Trust in artificial intelligence within production management – an exploration of antecedents

Abstract

Industry 4.0, big data, predictive analytics, and robotics are leading to a paradigm shift on the shop floor of industrial production. However, complex, cognitive tasks are also subject of change, due to the development of artificial intelligence (AI). Smart assistants are finding their way into the world of knowledge work and require cooperation with humans. Here, trust is an essential factor that determines the success of human-AI cooperation. Within this article, an analysis within production management identifies possible antecedent variables on trust in AI and evaluates these due to interaction scenarios with AI. The results of this research are five antecedents for human trust in AI within production management. From these results, preliminary design guidelines are derived for a socially sustainable human-AI interaction in future production management systems.

Practitioner summary: In the future, artificial intelligence will assist cognitive tasks in production management. In order to make good decisions, humans trust in AI has to be well calibrated. For trustful human-AI interactions, it is beneficial that humans subjectively perceive AI as capable and comprehensible and that they themselves are digitally competent.

Keywords:

• Production management
• cyber production management
• trust
• artificial intelligence
• machine learning

Introduction

Today's manufacturing companies need to master increasing complexity, to realise shorter delivery times and to react to dynamic changes (Kuhnle, Röhrig, and Lanza 2019). Current planning and control mechanisms (e. g. mathematical optimisation and heuristics) cannot efficiently deal with these complex, dynamic, and unpredictable environments (Stricker et al. 2018). Here, data-driven approaches from the field of artificial intelligence (AI) may provide valuable solutions (Stricker et al. 2018; Kuhnle, Röhrig, and Lanza 2019). Some argue that AI will increasingly take over complex cognitive tasks and support decision making (World Economic Forum 2015; McAfee and Brynjolfsson 2017). In this sense, AI is the counterpart to robotics in manufacturing companies: while robots facilitate physical doing of blue-collar workers, AI will support cognitive deciding of white-collar workers (McAfee and Brynjolfsson 2017). The long-term estimation is that AI will transform the system of production management to a cyber production management system where human and artificial intelligence cooperate successfully (Burggräf, Wagner, and Koke 2018).

De Visser, Pak, and Shaw (2018) point out that AI is different from previous automated systems: autonomously (model-based) instead of automatically (rule-based). Thus, AI can behave proactively, unexpectedly, and incomprehensibly for humans. Consequently, to establish productive and humane cooperation within a human-AI system under these circumstances, a particular design is required (Norman, Ortony, and Russell 2003; Haslam 2006; Riedl 2019; Shneiderman 2020). For humans to delegate control to others (either machines or fellow humans) requires trust in the respective counterpart. Accordingly, trust in the AI is the crucial factor for its broader acceptance (de Visser, Pak, and Shaw 2018; Siau and Wang 2018). Since AI is a new technology requiring new forms of interaction, existing models and theories on trust need to be verified in light of this new technology and revised if necessary (Detweiler and Broekens 2009; Farooq and Grudin 2016). The question of human trust in intelligent systems has been and is being investigated for several disciplines and applications. For example, in autonomous driving, the aim is to predict the driver’s trust based on factors such as gaze behaviour, culture, and gender (Hergeth et al. 2016; Hu et al. 2019), in order to adapt the vehicle’s behaviour automatically (Xu and Dudek 2015). In healthcare and homecare applications the tendency is to create transparent and adaptable systems (Rossi et al. 2017; Fischer, Weigelin, and Bodenhagen 2018). In safety-critical applications, such as airport security (Merritt et al. 2015) and military operations, the focus lies increasingly on the effects of errors on trust and performance (Rovira, Pak, and McLaughlin 2017).

The present article aims to empirically explore and examine antecedent variables on trust in AI within production management. Specifically, we derive an explanatory model, which describes important antecedent variables on trust and their operating mechanisms for the field of production management. This is a necessary first step towards the design of trustful human-AI cooperation in future cyber production management systems.

Background

In the future, AI will have the potential to take over knowledge work (McAfee and Brynjolfsson 2017), so that human activities may shift focus to problem-solving and monitoring activities, and more broadly on creative and emotion work (Dellermann et al. 2019). AI will support managers in coping with the increasing complexity and uncertainty in management. This requires interactive, cooperative decision-making mechanisms shared between artificial and human intelligence (McAfee and Brynjolfsson 2017). In line with Parasuraman, Sheridan, and Wickens (2000), Burggräf, Wagner, and Koke (2018) describes the forms of cooperation between AI and humans in production management in five stages: from sole human management via various degrees of assistance to fully autonomous decision-making by AI. Employing AI in production management has been subject of research since 1980 (Russell and Norvig 2016). This includes methods of operations research, such as fuzzy logic, genetic algorithms, or knowledge-based systems (Russell and Norvig 2016). Currently, there are three promising applications of AI in the field of production management. Firstly, in quality-based process control AI is able to reveal and learn dependencies between parameter combinations in process control and the resulting quality of the product, which are unknown today (Kuhnle, et al. 2017; Burggräf, Wagner, and Weißer 2020). If there is a digital twin for the real process, it is possible to use it as a simulation for reinforcement learning (RL) so that the algorithm learns the process control by itself (Jaensch et al. 2018). Thus, AI can replace classic control systems. Secondly, in predictive maintenance AI is capable of predicting machine failure based on the current machine status and machine history. The earlier this forecast is performed, the more time is available for preparation and downtime can be minimised. In the long run, predictive maintenance and further developments like prescriptive maintenance will even change the business model of the machines (Burggräf, Wagner, Dannapfel, et al. 2020). Thirdly, AI algorithms are increasingly being considered for scheduling/order dispatching. In this area of application, RL algorithms are receiving a significant amount of attention. They compete against the following: traditional algorithms of operations research (Gabel and Riedmiller 2008), heuristics of engineering science (e. g. first in first out, earliest due date) (Stricker et al. 2018; Kuhnle, Röhrig, and Lanza 2019), modified versions of themselves (Qu et al. 2016) and humans (Burggräf, Wagner, Koke, et al. 2020). Special characteristics of these current AI algorithms are that they can react very quickly to deviations, similar to heuristics or human decision making, and achieve a better solution quality than heuristics (Kuhnle, Röhrig, and Lanza 2019), but not as good as the slower optimisation methods (Gabel and Riedmiller 2008).

Despite all progress, it is unlikely that humans will be completely displaced in production systems in the near future. Thus, the new technology must be adopted by humans in order to maintain or better increase the overall performance. Due to its opaque, hard to predict and proactive nature, AI-based systems may be rather understood as a social counterpart than a classical tool by its users. Accordingly, people may think of it more akin to a (human) assistant or a team-mate. For this type of relationship trust is crucial (de Visser, Pak, and Shaw 2018; Hancock, Stowers, and Kessler 2019).

Trust has been widely studied in various disciplines. It seems that there is a common core of trust in all disciplines comprising of positive expectations and vulnerability (Rousseau et al. 1998). Lee and See (2004) summarised the core of trust as ‘the attitude that an agent will help achieve an individual’s goals in a situation characterized by uncertainty and vulnerability’. In information systems (IS) research on trust, the Integrative Trust Model of Mayer, Davis, and Schoorman (1995) is widely accepted. It was originally developed to describe interpersonal trust and therefore uses three dimensions of the trustee: ability, benevolence, and integrity. While previous authors have argued that these dimensions do not hold true for automated technology or software applications (Lee and See 2004; Söllner, Pavlou, and Leimeister 2013), recent authors claim that they are quite appropriate for autonomous technology (de Visser, Pak, and Shaw 2018). The reason for this shift is that autonomous, proactive, opaque, or anthropomorphic systems, such as robots or smart algorithms, often create the impression of a counterpart to be in dialog with rather than of a tool-like extension of one’s own body or mind, which in turn implies social interactions (Hassenzahl, et al. 2020). In other words, while ‘control’ is an attribute of a successful relationship to a tool, ‘trust’ is an attribute of a successful relationship to all those technologies experienced as a counterpart. More specifically, this success lies in the appropriate calibration of trust (de Visser, Pak, and Shaw 2018), that is, to avoid overtrust, which leads to misuse, as well as undertrust, which leads to disuse (Parasuraman and Riley 1997; Lee and See 2004; Drnec et al. 2016). While some variables/antecedents support the development of trust, other variables/antecedents diminish trust. Previous research on trust in non-human trustees revealed sets of various antecedents for different disciplines and non-human trustees, e. g. robotics (Hancock et al. 2011; Hancock et al. 2020), automation except for robotics (Schaefer et al. 2016), and software applications (Söllner, Pavlou, and Leimeister 2013). This research shows that the antecedents can be assigned to three groups: the trustor, the trustee, and the context/environment. Hoff and Bashir (2013, 2015) matched these three groups with the three layers of trust identified by Marsh and Dibben (2005): antecedents of the category trustor affect the dispositional trust; the antecedents of the non-human counterpart (trustee) affect the learned trust; and those of the environment affect the situational trust. First theoretical considerations assume that these groups and some of the antecedents hold true for AI as trustee (Siau and Wang 2018). However, to verify existing models and theories on trust in light of new technologies or new disciplines new research is needed (Detweiler and Broekens 2009; Farooq and Grudin 2016). Thus, the following section describes an approach to empirically examine trust in the new technology, AI, and within the new field of production management.

Research question and design

The research field addressed in this article is at the intersection of production management, artificial intelligence and trust. Consequently, all three facets are represented in the research question: ‘What does human trust in decisions of artificial intelligence depend on in the context of production management?’ The goal is to answer the research question with an explanatory model. First, it is necessary to understand the specifics of production management since a set of antecedents of trust for this field does not yet exist as it does in the described fields of robotics, automation, and software applications. For this reason, the existing sets of antecedent variables will be qualitatively mirrored and refined for AI and production management in the first phase. The second phase will quantitatively evaluate the resulting variables. For this purpose, a mixed-methods research strategy is used (Barton and Lazarsfeld 1984). It employs a qualitative approach as an auxiliary method for hypothesis generation with a subsequent quantitative approach for hypothesis testing. This design is characterised as follows: the results of the qualitative method are directly used in the quantitative method, so the overall analysis is not completely predefined. Due to the expected contribution to answer the research question, the second (quantitative) method is dominant over the first (qualitative) method. In summary, the two phases of this research design can be described as literature guided exploration (first phase) and statistical confirmation (second phase).

When carrying out the two phases and the discussion, two major topics of trust research are explicitly not dealt with, in order to reduce the complexity of the analysis. First, the distinction between different types of trust is excluded, e. g. contractual and relational in Rousseau et al. (1998). Second, the discussion of whether or not trust and distrust are independent constructs, e. g. Lewicki, McAllister, and Bies (1998), is not dealt with here.

Qualitative study for hypothesis development

In this first phase, we conducted a qualitative study to reveal potential antecedent variables influencing trust in the field of production management. We then structured and mirrored these identified antecedents by the existing antecedent sets of other fields and finally derive hypotheses.

Data collection method

The data was gathered in expert interviews (Meuser and Nagel 2009) and analysed with qualitative content analysis (Mayring 2014). The expert interviews aimed to systematically gather potential antecedents influencing trust in AI in production management. The interviews were intended to determine specific domain knowledge of production management and were therefore conducted as individual verbal interviews. To provide orientation and flexibility, the semi-structured interview form was used (Flick 2018). The guiding questions shown in Table 1 were transferred into an interview guideline, which led to interviews of 60 min on average. To ensure the objectivity of qualitative research, subjectivity must be limited by methodical control and reflection. For this purpose, a pre-test with two participants was carried out, which led to an improvement of the guideline.

Table 1. Guiding questions for the expert interviews.

Decisions and AI

How are decisions in production management made: based on evaluations or experiences or group discussions? Is the quality of the decisions reviewed? If it is possible, how? Imagine that in 30 to 40 years, AI will be able to make decisions that are currently reserved for humans. Which decisions in production management could be affected first? Can you imagine that an AI is not only subordinate (assistant) to humans, but equal (colleague)?

Trust

In your opinion, how important is trust in production management? In which decisions do you think greater trust is required and in which lower trust? What does this depend on? If the colleague is an AI, are there other/further antecedents on which trust could depend? What other enablers or barriers to trust in AI in production management might exist?

The expert interviews were conducted with four experts. They were carefully selected from the author’s professional and personal network. For us, it was important to select people who, in our eyes, are mentally able to imagine a future workplace with AI colleagues in production management. The first two interviewees are heads of production in the mechanical and plant engineering companies. One interviewee had 14 years, while the other had 20 years of experience in production management of companies with approximately 350 employees and high manufacturing depth. In recent years, both have implemented systems to automate previously cognitive processes within their companies, such as advanced planning and scheduling systems, and both are engaged in further automation based on AI. The last two interviews were conducted with employees of research institutions with a large consulting share in the field of production management. Both are managing groups of above 20 production researchers and have up to 25 years consulting experience. Both are working on questions of future AI applications for planning tasks in manufacturing companies. The combination of business- and science-oriented perspectives serves to generate a broad picture of potential antecedent variables of trust in AI in production management. The following analysis will show that multiple aspects were mentioned by all interviewees and that the hypotheses were well developed. Nevertheless, the small number of interviews needs to be mentioned as a limitation of this work.

Analysis approach

The interviews were analysed with qualitative content analysis according to Mayring and Fenzl (2014). In light of the objective of this first phase, the technique of summarising qualitative content analysis was employed (Mayring and Fenzl 2014). This technique consists of four sequential steps: paraphrasing of the most important text passages; generalisation by adjusting the abstraction levels of all paraphrases; first reduction of unimportant or double paraphrases; second reduction by bundling the remaining paraphrases (Mayring and Fenzl 2014).

During the inductive formation of the categories/codes, the transcripts were repeatedly read and the developed categories were iteratively adjusted. For this purpose, two researchers independently used the web application QCAmap from Mayring and Fenzl (2014). After that, they compared, discussed, and aligned their codes. The codes were then aggregated, renamed, and structured to a system of categories employing the theoretical models mentioned above, e. g. the Integrative Trust Model (Mayer, Davis, and Schoorman 1995). While some of the categories serve for the later formulation of human-AI decision scenarios (phase 2), six of them are identified as potential antecedent variables of trust in AI (see Figure 1). These factors have the potential to influence the trust in and subsequently the reliance on the system. Finally, it is expected to increase the performance of the human-AI team (Dzindolet et al. 2003; Lee and See 2004; Hoff and Bashir 2015; French, Duenser, and Heathcote 2018).

Figure 1. First potential antecedents for trust in AI within production management.

Organisational chart of the six antecedents divided into the three characteristic groups, first the AI group consisting of the two antecedents perceived capability and perceived understandability, second the human group consisting of the expert status and the digital affinity, and third the decision situation group consisting of the predictability and the error costs.

Hypotheses development

Similar to the trust model for interaction with an intelligent, autonomous robot of Matthews et al. (2020), we assumed that not all of our six antecedents directly influence the trust, but get mediated through other antecedents. Thus, we employed the (probably) most generic expression of a mediation model (Baron and Kenny 1986), the SOR model of Woodworth (1926), to structure our antecedents. Figure 2 summarises this structure and shows that trust is the response in the SOR model. The assignment of the independent variables and their presumed relationship to trust is described hereafter.

Figure 2. Model of hypotheses under investigation.

Illustration of the six antecedents and trust in the SOR model, as described above. In addition, there is an arrow going from each antecedent to the trust, with the label of the respective hypothesis. Here, the arrows for hypotheses H5 and H6 are drawn in dashed lines to indicate the assumed mediated effect.

The two antecedents of the category AI (trustee) do not describe the objective but attribute characteristics of the AI. For example, not the measurable accuracy of a prediction was assumed to foster trust, but instead the subjective perception of AI’s ability. The same applied for perceived comprehensibility of the AI’s decision. For both antecedents, the individual and subjective perception of a given explanation was examined. Thus, we assigned these two antecedents to the organism in the SOR model and hypothesised the following direct relationships to trust.

H1. Perceived ability of AI in production management increases trust.

H2. Perceived comprehensibility of AI in production management increases trust.

The assignment of the category human (trustor) was trivial since it represents almost exactly the organism of the SOR model. The antecedent expert status describes the person’s experience in production management and skills to solve tasks in this field. The digital affinity characterises the degree to which a person has natural attraction, interest, and pleasure in learning and using (the latest) digital methods and technologies. Digital affinity should not be interpreted as familiarity. While affinity reflects a personal trait, familiarity describes acquired knowledge through experience. In line with the interviewed experts, we assumed that both – expert status and digital affinity – are directly related to trust, but that more experienced and skilled persons in production management trust less because they themselves are the experts and therefore are more sceptical.

H3. The expert status of humans decreases his/her trust in AI.

H4. The digital affinity of humans increases his/her trust in AI.

The antecedents of the category decision situation are the predictability (or inverted uncertainty) and the costs of an error (or risk) of the decision situation. We assigned both to the stimulus in the SOR model because they are determined by external characteristics. Based on the expert interviews, we assumed that these antecedents directly affect the perception of AI (trustee) characteristics and thus are mediated through them to indirectly affect the trust. We expected that trust is higher in riskier and more uncertain decision situations because managers themselves fear their inability and making poor decisions under these conditions (Riabacke 2006).

H5. The predictability of the situation in production management indirectly decreases trust in AI, mediated through perceived ability and/or perceived comprehensibility.

H6. The error costs of the situation in production management indirectly increase trust in AI, mediated through perceived ability and/or perceived comprehensibility.

Quantitative study for hypothesis testing

In this second research phase, a quantitative study was used to statistically test the significance of the hypotheses. Guided by a linear approach, the discovered antecedents were examined.

Data collection method

The quantitative study aimed to test the developed hypotheses. As a method for data collection, an online survey was conducted to acquire a large number of heterogeneous participants. Since the potential antecedents represent latent variables and are therefore not directly measurable, it was necessary to operationalise them with a measurement system and thus transfer them to a questionnaire (Kumar 2014). The measurement system is described in Table 2 (and the formulation of the items is appended in Appendix A). In line with prior research (Mcknight et al. 2011; Lankton, McKnight, and Tripp 2015; Gulati, Sousa, and Lamas 2018) we modelled trust as a reflective construct. Moreover, the other antecedents were also modelled as reflective constructs, according to the decision rules of Jarvis, Mackenzie, and Podsakoff (2003). To assure reliability and validity we used established items, modified established items, or followed the suggestions of Morgado et al. (2017) for the development of new items.

Table 2. Measurements to operationalise the variables.

Variable	No. of items	Scale	Source
Trust	3	5-stage Likert	Beierlein et al. (2014)
Perceived ability	3	5-stage Likert	Schyns and Collani (2014)
Perceived comprehensibility	2	5-stage Likert	self-developed
Digital affinity	4	5-stage Likert	self-developed with Karrer, Glaser, and Clemens (2009) and Neyer, Felber, and Gebhardt (2016)
Expert status	1	5-level branching	self-developed with Benner (2017)
Predictability	1	3 scenarios	self-developed
Error costs	1	3 scenarios	self-developed

While all items were surveyed via a self-assessment of the participants, the variables predictability and error costs were operationalised by four scenarios (decision situations), that were derived from the expert interviews of phase one. Each scenario determined a specific value for the two variables, e. g. low, average, or high. The scenarios described an AI assistant recommending a specific decision in a defined situation of production management. Scenario 1 (layout optimisation) dealt with the repositioning of machines in the layout of a production. This was a decision of average predictability and average error costs. In scenario 2 (order acceptance), the AI assistant advised rejecting an order of 800,000 € because no capacities were available, which corresponded with high predictability and high error costs. Scenario 3 (job scheduling) described an AI assistant that scheduled the job sequence in the production shop. This task had high predictability and low error costs. In scenario 4 (production expansion) the AI assistant recommended expanding the production plant for 800,000 € to increase technological flexibility. Here, low predictability and high error costs were present. In the questionnaire, each of the four AI assistants was given a name to make it easier to distinguish between the scenarios (see Appendix B for the full descriptions of the scenarios).

In total, the questionnaire comprised 20 items, partially reverse-coded. Within the questionnaire, the items were divided into different blocks. Hereby, the items of one block were homogenous and the order of the items in one block supported an easy cognitive processing: the items build up logically on each other, and simple and heavy items alternate (Kumar 2014). The questionnaire designed in this way was subjected to a pre-test with three researchers, who have experience in production management. Besides, the pre-test also provided information on the duration it takes to answer it: approx. twelve minutes. The target group of the survey was current and former employees of manufacturing companies, previous experience with AI was not required. Potential participants were asked via personal messages and posts on different social networks to participate in this study. The questionnaire was published online from 08/01/2018 to 08/14/2018. At the end of the survey period, 182 respondents participated in the survey (gross sample). Of those 52 respondents did not complete the survey, so the retention rate was 71.4% and the net sample consisted of 130 respondents (43 females and 87 males), with a mean age of 32 years (SD = 10). Regarding the age distribution, accumulation was observed in the age range from 25 to 30 years (41.6%), while 21.5% of the participants were below and 36.9% above this age range. 67% of the participants had a university degree (bachelor, master, diploma, or PhD), while 33% were without an academic degree (skilled worker or technician). Also, 67% of the participants had professional experience in a manufacturing company, while 33% did not have this experience. The values of the net sample are publicly available on Mendeley Data, see Saßmannshausen (2020).

Analysis approach

As the next step, the data had to be prepared. This included the coding of verbal response options to numerical values. The following analysis was performed with PLS using the software SmartPLS version 3.3.2 (Ringle, Wende, and Becker 2015). The rationale for choosing PLS as a method for structural equation modelling (SEM) was threefold. First, according to Hair et al. (2016) as well as Jannoo et al. (2014) PLS is superior to covariance-based SEM methods (CBSEM), when the goal lies in explorative theory development and predictive relevance of antecedent variables – as it is the case here. Second, the choice of the proper SEM method depended on the underlying nature of the data. Sarstedt et al. (2016) developed a framework based on previous and own studies. For common factor data and reflective constructs, the CBSEM method is better suited, but PLS also provides valid results. In contrast, PLS is perfectly suited when the data follow a composite nature, while CBSEM should not be used here (Sarstedt et al. 2016). Thus, the use of PLS is preferred cases, when data nature is unclear or unknown because it can handle both. As an indicator of the data’s nature, the standardised root mean square residual (SRMR) is proposed by Sarstedt et al. (2016). As a threshold, they reference to Hu and Bentler (1998) and recommend that values of 0.08 or smaller indicate composite data. The data of the present study had an SRMR of 0.08, which indicated the use of PLS. The third reason for choosing PLS lied in its advantages to handle unusual data characteristics (e. g. non-normal distributions) and complex models (six or more antecedents) – as it is mostly the case in social science (Hair et al. 2014, Sarstedt, Ringle, and Hair 2017). Therefore, we investigated the data distribution using Shapiro-Wilk’s test (p < 2.2e⁻¹⁶, should be > 5% for normal distribution) and optical inspection of the density plot as well as QQ plot. The results showed that the data was not normally distributed. Consequently, PLS was preferred over CBSEM for the analysis.

The PLS assessment followed the two-stage approach of Sarstedt, Ringle, and Hair (2017). The first stage was the evaluation of the measurement model (Sarstedt, Ringle, and Hair 2017). This was done in four steps by examining item loadings, internal consistency, convergent validity, and discriminant validity. In step one, the item loadings should at least be 0.7 (Hair et al. 2016; Sarstedt, Ringle, and Hair 2017). Consequently, we dropped three items, that did not fulfil this criterion (items Trust_3, Digi_3 and Digi_4). For the internal consistency (IC) in step two, we reported three coefficients. Two of them are the most common coefficients according to Cho (2016): the congeneric reliability (also composite reliability and upper bound of IC) and the tau-equivalent reliability (also Cronbach's Alpha and lower bound of IC). As third coefficient for IC we employed the new reliability measure introduced by Dijkstra and Henseler (2015): Roh A. As shown above our data was of composite nature, which caused the tau-equivalent reliability to underestimate the IC (Dijkstra and Henseler 2015), which was even reinforced by our scales with few items (Cortina 1993). Furthermore, we could assume that the congeneric reliability overestimated the IC because the item loadings in PLS tend to be upward-biased (Dijkstra and Henseler 2015). Consequently, we focussed on Roh A when evaluating the coefficients, but also considered the other two coefficients slightly, see Table 3. Values above 0.7 indicate good IC, while values between 0.6 and 0.7 are acceptable for exploratory research (Nunnally 1967; Hair et al. 2016). For Roh A all variables passed this threshold. Since the tau-equivalent reliability underestimated the IC it was not surprising that some variables would not have passed this threshold if regarded without differentiation. Given that Roh A exceeded the threshold there is no need for action such as dropping further items. In step three we evaluated the convergent validity by calculating the average variance extracted (AVE). The AVE should be 0.5 or higher (Hair et al. 2016), as it was for all variables under investigation (see Table 3).

Table 3. Evaluation of measurement model.

Variable	Item	Loading	Tau-equivalent reliability	Roh A	Congeneric reliability	AVE
Evaluation step		one	two	two	two	three
Trust	Trust_1	0.775	0.632	0.731	0.838	0.723
	Trust_2	0.919
Perceived ability	Abi_1	0.848	0.728	0.782	0.840	0.638
	Abi_2	0.788
	Abi_3	0.757
Perceived comprehensibility	Comp_1	0.894	0.532	0.604	0.805	0.677
	Comp_2	0.742
Digital affinity	Digi_1	0.886	0.646	0.728	0.861	0.757
	Digi_2	0.852
Expert status	Exp_1	1.000	1.000	1.000	1.000	1.000
Predictability	Pred_1	1.000	1.000	1.000	1.000	1.000
Error costs	Cost_1	1.000	1.000	1.000	1.000	1.000

The fourth step assessed the discriminant validity to ensure that the variables were unique and distinct from the other variables (Sarstedt, Ringle, and Hair 2017). Here the Fornell-Larcker criterion and the cross-loadings are the most frequently used approaches (Henseler, Ringle, and Sarstedt 2015). Table 4 shows the assessment of the Fornell-Larcker criterion, where the bolt values on the diagonal should be greater than the values in the corresponding row and column (Hair et al. 2016) – which is fulfilled here. In addition, Table 5 shows the cross-loading of the items. Here, the loading of each item to the assigned variable should be higher than the loading to any other variable (Gulati, Sousa, and Lamas 2018) – which is also fulfilled here. Thus, our instrument meets both criteria. According to Henseler, Ringle, and Sarstedt (2015) we also assessed the heterotrait-monotrait ratio (HTMT) to prove the discriminant validity. The HTMT should be smaller than 0.9 for conceptual similar variables/constructs and smaller than 0.85 for distinct variables/constructs (Henseler, Ringle, and Sarstedt 2015). As shown in Table 6 all HTMT ratios fulfil this requirement. Summing up the four steps for evaluation of the measurement model, our instrument met all criteria, after dropping three items, and had good internal consistency/reliability, convergent, and discriminant validity.

Table 4. Evaluation of Fornell-Larcker criterion.

Variable	Trust	Perc. ability	Perc. comp.	Digital affinity	Expert status	Predictability	Error costs
Trust	0.850
Perceived ability	0.578	0.799
Perceived comprehensibility	0.418	0.505	0.823
Digital affinity	0.198	0.120	0.063	0.872
Expert status	−0.122	−0.091	0.012	−0.371	1.000
Predictability	−0.167	−0.266	−0.271	0.000	0.000	1.000
Error costs	0.121	0.241	0.178	0.000	−0.000	−0.455	1.000

Table 5. Evaluation of cross-loadings.

Item	Trust	Perc. ability	Perc. comp.	Digital affinity	Expert status	Predictability	Error costs
Trust_1	0.775	0.359	0.261	0.128	−0.116	−0.112	0.064
Trust_2	0.919	0.587	0.423	0.197	−0.099	−0.164	0.130
Abi_1	0.564	0.848	0.529	0.077	−0.024	−0.328	0.270
Abi_2	0.382	0.789	0.348	0.107	−0.088	−0.166	0.151
Abi_3	0.396	0.758	0.271	0.118	−0.140	−0.079	0.118
Comp_1	0.405	0.477	0.895	0.072	0.028	−0.261	0.149
Comp_2	0.264	0.338	0.744	0.023	−0.017	−0.174	0.147
Digi_1	0.182	0.090	0.078	0.888	−0.316	−0.000	0.000
Digi_2	0.162	0.120	0.029	0.856	−0.333	−0.000	0.000
Exp	−0.122	−0.091	0.012	−0.371	1.000	−0.000	−0.000
Pred	−0.167	−0.266	−0.271	0.000	−0.000	1.000	−0.455
Cost	0.121	0.241	0.178	0.000	−0.000	−0.455	1.000

Table 6. Evaluation of heterotrait-monotrait ratio (HTMT).

Variable	Trust	Perc. ability	Perc. comp.	Digital affinity	Expert status	Predictability	Error costs
Trust	–
Perceived ability	0.777	–
Perceived comprehensibility	0.663	0.736	–
Digital affinity	0.287	0.179	0.137	–
Expert status	0.157	0.122	0.037	0.450	–
Predictability	0.203	0.278	0.359	0.000	0.000	–
Error costs	0.142	0.261	0.245	0.000	0.000	0.455	–

The second stage of the PLS assessment was the evaluation of the structural model, which consists of five steps. For these steps, we used bootstrapping with 5,000 sub-samples as a non-parametric approach, especially for the evaluation of the mediation effects. In the first step, we assessed the multi-collinearity of the antecedents by use of the variance inflation factor (VIF). Depending on the literature the VIF should be smaller than ten (Hair et al. 2014) or even smaller than five (Sarstedt, Ringle, and Hair 2017). Table 7 shows that the variables met this requirement. The second step evaluated the path coefficients (effects of the antecedents) and their significance. As shown in Table 7 the direct effects of perceived ability, perceived comprehensibility and digital affinity were all significant. Here, the perceived ability had the strongest effect with 0.48, which means that the increase of one of perceived ability by one standard deviation is accompanied by the increase of trust by 0.48 standard deviations, whereas all other antecedents remain constant. In the case of the antecedents predictability and error costs, there were significant indirect effects. Since the direct effects were not significant, this is a full mediation, according to Nitzl, Roldan, and Cepeda (2016). These mediations were further analysed, as the indirect effects were formed via two moderators (perceived ability and perceived comprehensibility). The indirect/moderating effect of predictability in Table 7 is the combination of the two sub-effects: the indirect effect via perceived ability (−0.095, p < 0.001) and the indirect effect via perceived comprehensibility (−0.042, p < 0.01). In contrast, only the sub-effect of error costs via the mediator perceived ability was significant (0.073, p < 0.01), while the sub- effect via perceived comprehensibility was not significant (0.012, p > 0.05).

Table 7. Evaluation of structural model with trust as independent variable.

Variable	VIF	Path coefficient with p-value		Indirect effect with p value		f²	R²	Q²
Perceived ability	1.423	0.480	0.000***			0.264	0.843 (adj. 0.377)	0.249
Perceived comprehensibility	1.387	0.174	0.000***			0.037
Digital affinity	1.172	0.117	0.002**			0.021
Expert status	1.170	−0.037	0.271			0.003
Predictability	1.334	−0.006	0.881	−0.137	0.000***	0.002
Error costs	1.286	−0.028	0.501	0.085	0.001**	0.003

**p ≤ 0.01: very significant; ***p ≤ 0.001: highly significant.

In the next, third, step we assessed the individual contribution of the antecedents. Thus, we reported the f² (effect size), see Table 7. In line with the analysis above, the three significant direct paths showed effect sizes different from zero, which can be labelled as small to medium (Cohen 1988).

Following Hair et al. (2016), in the next steps, we analysed the overall suitability of the model. Thus, we calculated the coefficient of determination in step four. With an R² of 0.843 (adjusted R² = 0.377), we nearly explained 40% of the variance of trust, which can be interpreted as weak to moderate (Henseler, Ringle, and Sinkovics 2009; Sarstedt, Ringle, and Hair 2017) or even as large effect size (Cohen 1988; Kotrlik and Williams 2003). To support this R² we calculated the statistical power, which should be higher than 0.8 for α  =  0.05 (Nitzl, Roldan, and Cepeda 2016). Using G*Power version 3.1.9.6 our statistical power was 1.000 (with α  =  0.05, n = 130 and six predictors), which is substantial. As the last and fifth step of the evaluation of the structural model we reported the predictive accuracy of the model, Q². The Q² was calculated using a blindfolding approach with a recommended omission distance of seven (Hair et al. 2017). Our Q² of the variable trust was 0.249. This value fulfilled the requirement to be above zero (Sarstedt, Ringle, and Hair 2017), so the model’s prediction of trust is acceptable and relevant. After this presentation of the analysis results, the following section outlines what these results mean regarding the developed hypotheses.

Hypotheses summary

We developed six hypotheses based on the literature, enriched by qualitative information from the conducted expert interviews. The significance of these hypotheses can be derived from the analysis results presented above. For a better overview, we have summarised the relevant findings for the hypotheses in Figure 3.

Figure 3. Coefficients of the structural model.

Illustration of the six antecedents and trust with an arrow going from each antecedent to trust labelled with the respective effect size and significance. Significant effects on trust with their respective effect size are digital affinity (0.117), perceived ability (0.480), perceived comprehensibility (0.174), predictability as indirect effect (−0.137), and error costs also as indirect effect (0.085).

Both attributed characteristics of AI (trustee) – perceived ability and perceived comprehensibility – have a significant positive effect on trust. Thus, their respective hypotheses H1 and H2 hold true for this study. For both assumed relationships of the characteristic of the human (trustor) and trust, only one is statistically significant. While the expected negative effect of expert status on trust seems to be correct, this effect is not significant and therefore H3 is not confirmed. In contrast, H4 is significant, which means that a higher digital affinity leads to higher trust.

The two assumed indirect effects of the decision situation, namely H5 and H6, hold true in the present study. Since the corresponding direct effects on trust are not significant, full mediation is present. For the mediation the two mediators perceived ability and perceived comprehensibility were assumed. Looking at the partial mediations through only one of these mediators, we see that the antecedent predictability has a significant indirect effect on trust through both mediators. Thus, the sub-hypotheses H5a and H5b hold true. In contrast, for the antecedent error costs only the mediation through perceived ability is significant (H6a), but the one through perceived comprehensibility is not (H6b). The results of the hypothesis testing are summarised in Table 8.

Table 8. Results of hypothesis testing.

Hypothesis	Statement	Significant
H1	Perceived ability of AI in production management increases trust.	Yes
H2	Perceived comprehensibility of AI in production management increases trust.	Yes
H3	The expert status of humans decreases his/her trust in AI.	No
H4	The digital affinity of humans increases his/her trust in AI.	Yes
H5	The predictability of the situation in production management indirectly decreases trust in AI, mediated through perceived ability and/or perceived comprehensibility.	Yes
H5a	Mediation only through perceived ability	Yes
H5b	Mediation only through perceived comprehensibility	Yes
H6	The error costs of the situation in production management indirectly increase trust in AI, mediated through perceived ability and/or perceived comprehensibility.	Yes
H6a	Mediation only through perceived ability	Yes
H6b	Mediation only through perceived comprehensibility	No

Discussion

The test of the hypotheses and the structural model show that the antecedents predictability and error costs have no significant direct effect on trust, which is in line with a meta-analysis of factors influencing trust in automation (Schaefer et al. 2016). However, our study supports the expectation of Lee and See (2004) that there should be some effect. Instead, these antecedents contribute to the explanation of trust with their fully mediated indirect effects. This shows that the employment of the SOR model as a structural framework for trust antecedents is quite useful. In light of Hoff and Bashir (2015)’s trust model, our findings indicate that the two identified antecedents of the group environment do not affect the situational trust as implied by the model, but indirectly affect the learned trust. Since the mediations through perceived ability are stronger than through perceived comprehensibility it seems that trust depends on the human’s understanding of how the situation affects the AI’s capabilities (Lee and See 2004).

Considering the category AI (trustee) characteristics, both antecedents contribute to the explanation of trust. In IS research, the most mentioned antecedents of this category are ability, benevolence, and integrity (Mayer, Davis, and Schoorman 1995) from human-human trust. Also, their further developments gain more attention, like performance, process, and purpose (Lee and See 2004) or performance, helpfulness, and predictability (Söllner, Pavlou, and Leimeister 2013). So, ability is known as a traditional and strong influencing factor (Hancock et al. 2011; Schaefer et al. 2016). In this study, the perceived ability has the strongest effect of all analysed antecedents (β = 0.48, p < 0.001). This effect indicates that an increase of one standard deviation in perceived ability goes along with a rise in trust of 0.48 standard deviations – assuming all other antecedents are held constant. Previous studies also underline that the use of perceived ability instead of objective/true ability was a suitable choice. Thus, our first recommendation for AI designers to foster trust is to make the AI appear capable. For instance, a certain ability is perceived depending on the features of the IT system’s interface, even if these features are not related to the true abilities (Tseng and Fogg 1999). At the beginning of an AI implementation, it is also conceivable to verify the decisions of the AI by a second (e. g. rule-based) system or a human being in order to avoid errors and error messages and thus to increase the perceived ability. This serves as a great adjustment mechanism for trust calibration.

The variable perceived comprehensibility is a novel trust antecedent that has become increasingly important since humans interact with intelligent and autonomous systems (Matthews et al. 2020) – as the efforts of governments and companies to explainable AI underline (Cutler, Pribić, and Humphrey 2019; Fjeld and Nagy 2020). The effect on trust is not as strong as the one of perceived ability: when perceived comprehensibility increases by one standard deviation, trust rises 0.174 standard deviations – assuming all other antecedents stay constant. This leads to the second design recommendation: subjectively better explanations or more expectable decision-making behaviour of the AI lead to higher trust. Consequently, there is a need for explainable AI (XAI). Waltl and Vogl (2018) provide a taxonomy of different learning systems and their natural contribution to explainability. As neural networks are not transparent by nature there are some efforts to support their explainability (see e. g. Ribeiro, Singh, and Guestrin (2016) or Schaaf, Huber, and Maucher (2019)). Moreover, trust can be further increased using a visually represented agent communicate the explanation rather than just displaying it (Weitz et al. 2019). In the category of human (trustor) characteristics the expert status of the human in the field of production management seems not to contribute to the explanation of trust. In other IS research similar constructs like ability or competence of the human affect trust (Schaefer et al. 2016; Siau and Wang 2018). Although our study does not support it, more experienced production managers could tend to trust less. This relationship then could be explained by a study of Kantowitz, Hanowski, and Kantowitz (1997) on vehicle navigation aid. The more familiar a person was with a city, the more negative navigation errors were judged, in terms of trust and reliability. Therefore, an experienced production manager may be more critical regarding the decisions of an AI assistant (risk of undertrust) than an inexperienced production manager, who might even tend to overtrust.

Digital affinity is a new trust antecedent that has not yet been considered in IS trust research on autonomous systems. An analogous antecedent, however, was found in research on flight automation: the general accustoming of pilots to the use of automated solutions (Riley 1996). It seems that familiarity with technology leads to higher trust, which is supported by the affinity to that technology. Here, digital affinity shows a strong and statistically significant effect. Our third recommendation is addressed to designers of human-AI systems. Since a trustful relationship leads also to a higher team performance, it is advisable to introduce the AI to digitally competent employees. In addition, this trait might become a criterion for staff recruiting.

The gained insights of trust in AI in production management can be used, for example, to introduce an AI assistant to employees and thus foster the trust in the AI. Here, the perceived ability seems to be a mandatory requirement. However, this model currently explains only about 40% of the trust variance. Thus, there is a need for further research to improve the model, partly because we built it on a small number of expert interviews. For example, further antecedents that were not considered in this study should be added, like the visual representation of the AI (Siau and Wang 2018) or the culture of an organisation (Lee and See 2004). Moreover, a new category of antecedents could be introduced next to the categories of the trustor (human), the trustee (AI), and the decision situation: the category of interaction. Employing the trust model of Hoff and Bashir (2015) the antecedents of this new category could affect the dynamic learned trust – a result of dividing the learned trust into initial learned trust and dynamic learned trust. Here, adaptive or adaptable shifts between human and AI decision making/control could be introduced, depending on the human’s trust and capabilities in the situation (Lee and See 2004). All this will lead to a more accurate model and the verification of the presented results. As an extension to the refinement of the antecedent side of trust, we should also take a look at the consequence side. As seen in Hoff and Bashir (2015)'s trust model, we could explore how trust affects reliability, performance, and satisfaction.

It should be noted that this study was an imagination study, so the participants did not face any consequences. This could have led to a distortion of the response behaviour, following the risk compensation theory (Hedlund 2000). Therefore, further research should address a corresponding experimental and experienceable design within production management. In such experimental settings, the linearity of the relationships assumed here could also be questioned, since there are indications that, for example, the relationship between ability and trust is not linear (Muir and Moray 1996; Kramer 1999). Additionally, the stability/robustness of trust in different settings can be examined to gain deeper insights, as indicated by McKnight, Cummings, and Chervany (1998).

Conclusion

The objective of this study was to discover and evaluate antecedent variables of trust in AI for production management. Here, the antecedent of digital affinity stands out, as it has not yet been included in any set of antecedents for trust between humans and AI. Other examined antecedents are similar to those in the literature on robots and automation, with the difference that we have studied the perceived (not the objective) characteristics of AI. Since this perception happens inside humans, we employed the SOR model to structure the antecedents. Summing up, it was possible to contribute to answer the guiding question: ‘What does human trust in decisions of artificial intelligence depend on in the context of production management?’ The results serve as first design guidelines for a socially sustainable human-AI interaction in production management. Autonomous and independent decision-making AI assistants will offer a sensible and possibly necessary addition to human managers in the future. Because a broad explanation of trust was not achieved, the inclusion of further antecedents should be the topic of further research to ensure the human-centered design of future production management systems.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Appendix A. Operationalisation of variables for online survey

 

Table

Variable	Item	Scale	Source
Trust	I am convinced that the AI has good intentions.	totally true rather true neither nor rather not true not true at all	Beierlein et al. (2014)
	You cannot rely on the AI. (-)
	In general, the AI can be trusted.
Perceived ability	The AI is up to the needs of the situation.	totally true rather true neither nor rather not true not true at all	Schyns and Collani (2014)
	If the AI is confronted with a new situation, it knows how to deal with it.
	For every problem the AI has one or more ideas for solutions.
Perceived comprehen-ibility	I do not understand the decision of the AI. (-)	totally true rather true neither nor rather not true not true at all	self-developed
	The explanation helped me to understand the AI.
Digital affinity	I am very interested in the latest digital developments, even though I have no intention of buying.	totally true rather true neither nor rather not true not true at all	self-developed with Karrer, Glaser, and Clemens (2009) and Neyer, Felber, and Gebhardt (2016)
	I love to use and own the latest electronic devices.
	For me, dealing with new PC programs or apps is usually too much of a challenge. (-)
	It is easy for me to learn how to use new devices, programs and apps.
Expert status	In your career to date, have you gained professional experience in the field of production?	Branching questions with yes/no or always/never answers	self-developed with Benner (2017)
Predictability	Stages of low, mean and high predictability within the scenarios	Different Scenarios: Layout optimisation (mean predictability and mean error costs) Order acceptance (high predictability and high error costs) Job scheduling (high predictability and low error costs) Production extension (low predictability and high error costs)	self-developed descriptions of different decision situations
Error costs	Stages of low, mean and high error costs within the scenarios

 

Appendix B. Description of the decision scenarios for the survey

 

Table

Scenario	Introduction	Description
Scenario 1 layout optimisation	For the tasks and decisions concerning material flow optimisation in your production, you use the artificial intelligence FLOW.	FLOW recommends you to reposition several machines in your production shop, which would require an investment of 80,000 €. FLOW justifies the decision by saying “The new machine arrangement enables a more efficient and flow-oriented logistics between the machines. Stocks on the machines can be reduced and thus the overall throughput time can be decreased”.
Scenario 2 order acceptance	The artificial intelligence DOOR helps you with the tasks and decisions to accept orders for your production.	DOOR recommends you not to accept a new order worth 800,000 €. DOOR justifies the decision with “The production is already overloaded for the next five months, which is balanced by overtime and working weekends. This new order should be ready in four months, which is not achievable”.
Scenario 3 job scheduling	The artificial intelligence PRIO assists you with the tasks and decisions for planning the job sequence.	PRIO recommends the rescheduling of an order from machine 6 to machine 2. PRIO justifies the decision by stating “Machine 2 has free capacity due to delivery delays so it can relieve machine 6. In total, orders can thus be accelerated”.
Scenario 4 production expansion	In the tasks and decisions that affect the entire production site, you are supported by the artificial intelligence SITE.	SITE recommends that you invest in the purchase of a metal 3 D printer and the necessary extension of the production shop for a total of 800000 €. SITE justifies its decision with “To enter a strategic market niche, it is necessary to invest in this new production technology”.