Search

For decades, efforts have been made to reduce one of the most central causes of organizational stress, namely workload, which describes the amount or difficulty of employees’ work (Bowling & Kirkendall, 2012). With the rapid development of technology, newer and newer technologies are being developed to reduce workload, and thus counteract stress. New technologies such as software agents are increasingly used to perform complex tasks (O’Neill et al., 2022), and are supposed to decrease workload of humans (Wright et al., 2016; Bradshaw, 1997). Those software agents differ in their level of autonomy (LOA), and, consequently, in the degree to which they assist with or takeover tasks. Besides workload, there are other dangerous overload conditions for employees that can lead to negative consequences such as stress, namely information overload (IO; Graf & Antoni, 2022), which describes a state of being overwhelmed by information, where one perceives that information demands exceed one’s information processing capacity (Antoni & Ellwart, 2017). Whether the use of agents with different LOAs and their different task support only influences the workload, or also the information processing and therefore IO, has not yet been clarified. The more autonomous and thus the more the software agent takes over work, the less operational work is done by humans, which should be reflected in lower workload (Prewett et al., 2020). Research relates this decrease in workload with higher autonomy of robots and systems, but also unclear findings are present (Selkowitz et al., 2017; Omnasch et al., 2014). Moreover, information processing could decrease if the agent acts autonomously, and its processes are not followed. However, monitoring of the agents’ actions might create additional information load, because high, and/ or divided attention and therefore information processing is still needed (Wickens & Carswell, 2021). If the software agent acts at high speed with high autonomy, IO might even increase due to the fast-paced monitoring task. Therefore, different LOA could affect IO differently, since different amounts and complexities of information need to be incorporated depending on agent behavior. Research suggested that software agents may reduce human IO (Berghel, 1997; Bradshaw, 1997, Maes, 1997) - however, if the use of different LOAs of software agents leads to less information to be processed and thus there is less risk to perceive IO has not been clarified to date in a randomized-controlled study. Therefore, this research transfers workload and IO to the context of software agents in human-autonomy contexts and observes the (possible) differential influence of LOA. Both high workload and information overload (IO) are two processes established in research that represent an imbalance between environmental requirements and internal capacities (Graf & Antoni, 2022; Meijman & Mulder, 2013, p. 5-6). Stress perceptions can result from such an imbalance between external demands and internal regulatory capabilities (Koolhaas et al, 2011), particularly if loss of control and negative consequences are perceived. In the literature, mixed findings are present that show that higher LOA might increase the perception of technology related stress and exhaustion (Ulfert et al, 2022), decrease it (Pollak et al., 2020), or do not influence it at all (Sauer et al., 2013; Szalma & Taylor, 2011), despite that software agents may reduce workload (Omnasch et al., 2014). These mixed findings leave open, whether there is a higher experience of stress when using different LOA software agents, and what reasons lie behind this. Therefore, this study focusses on two research questions, namely, 1) how information overload is influenced by different LOA, and if these effects are differential from effects on workload, and 2) how stress perceptions (i.e., stress, technostrain) are influenced by different LOA, and if those effects differ.

Human-autonomy or human-agent teams (HATs) are socio-digital systems “involving one or more humans and one or more autonomous agents, wherein each human and autonomous agent is recognized as a unique team member occupying a distinct role on the team, and in which the members strive to achieve a common goal as a collective” (O’Neill et al., 2002). To be considered autonomous, the agents must meet at least partial levels of self-directed behavior (agency) and agent autonomy. O’Neill and colleagues (2022) provide a framework for three levels (no, partial and high) of agent autonomy (cf. Table 1). It builds on the concept of levels of automation (LOAs) suggested by Parasuraman and colleagues (2000) ranging from low (LOA 1: The computer offers no assistance human have to take all decisions) to high (LOA 10: The computer decides everything, act autonomously, ignoring the human). No agent autonomy encompasses LOA one to four, where dynamic and / or adaptive agents either offer no assistance or provide suggestions and increasing decision support, and the human team member can, but does not have to, decide and act according to the agents’ suggestions. Partial agent autonomy encompass medium levels of autonomy (LOA 5-6), where adaptive agent suggest a decision and executes it if the human approves it or allow the human team members to veto the adaptive agents’ decisions for a specific amount of time before they act autonomously. High agent autonomy encompasses LOAs seven to ten where adaptive agent act autonomously either informing humans necessarily or if asked or if adaptive agent decide to or ignore humans at all. In this regard, researchers distinguish between adaptive and adaptable adaptive agent: An AI that automatically changes its level of autonomy is an adaptive agent, while an adaptive agent whose LOA is set and adjusted by a user is an adaptable system (Calhoun, 2022). In this study we investigate adaptive agent that automatically changes its LOA based on defined performance and stress criteria of the human without the possibility of a human’s intervention to the decision. The aim of the study is to investigate the influence of agent knowledge and perceived mental model congruence on explicit coordination needs and attitudes (trust, acceptance) towards this adaptive agent. Within a computer simulation (Timm et al., 2022) humans must process incoming emergency calls and decide rescue stations, vehicles, and response teams along requirements and availabilities. The adaptive agent takes over subtasks of the human at different LOA levels by measuring the human's performance indicators as the degree of stress and adaptively and dynamically adjusting its support. Adaptive agent knowledge and perceived mental model congruence refers to individual mental representations but includes distinctive aspects. In this study, adaptive agent knowledge relates to the human understanding of adaptive agent reasoning to decide the support level. Adaptive agent reasoning knowledge includes the understanding of specific input data the adaptive agent collects to monitor stress and performance, and an understanding of how the adaptive agent processes and decides about human performance as indicators of stress to adaptively adjust support. Prior the task phase, humans can have high adaptive agent knowledge with elaborated understanding of input, reasoning and decision processes of the adaptive agent or low adaptive agent knowledge with just an information about the role and goals. Mental model congruence relates to the human awareness during the task if the adaptive support of the adaptive agent is consistent with one's own perceptions of high stress and need for assistance. High mental model congruence indicates a perception that human and adaptive agent have a similar understanding of when human needed more or less support because of overload and stress. Adaptive agent knowledge and perceived mental model congruence may influence different coordination needs and attitudes towards the adaptive agent during action phase and afterwards during transition phase. Need for explicit coordination refers to communication desires during the task (need for pull/push coordination during action phase) and the need to reflect the human-agent interaction after the task (need for reflection during transition phase). Moreover, adaptive agent knowledge and perceived mental model congruence may influence perceived agent trust and acceptance. Trust relates to the feeling of being able to fully rely on the decisions of the agent. Agent acceptance relates to humans evaluation of whether the agent meet expectations with regard to performance goals and efforts invested.

Recent research has described the phenomenon of human-autonomy or human-agent teams (HATs). O’Neill and colleagues (2022) define HATs as “involving one or more humans and one or more autonomous agents, wherein each human and autonomous agent is recognized as a unique team member occupying a distinct role on the team, and in which the members strive to achieve a common goal as a collective”. O’Neill et al. (2002) describe autonomous agents by a degree of self-government, self-directed behavior (agency), and taking on a unique role or set of tasks. Thus, humans and agents can work on subtasks independently but have to collaborate interdependently to achieve common team goals (O’Neill et al. 2022). To be considered autonomous, the agents must meet at least partial levels of self-directed behavior (agency) and agent autonomy. O’Neill and colleagues (2022) provide a framework for three levels (no, partial and high) of agent autonomy. It builds on the concept of levels of automation (LOAs) suggested by Parasuraman and colleagues (2000) ranging from low (LOA 1: The computer offers no assistance human must take all decisions) to high (LOA 10: The computer decides everything, act autonomously, ignoring the human). No agent autonomy encompass LOA one to four, where software agents (SA) either offer no no assistance or provide suggestions and increasing decision support, and the human team member can, but does not have to, decide and act according to the SA's suggestions. Partial agent autonomy encompass medium levels of autonomy (LOA 5-6), where SA suggest a decision and executes it if the human approves it or allow the human team members to veto the SA’ decisions for a specific amount of time before they act autonomously. High agent autonomy encompass LOAs seven to ten where SA act autonomously either informing humans necessarily or if asked or if SA decide to or ignore humans at all. Although, these LOAs refer to decision making and executing, Wickens et al (1998) proposed that also information acquisition and analysis as prior phases of information processing can be differentiated accordingly. Sheridan and Verplank (1978) argue that when SA assist humans in the information processing process and extend human capabilities or relieve task demands humans share control, while humans trade control when SA back up or replace them. Thus, sharing control starts at LOA 2 when SA have no autonomy but assist information processing and, for example, suggest decision or action alternatives, while high SA autonomy implies trading control. Furthermore, as in HAT humans and agents work on subtasks independently but have to collaborate interdependently to achieve common team goals, multiple tasks with different levels of agent autonomy and control have to be considered. In their meta-analytic study, Onnasch et al. (2014) report that increasing automation and agent autonomy can lead to routine-failure trade-offs. This trade-off describes that higher LOA yields better HAT performance and lower human workload, but leaving the human team member out of the loop leads to performance impairments compared to manual performance if the software agent drops out. This is also called the “lumber jack effect”, which refers to the falling of forest trees, “the higher they are, the farther they fall.” The results of their post-hoc analysis indicates the more the software agent decides on, selects, and implements actions, the lower the performance (compared to the prior manual performance) if the automation drops out. They conclude that each step of increase of LOA, the risk of lower performance increases due to human forgetting and the loss of routines when returning to manual performance. To minimize this risk, they recommend that human operators should be involved, at least to some degree, in the selection of decisions and actions as well as in the implementation of actions. Onnasch et al. (2014) offer several explanations for these routine-failure trade-offs. One may be task unfamiliarity and loss of manual skills, i.e. due to forgetting and the loss of routines. Another is losing situation awareness as a result of reduced system monitoring. They refer to the generation effect (Slamecka & Graf, 1978) to explain that as long as a person actively chooses his actions based on automated information acquisition and integration by the software agent rather than only accepting a proposed action this supports the return to manual processes. Therefore, they assume that the distinction between situation assessment and action support is critical for the performance decrement. However, the number of studies analyzed was limited and LOA was based on ranking studies. Therefore, the authors call for experimental research to confirm their findings. The underlying mediating information processing is also not very well elaborated. Whether the performance drop after the return to manual processes is due to the supposed generation effect or due to task unfamiliarity and forgetting or losing situation awareness as a result of reduced system monitoring remains unclear. Furthermore, comparing apriori to aposteriori performance, task knowledge and work load instead of comparing relative effects of LOA interventions with a no LOA condition do not take into account ongoing learning processes during experimental tasks, which may not yet be completely routinized. To sum up, neither the conditions nor the underlying processes of the “lumber jack effect” are clear and the empirical evidence is yet scare. Contribution This study contributes to research firstly by comparing the effects of different LOA on performance and workload within HATs and relative performance and task knowledge decrements, and workload increase when returning to manual performance. Secondly, it clarifies the underlying processes by analyzing the effects of task familiarity, active human information processing (generation effect) and system monitoring. Thirdly, it considers multiple tasks with different levels of agent autonomy and control.

To provide implications whether insights of human teamwork research is transferable to human-agent collaboration, this study investigates the effects on three different levels of autonomy on aspects characterizing autonomous agents (i.e., agents’ agency, perception as a mate or a tool) and human-autonomy teams (task interdependence, uniqueness of roles, team identification). Research Question: At what level of autonomy does a human perceive an autonomous agent as a teammate and their collaboration as a team?

The FCI (Fire, Crime, and Injury) is a distributed platform for analyzing collaboration between humans and intelligent agents. The FCI is a serious game in which players must cooperate to process and control emergency operations. Players can take on the roles of the fire department, police, paramedics, or control center. An agent can automate individual tasks or entire roles. Furthermore, it is possible to vary the degree of interdependence of the tasks in team experiments. For psychological team research, FCI offers an experimental platform to investigate team states and processes in Human-Agent-Collaboration. For artificial intelligence research, FCI provides the opportunity to generate data on human decision-making behavior and contributes to the extension of cognitive agent models.

Research Question: 1. Does meta-knowledge lead to lower interaction-related activities (i.e., searching information of “who knows what”) and, in turn, to better team performance? 2. Which teams (e.g., high vs. low meta-knowledge) under which environmental conditions (e.g., with which instruction a software agent is implemented) benefit the most from automated interaction among team members? 3. Which teams (e.g., high vs. low meta-knowledge) under which environmental conditions (e.g., with which instruction a software agent is implemented) adapt best to a software agent failure?

Research Question: 1. Does the automation of team interaction processes influence social presence and activity awareness of other team members? 2. Does the automation of team interaction processes influence team commitment and willingness to do additional work? 3. Are social presence and activity awareness of other team members associated with team commitment and willingness to do additional work?

In a work environment where humans and digital agents work together interdependently, humans must trust and accept the digital agent (O’Neill et al., 2020). Following the definitions of technology acceptance models, digital agent acceptance is indicated as high effort expectancy (the digital agent does not lead to additional work) and performance expectancy (the digital agent meets the team goals). Usability research shows, that knowledge about a digital tool is relevant for user acceptance (Dockweiler et al., 2015; Dockweiler & Hornberg, 2017). Moreover, user knowledge about the digital tool is essential to calibrate trust in that system (Khastgir et al., 2018). However, research on individual knowledge representations about digital agents is scarce. There is first evidence that knowledge about the specific activities of a digital agent is beneficial with regard to motivation and performance regarding the activity of team members (Haines et al., 2021). Moreover, digital agents and humans should share a mental map of the task and team roles (Fan & Yen, 2011). The development and relevance of individual knowledge about the digital agents' roles and tasks has not been explored so far. In this study, we want to investigate the influence of digital agent-related role task knowledge on agent trust and acceptance before and after a first interaction. It is expected that high digital agent-related role and task knowledge will relate to higher agent trust and acceptance before and after the task. Moreover, the study explores the development of role task knowledge during the task. Consistent with usability and team research (e.g., Vekantesh et al., 2016, Haines et al., 2021) user experiences will influence knowledge development over time. The prerequisite for this is, however, the awareness of task performance of the digital agent which is in many cases rarely physically visible but performs task operations in an opaque manner. In this study we postulate that digital agent activity awareness will moderate the development of role-task knowledge. More specifically, higher visibility will lead to higher increases of role-task knowledge between pre- and post-action condition. The study consists of a laboratory experiment in which teams of three people work on an interdependent task. In a first step, the digtal agent is introduced and participants are informed about its basic functions and answer questionnaires (T1, pre-action). During the interdependet task (action), which is played on the FCI, participants are supported by the digital agent, which takes over some of their tasks. After the task, participants again answer the questionnaires (T2, post-action).