Automation and decision support for personnel selection
Automated support systems for high-level cognitive tasks are an emerging topic in research and practice, especially within personnel selection. For instance, Campion et al. (2016) used an automated system to emulate human ratings of applicant motivational letters. Such a system processes applicant information and could provide hiring managers with rankings of candidates to aid during hiring processes. Other studies have investigated the use of automated job interviews (Naim et al., 2018; Suen, Chen, & Lu, 2019). Those studies argue that it might be possible to automatically process interviews to predict interview performance which would make it feasible to screen thousands of applicants and present an evaluation of the most suitable applicants to hiring managers (Langer, König, & Hemsing, 2020).
The emerging use of such systems especially for personnel selection might be due to the complexity of personnel selection. Hiring managers gather and integrate a large variety of information (e.g., results from cognitive ability tests, interviews) from a potentially large number of applicants. Additionally, they screen the most suitable applicants, decide which applicants to hire, consider a variety of organizational goals (e.g., cost, diversity), and simultaneously need to adhere to legal regulations (e.g., regarding adverse impact) (König, Klehe, Berchtold, & Kleinmann, 2010). Automated decision support systems might help with gathering and combining of information for large numbers of applicants, thus making selection procedures more efficient, and there is hope that they can attenuate adverse impact (Campion et al., 2016; see, however, Raghavan, Barocas, Kleinberg, & Levy, 2020 showing challenges of automated personnel selection regarding adverse impact).
To get a clearer understanding of automated systems for personnel selection, we consider research in automation and on decision support systems. Following Sheridan and Parasuraman (2005, p. 89), automation refers to “a) the mechanization and integration of the sensing of environmental variables (by artificial sensors), b) data processing and decision making (by computers), c) mechanical action (by motors or devices that apply forces on the environment), d) and/or ‘information action’ by communication of processed information to people.” We focus on the data processing and decision-making aspects of automation followed by information action where systems provide processed information to human decision-makers. This kind of automation is reflected by four broad classes of functions of automation: gathering of information, filtering and analysis of information, decision recommendations, and action implementation (Parasuraman, Sheridan, & Wickens, 2000; Sheridan, 2000). With more of these functions realized within systems, the extent of automation increases. For instance, gathering of information could imply automatic transcription of video interviews (Langer, König, & Krause, 2017). Filtering and analysis of information could mean highlighting keywords within these transcripts. Systems that provide decision recommendations could be trained on previous data in order to fulfill classification (e.g., distinguishing suitable vs. non-suitable applicants) or prediction tasks (e.g., what applicants will most likely be high-performing employees). The outputs of those tasks can then be presented to human decision-makers. This is precisely the kind of automated system that the current paper is referring to: automated systems that gather and analyze information as well as derive inferences and recommendations based on these steps. Such systems can aid hiring managers in a large range of their duties (e.g., gathering and integrating of information, applicant screening), and their outputs can be used as an additional (or alternative) source of information for decision-makers.
Further research on automated information processing and decision support systems has its roots in the literature on mechanical versus clinical gathering and combination of information (Kuncel, Klieger, Connelly, & Ones, 2013; Meehl, 1954). In line with this research, automated processes from the era of machine learning and artificial intelligence can be perceived as a more complex and sophisticated way of mechanical gathering or combination of data. Specifically, some of the novel approaches use sensors (e.g., cameras, microphones) to extract information from interviews, use natural language processing to extract the content of applicant responses, and apply machine learning methods to evaluate applicant performance (how this novel kind of mechanical information gathering and combination compares to traditional mechanical approaches and to clinical approaches regarding validity is an open question; Gonzalez, Capman, Oswald, Theys, & Tomczak, 2019). Research on mechanical combination of information can provide insights into potential effects of using modern information processing and decision support automation (see, e.g., Gonzalez et al., 2019; Longoni, Bonezzi, & Morewedge, 2019). On the one hand, it has shown that mechanical gathering and combining of information (e.g., combination using ordinary least squares regression) can improve decision quality when compared to clinical gathering and combining of information (e.g., intuition-based combination of information) (Kuncel et al., 2013). On the other hand, this literature found that people are skeptical of the use of mechanical gathering and combining of information (Burton et al., 2019). Some researchers even refer to the reluctance to use mechanically combined information as algorithm aversion (Dietvorst et al., 2015; but cf. Logg et al., 2019) and indicate that using automated systems within jobs might affect people’s behavior and reactions to the job (Burton et al., 2019).
Information processing and decision support systems and their effect on performance and efficiency
There are several ways in which automated information processing and decision support systems for personnel selection can be implemented. Beyond the specific tasks allocated to an automated system (e.g., information gathering, analysis of information), the timing of when to provide hiring managers with the outputs of a system is a crucial implementation choice (Silverman, 1992; van Dongen & van Maanen, 2013). Two of the main points in time to integrate decision support systems are systems that provide their outputs (e.g., rankings of applicants) before a human decision-maker processes available information (support-before-processing systems) and systems that provide their outputs after an initial processing of information (support-after-processing systems) (Endsley, 2017; Guerlain et al., 1999; Silverman, 1992). For instance, automated personnel selection systems can come with automated support-before-processing (see Raghavan et al., 2020 who provides an overview on providers of automated personnel selection solutions). In general, the respective system analyzes applicant data and derives outputs (e.g., evaluations of performance, personality; applicant rankings) (Langer et al., 2019). Hiring managers receive these outputs together with further applicant information. Thus, they receive the output of the automated system and can have an additional look into further applicant information. This means decision-makers could decide to fully rely on the recommendation provided by the system, but they can also use it as an additional source of information to integrate together with further applicant information to reach a decision (Kuncel, 2018). Potential advantages of these systems are that they can increase efficiency (Onnasch, Wickens, Li, & Manzey, 2014) and serve as a source of mechanically combined information to enhance decision quality (Kuncel et al., 2013)—given an adequate validation of the system. Potential disadvantages are that they can induce anchoring effects so hiring managers might only give attention to best-ranked applicants (Endsley, 2017). Additionally, research from classical areas of automation indicates that people tend to initially perceive such systems as highly reliable which can lead decision-makers to follow recommendations without consulting additional, potentially contradictory information, and without thoroughly reflecting on applicants’ suitability (i.e., they might overtrust the system; Lee & See, 2004). Finally, for support-before-processing systems, people can feel “reduced to the role of […] recipient of the machine’s solution” (Silverman, 1992, p. 111). This feeling can diminish user acceptance and might be one reason for perceived loss of reputation when using such systems because decision-makers perceive they have less opportunity to show their expertise (Arkes, Shaffer, & Medow, 2007; Nolan et al., 2016).
Partly due to the latter issues, support-after-processing systems (Endsley, 2017; Guerlain et al., 1999) evolved as an alternative to support-before-processing systems. They would also process applicant information and provide an evaluation of applicants. However, these systems serve as an additional source of mechanically combined information that decision-makers can use after they have processed available information (Silverman, 1992). Additionally, such systems can be designed to provide feedback or criticize human decisions (Sharit, 2003; Silverman, 1992). Rather than proving the correctness of human decision, those systems serve as an opportunity to reflect on an initial decision and as an additional point of view on the decision. Up to date, research on those systems comes primarily from medical research and practice (Longoni et al., 2019; Silverman, 1992). For instance, in cancer diagnosis, a physician would first analyze available data (e.g., MRI images; Langlotz et al., 2019) and come up with a diagnosis (or a therapy plan; Langlotz & Shortliffe, 1984). The respective systems could use this diagnosis as an input and either provide the physician with the diagnosis it would have given or with information regarding what part of a diagnosis seems to be inconsistent with available data (Guerlain et al., 1999). However, because there is already evidence that issues with support-before-processing systems (e.g., perceived loss of reputation) might translate to managerial tasks (Nolan et al., 2016), the use of support-after-processing systems will likely not remain restricted to medical decision-making. Optimally, they encourage more thorough information processing and finding better rationales for decisions (Guerlain et al., 1999). For instance, such systems in personnel selection might counterbalance human heuristics by making decision-makers aware of overlooked or hastily rejected candidates (Derous, Buijsrogge, Roulin, & Duyck, 2016; Raghavan et al., 2020). Further potential advantages are that those systems, similar to support-before-processing systems, serve as an additional mechanical source of information combination, thus potentially enhancing decision quality (Guerlain et al., 1999). Furthermore, and in contrast to support-before-processing systems, there are no initial anchoring issues when using support-after-processing systems (Endsley, 2017).
One disadvantage of support-after-processing systems is that they do not increase efficiency of decision-making as decision-makers still need to process initially available information. In contrast, they can even increase the time necessary to reach decisions as decision-makers are encouraged to consider additional information and alternative perspectives on available information (Endsley, 2017). Furthermore, over-trusting effects cannot be ruled out. However, instead of solely relying on the recommendation by an automated system (as could be the case for support-before-processing systems), decision-makers themselves would have processed and combined different sources of information. This could give them a better rationale whether or not they want to follow the system’s recommendations (Sharit, 2003). Given the proposed advantages and disadvantages of the different decision support systems, we propose (see Table 1 for an overview on contrasts for the hypotheses):
Hypothesis 1Footnote 1: Participants who receive support (by a support-before-processing system or a support-after-processing system) will show better performance in the tasks than participants who receive no support (i.e., the no-support group).
Hypothesis 2Footnote 2: Participants in the support-before-processing group will complete the decision in less time than the other groups.
Effects on knowledge and task characteristics
In a review on the topic of algorithm aversion, Burton et al. (2019) argue that using automated support systems might have a variety of effects on decision-makers. For instance, decision-makers might feel reduced autonomy when receiving decision support, expectations towards systems might not be met by respective systems (Dietvorst et al., 2015; Highhouse, 2008), and certain design options within such systems (e.g., how and when to present a recommendation) might not match to human information processing, thus contributing to reluctance to use such systems.
This implies that using such systems for managerial work may affect knowledge and task characteristics (e.g., during information processing and decision-making tasks), and we propose that work design research (Hackman & Oldham, 1976; Morgeson & Humphrey, 2006; Parker & Grote, 2020) can help to understand these implications. Specifically, the integrated work design framework (Morgeson et al., 2012) proposes a variety of knowledge and task characteristics that affect important attitudinal, behavioral, cognitive, and well-being outcomes. Knowledge characteristics refer to the demands (e.g., cognitive demands) that people experience while fulfilling tasks (Morgeson & Humphrey, 2006). Task characteristics relate to the tasks that have to be accomplished for a particular job and how people accomplish these tasks.
In relation to decision support systems, the knowledge characteristic information processing and the task characteristics autonomy, task identity, and feedback from the job are especially important. The amount of information processing reflects the degree to which a job requires processing, integration, and analysis of information (Morgeson & Humphrey, 2006). Autonomy indicates the degree to which employees can fulfill tasks independently and can decide how to approach tasks (Hackman & Oldham, 1976; Morgeson et al., 2012). Task identity describes the degree to which tasks can be fulfilled from the beginning to the end versus only working on specific parts of the task (Hackman & Oldham, 1976; Morgeson et al., 2012). Finally, feedback from the job is defined as the degree to which employees receive information about their performance in the job from aspects of the job itself (Hackman & Oldham, 1976; Morgeson et al., 2012). Variations in these characteristics affect psychological states such as experienced meaningfulness, perceived responsibility for work outcomes, as well as work satisfaction and performance (Chung-Yan, 2010; Hackman & Oldham, 1976; Humphrey et al., 2007). While Humphrey et al. (2007) emphasize the lack of empirical research regarding knowledge characteristics, previous research shows that, in general, more demanding information processing requirements, a higher level of autonomy, task identity, and feedback from the job relate to more positive outcomes (Morgeson et al., 2012; but see Chung-Yan (2010) and Warr (1994), indicating that “the more the better” is not necessarily true for all characteristics).
Information processing and decision support systems for personnel selection might affect those knowledge and task characteristics. Support-before-processing systems show their assessment of applicants before hiring managers process applicant data. This might reduce information processing requirements (e.g., integrate information, compare applicants). Regarding task identity, hiring managers might feel that they do not really complete the entire selection task. Moreover, they might perceive a loss of autonomy as the system already implies which applicants to favor (Burton et al., 2019; Langlotz & Shortliffe, 1984; Nolan et al., 2016). All of this would be in line with findings and speculations of previous research that indicated that hiring managers tend to favor selection methods where they can show their expertise and those that allow for more autonomy (e.g., unstructured vs. structured interviews; using clinical vs. mechanical combination of information) (Burton et al., 2019; Highhouse, 2008).
In contrast, support-after-processing systems might require more information processing, allow for more perceived task identity, and grant a higher level of autonomy. They allow hiring managers to independently analyze and integrate information, reach an initial decision about applicants, and only then provide them with additional information or novel perspectives for their final decision. Furthermore, hiring managers might use recommendations from support-after-processing systems as feedback on task performance. Specifically, when hiring managers have no prior experience with an automated system, they may believe that the system is working as intended (Madhavan & Wiegmann, 2007). The purpose of an automated personnel selection system is to evaluate applicants’ job fit, so if people initially believe that an automated system is able to do this, they might compare their own decisions to the recommendations by the system in order to get an idea about their own performance.
Following, we will build hypotheses that argue for how the aforementioned variations in knowledge and task characteristics affect five important psychological factors at work: enjoyment, monotony, satisfaction with the decision, perceived responsibility, and self-efficacy.
Effects on enjoyment and monotony
Enjoyment and monotony of a task are important for well-being at work (Taber & Alliger, 1995). Enjoyment of a task is present if people are happy doing a task, whereas monotony indicates that people perceive a task to be repetitive and boring (Smith, 1981). The proposed differences in information processing requirements, task identity, and perceptions of feedback might affect enjoyment and experienced monotony with the task (Smith, 1981). In the case of a support-before-processing system, the task might appear less cognitively demanding, and decision-makers might think that most of the task was already fulfilled by the system. In contrast, using support-after-processing systems upholds information processing requirements, and task identity. Additionally, if people perceive that they are provided with evidence on how well they performed, this can increase enjoyment within a task (Sansone, 1989). Thus, we propose:
Hypothesis 3a: Participants in the support-before-processing system group will perceive less enjoyment and more monotony with the task than the other groups.
Hypothesis 3b: Participants in the support-after-processing system group will perceive more enjoyment and less monotony with the task than the no-support group.
Effects on satisfaction with the decision
For knowledge workers (e.g., hiring managers), a large share of daily work consists of information processing and decision-making (Derous et al., 2016). Satisfaction with decisions thus likely contributes to workers’ overall job satisfaction (Taber & Alliger, 1995). Satisfaction with decisions is especially important in situations where the consequences and the quality of a decision do not immediately become apparent (Houston, Sherman, & Baker, 1991; Sainfort & Booske, 2000). This is the case for personnel selection, where the quality of the decision is determined by an applicant’s future job performance (Robertson & Smith, 2001). When the consequences of decisions will only become visible in the long term, satisfaction with decisions might arise if people are convinced by their decision and satisfied with the decision-making process (Brehaut et al., 2003; Sainfort & Booske, 2000). Especially when people autonomously process information, they might become more convinced by their decision and thus more satisfied with it (Sainfort & Booske, 2000). Additionally, if people believe they have made a good decision (i.e., if they get any kind of perceived feedback on their performance), this might increase satisfaction with the decision (Sansone, 1989). To be clear, information processing requirements, task identity, task autonomy, and feedback might be more pronounced in the case of support-after-processing systems; thus, we propose:
Effects on perceived responsibility
Research from different areas of decision support shows that using standardized or computerized processes can reduce users’ perceived responsibility (Arkes et al., 2007; Lowe, Reckers, & Whitecotton, 2002; Nolan et al., 2016). According to Nolan et al. (2016), perceived responsibility in personnel selection may decrease because decision-makers believe that task fulfillment cannot be fully attributed to themselves. Instead, some of the credit goes to the automated or standardized support process. Aligning with this, work design research assumes that reduced autonomy leads to less perceived responsibility (Hackman & Oldham, 1976). Furthermore, Langer et al. (2019) speculated about situations where hiring managers who are supported by a system follow the advice of the system. Hiring managers might perceive that the decision was already predetermined, so even if they were to thoroughly evaluate all available information, by following the system’s recommendation, their perceived responsibility could reduce, especially for those using a support-before-processing system. This is why we propose:
Effects on self-efficacy
Self-efficacy reflects a person’s belief that through their skills and capabilities, they can show a certain task performance (Bandura, 1977; Stajkovic & Luthans, 1998). Self-efficacy is an important work-related variable as it affects job performance and work satisfaction (Judge, Jackson, Shaw, Scott, & Rich, 2007; Lent, Brown, & Larkin, 1987; Stajkovic & Luthans, 1998).
Self-efficacy can be task specific or general (Jerusalem & Schwarzer, 1992; Schwarzer & Jerusalem, 1995). Specific self-efficacy strongly depends on a specific situation (Trippel, 2012), whereas general self-efficacy would generalize to a wider variety of situations (Jerusalem & Schwarzer, 1992). In the current study, specific self-efficacy relates to the perceived self-efficacy to perform the selection task at hand, whereas general self-efficacy relates to participants’ self-efficacy in personnel selection as a whole.
Successfully experiencing and independently fulfilling a task should strengthen both specific and general self-efficacy (Bandura, 1977). This should be especially pronounced for more demanding tasks (e.g., tasks that afford more information processing). Furthermore, receiving evidence of good performance can increase self-efficacy (Bandura, 1977). Describing the differences between the experimental groups in the terminology of work design research, only in the support-after-processing system and in the no-support group, participants may perceive that they autonomously fulfilled an entire cognitively demanding task. Additionally, participants in the support-after-processing condition might interpret the information by the system as an indicator of their task performance. All of this could strengthen specific and general self-efficacy.
In contrast, participants supported by the support-before-processing system might believe that they did not fulfill the entire task themselves, might perceive the task to be less demanding, and might experience less autonomy (Burton et al., 2019). Those participants could also believe that they were only following the advice of the system (Endsley, 2017; Langer et al., 2019). If this is the case, they might not perceive that they contributed to the task beyond what was already given by the system so completing the task would neither strengthen their feelings of being capable of performing the task at hand nor would it translate to self-efficacy in personnel selection in general.