985; Tyler et al., 1985; Sheppard and Lewicki, 1987; Tyler, 1988; Sheppard et al., 1992). In organizational and legal contexts, procedural justice is concerned with people’s fairness perceptions regarding the processes or rules applied throughout the decision-making process (Thibaut and Walker, 1975). When it comes to the allocation of scarce resources like getting a job or securing a loan, procedures will be seen as fairer when they are impartial, i.e., if procedures are applied consistently and without biases; are based on accurate information; are correctable and ethical, and are representative of relevant parties involved in the decision (Leventhal, 1980). Impartiality means that when people are making moral decisions (e.g., allocating scarce goods and resources) they should not give any special treatment to themselves, or to members of their own ingroup, and instead take a neutral and unbiased position (Cottingham, 1983).

Decades of organizational justice research show that impartiality perceptions are positively associated with cooperation, performance, or job satisfaction whilst reducing potentially damaging behaviors and attitudes such as retaliation, complaints, or negative word-of-mouth (Cohen-Charash and Spector, 2001; Colquitt et al., 2013; Colquitt and Zipay, 2015). Furthermore, knowing that procedures are impartial can make people more accepting of authorities, laws, and policies, even when the outcomes are disadvantageous to them (Tyler, 1994; Sunshine and Tyler, 2003).

One of the greatest hopes regarding algorithmically-driven decisions in organizational contexts lies in AI’s ability to suppress or even eliminate common human biases that threaten the enactment of fair procedures (Graso et al., 2019). AI has the potential for standardizing decision-making processes, thereby eliminating many of the idiosyncrasies that can lead human decision-makers to depart from impartiality (Grace et al., 2018; Raisch and Krakowski, 2020). Giubilini and Savulescu (2018), for example, argue that AI could serve as an “artificial moral advisor” because of its ability “to take into account the human agent’s own principles and values” whilst making consistent judgments without human biases and prejudice. In principle, using AI in allocation decisions should thus increase impartiality by “standardizing decision procedures and reducing potential conflicts through highly consistent and supposedly unbiased decisions” (Ötting and Maier, 2018), which directly correspond to tenets of fair procedures (Leventhal, 1980).

However, empirical evidence regarding laypeople’s perceptions of impartiality in algorithmically-driven decisions is limited, and findings in adjacent domains are ambiguous (for an overview, see Castelo et al., 2019; Glikson and Woolley, 2020). For example, Newman and colleagues find that algorithm-driven (vs. human-driven) hiring and lay-off decisions are perceived as less fair because people view algorithms as reductionist and unable to contextualize information (Newman et al., 2020). Yet, the authors do not test for perceptional differences between human and AI deciders. Furthermore, Ötting and Maier (2018) found that procedural justice had a positive impact on employee behaviors and attitudes, irrespective of whether the decider was a human, a robot, or a computer system. One limitation of their study was that the authors manipulated procedural justice (fair vs. unfair) and did not explicitly measure people’s baseline perceptions regarding human vs. AI deciders. Other evidence suggests that people might associate greater impartiality with AI-based decision procedures. For example, people generally perceive robots and artificial intelligence systems as consistent and reliable (Dzindolet et al., 2003), and they are more likely to rely on algorithmic advice than on human advice, particularly when their expertise is limited (Logg et al., 2019). In comparison, there is ample evidence to suggest that human decisions are often biased, for example, by prejudice (Hebl et al., 2020) or favoritism toward people who are close to them (Hughes, 2017). It is, therefore, reasonable to posit that people are more likely to attribute greater impartiality to AI than to a human decision-maker because they will view the former as having more of the attributes that characterize an unbiased decision-maker. Formally, we thus argue that

We are not assuming that people will believe that AI is entirely impartial, only that they will be viewed as more capable of approaching this standard than humans. If so, based on considerable evidence suggesting that people value impartiality and bias-suppression in decision-makers (Leventhal, 1980), people should prefer an AI over human-decision makers in contexts in which impartiality by human deciders is potentially jeopardized. That is because implementing AI holds the potential to remove subjective (and potentially biased) judgments from allocation decisions and instead make those decisions on more objective or quantifiable competence criteria. It seems possible, therefore, that laypeople who may be subjected to biased evaluations, especially when these judgments are based on prejudice or stereotypes, should prefer AI decision-makers over humans. Formally, we posit that

In an exploratory study, we first test whether people perceive allocation decisions that involve AI as more impartial compared to procedures that are led by humans (Study 1). Next, we test whether laypeople’s preference for AI (vs. human) decision-makers shifts if they are prompted to think that a human decision-maker might be partial (Study 2 and 3). Our studies comply with ethical regulations regarding human research participants, and we received ethical approval from the Human Research Ethics Committee at a major European university. We obtained informed consent from all participants. We informed them that participation was voluntary and that they could stop their participation at any time. We recruited all participants from Prolific Academic (Peer et al., 2017). All our studies contain measures of gender and age. Demographic information and sample sizes for each study are presented in Table 1. Except for Study 1, all studies were pre-registered.

All measures, manipulations, and exclusions are disclosed in the “Materials and Methods” section and Supplementary Material. We determined sample sizes by assuming a medium effect (Cohen’s d = 0.50), and we conducted a power analysis to calculate the required number of participants per condition to obtain a power of 0.95. Data collection was stopped once the pre-determined sample size was reached. All studies included attention or comprehension checks, which resulted in the exclusion of participants who failed those checks. The complete stimulus material and data can be publicly accessed in Supplementary Material. Test statistics presented in this research are all two-sided.

Study 1 showed that people perceive AI (vs. human) decision-makers to be more impartial. Based on research we reviewed that shows that people value impartial and unbiased decision-makers in allocation decisions, laypeople’s preferences for AI decision-makers should thus shift when they believe that a human decision-maker might be partial. We test this possibility in Study 2 by examining people’s preference for an AI over a human when the human is likely to be biased in favor of them, against them, or when they know the human is biased but are uncertain of the direction. We expected people to prefer AI (vs. human) decision-makers when they know that a person is negatively (vs. positively) biased against them.

In our final study, we examined whether people’s preferences for an AI (vs. human) might reverse when they believe that a human decision-maker might be partial against them because of their social status within their profession. We selected respect as a form of one’s social resource (Foa and Foa, 1980; Brown et al., 2020) which may make a human more partial than AI. The study thus aimed to replicate and advance findings from Study 2.

Technological advances are expected to result in increased use of AI in decisions that distribute scarce goods and resources between competing parties. The transition toward AI-led decision-making raises important moral and ethical questions for businesses, many of which concern algorithmic fairness and transparency (Rahwan et al., 2019; Raisch and Krakowski, 2020). But despite the critical importance of these issues, we still have a limited understanding of how people perceive allocation decisions in which human deciders are replaced by artificial forms of intelligence. While a growing body of work has explored the cognitive-affective factors behind people’s algorithm aversion (Khalil, 1993; Glikson and Woolley, 2020; Raisch and Krakowski, 2020; Robert et al., 2020), few studies have investigated the role of impartiality in human-AI interactions (Leventhal, 1980). Our study addresses this paucity and contributes to the literature by highlighting an important boundary condition to laypeople’s algorithm aversion.

First, our study sheds new light on the role of impartiality in human-AI interactions. Since AI is not freighted with some of the characteristics that can lead humans to stray from impartiality, it holds the promise of enhancing the accuracy, consistency, and incorruptibility from the social influence of many decision procedures. Importantly, impartiality has been shown to influence the acceptance and legitimacy of decision procedures (Tyler et al., 1985). People consistently value impartiality and prefer deciders who make allocation decisions without biases, prejudices, or previously determined personal preferences (Miller, 2001; Shaw and Olson, 2014). While we provide further evidence that people prefer humans over AI in decisions concerning them (Study 1), we consistently show that laypeople associate greater impartiality with AI (vs. human) decision-makers. This finding is noteworthy because technologists, scholars, and policy-makers have often raised concerns regarding the prevalence of partiality in AI that stems from historically biased data and poorly designed algorithms (Khalil, 1993; De Cremer, 2020; Robert et al., 2020). Our findings suggest that laypeople perceive AI as more capable of achieving impartiality than humans (Helberger et al., 2020). This is not to say that laypeople believe that AI is completely unbiased—it merely suggests that despite the many flaws and limitations within algorithmic decision making, laypeople still perceive AI to be less biased than humans.

Secondly, our findings show that impartiality concerns constitute an important boundary condition to people’s algorithmic preferences (Study 2 and 3). We show that when laypeople are concerned about the negative biases of human deciders, their preferences shift toward AI decision-makers. This is because they emphasize impartiality and bias suppression, which AI is perceived to be more capable of than human decision-makers. In other words, people who are potentially subjected to negative biases show greater preferences for AI deciders because it increases impartiality and removes biases, which might curb their chances of obtaining desired outcomes (e.g., securing a job or getting admitted to a university). The only exception is when human decision-makers are partial in people’s favor, in which case most people prefer the human over an AI. For a person who potentially benefits from a partial decider, choosing an AI to make decisions about them might even be self-destructive. This finding also has important implications for AI ethics. While previous studies have suggested that designing policies and regulations that continue to build trust in AI is likely to enhance further the acceptance and legitimacy of AI-led decisions (Khalil, 1993; Glikson and Woolley, 2020), we have identified an important caveat to this goal. Namely, despite its positive features and potential for standardizing and debiasing decision procedures (Zou and Schiebinger, 2018), people might not actually wish to endorse AI if they believe that partial decision-makers will help them to attain desired outcomes. Therefore, AI’s capabilities are likely to be valued more by those who experience negative evaluations or even prejudice in intra-human interactions.

In our studies, we only assessed prejudice (Study 2) and respect based on one’s status (Study 3) as examples of partiality that may lead people to endorse AI decision-makers. Future research could examine whether other impartiality violations might lead people to support AI against human-human interactions. Future research could also investigate the moderating role of related constructs like power (Fast and Schroeder, 2020). For example, in some instances, an AI might be perceived as less of a threat to one’s position in the organizational hierarchy. Indeed, research has shown that people prefer to be replaced by robots (vs. humans) when their job loss is at stake (Granulo et al., 2019).

Our study also offers managerial implications. Despite significant value being placed on justice, fair procedures are still frequently ignored, and decision-makers routinely deviate from principles of impartiality that they often claim to value (Graso et al., 2019). When this is the case, implementing AI-led decision procedures might provide some way to improve the accuracy, consistency, and impartiality of organizational decision procedures. However, such changes might also be met by resistance from employees and other stakeholders who have little to gain from such changes. While people who are concerned with being evaluated negatively are more likely to endorse such changes, people who benefit from partiality in the organizations are more likely to resist handing over decisions to non-human entities. Indeed, our study suggests (and future research can attest) how social resource-rich groups may be less invested in endorsing impartial decision-making tools such as AI. Future research should further explore how resistance to AI decision-makers, specifically among powerful individuals, can be overcome via algorithmic design or supporting procedures and policies.

In summary, our set of pre-registered studies conducted across diverse and complementary contexts has notable strengths. It builds on existing findings and advances our understanding of people’s perceptions of AI vs. human decision-makers. Furthermore, we have identified perceived impartiality as an important boundary condition to people’s algorithmic preferences in allocation decisions. Our simple design involved repeatedly contrasting the two decision-makers, and it allowed us to identify people’s underlying and reflexive assumptions regarding the impartiality of AI as decision-makers, and how they influence preferences for AI in different contexts. Nonetheless, our study has caveats and limitations, which we discuss with the hope of encouraging future research.

First, we only assessed people’s perceptions of AI systems that do not influence participants directly. Throughout our studies, we asked participants to assume that these decisions affect them. Still, as our participants were online platform users, there were no real consequences to their strongly endorsing AI or human decision-makers. Scarce research resources permitting, we recommend that future studies assess whether impartiality influences people’s preference for human or AI decision-makers when they themselves are invested in the outcome in question. An example would be giving employees in a large company an option to choose an AI or a human decision-maker when assessing promotions, raises, or bonuses. Furthermore, emerging research shows that in real-life scenarios, people might fail to reliably detect differences between algorithmically-generated and human-generated content, and that stated preferences might therefore diverge from actual behavior (Köbis and Mossink, 2021).

Second, while this is not explicitly tested in this study, it is possible that the use of AI in allocation decisions would have stronger support among people who believe that they might be disadvantaged or who experience prejudice because of their race, gender, or sexual orientation (Zou and Schiebinger, 2018). On the other hand, this support might be tempered by concerns that a more impartial decision-maker might not produce the outcomes they desire. We leave it to future research to investigate this possibility.

Finally, we focused on people’s perceptions of the human vs. AI decision-makers, but we did not examine attitudes toward the calibrators behind the AI. Any algorithm-based system is only as good as its inputs, and those inputs are only as good as the person calibrating the system. Perhaps people’s trust in AI as an impartial entity is inextricably linked with their trust in the AI’s calibrator. The appeal of certain AI systems is that they are self-correcting and capable of learning (Silver and Silver, 2017; Silver et al., 2017) which should presumably increase people’s perceptions of AI as distinct entities with minds of their own (Bigman and Gray, 2018; Bigman et al., 2019). Alternatively, we may witness a reality in which AI will remain simple reflections of their human masters and their calibrating powers, incapable of ever achieving true impartiality.

The original contributions presented in this study are included in the article/Supplementary Material, further inquiries can be directed to the corresponding author/s.

The studies involving human participants were reviewed and approved by the UCD Office of Research Ethics, University College Dublin, Belfield, Dublin (No. HS-E-19-102-CLAUDY). The patients/participants provided their written informed consent to participate in this study.

MC co-designed and ran the studies, wrote up the results, and wrote parts of introduction and discussion. KA provided overall guidance regarding the direction of the study, co-designed studies, and wrote parts of introduction and discussion. MG provided input on theoretical framing and study design and wrote parts of introduction and discussion. All authors contributed to the article and approved the submitted version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.