Beyond AI-powered context-aware services: the role of human–AI collaboration

Abstract

Purpose

Artificial intelligence (AI) has gained significant momentum in recent years. Among AI-infused systems, one prominent application is context-aware systems. Although the fusion of AI and context awareness has given birth to personalized and timely AI-powered context-aware systems, several challenges still remain. Given the “black box” nature of AI, the authors propose that human–AI collaboration is essential for AI-powered context-aware services to eliminate uncertainty and evolve. To this end, this study aims to advance a research agenda for facilitators and outcomes of human–AI collaboration in AI-powered context-aware services.

Design/methodology/approach

Synthesizing the extant literature on AI and context awareness, the authors advance a theoretical framework that not only differentiates among the three phases of AI-powered context-aware services (i.e. context acquisition, context interpretation and context application) but also outlines plausible research directions for each stage.

Findings

The authors delve into the role of human–AI collaboration and derive future research questions from two directions, namely, the effects of AI-powered context-aware services design on human–AI collaboration and the impact of human–AI collaboration.

Originality/value

This study contributes to the extant literature by identifying knowledge gaps in human–AI collaboration for AI-powered context-aware services and putting forth research directions accordingly. In turn, their proposed framework yields actionable guidance for AI-powered context-aware service designers and practitioners.

Citation

Jiang, N., Liu, X., Liu, H., Lim, E.T.K., Tan, C.-W. and Gu, J. (2023), "Beyond AI-powered context-aware services: the role of human–AI collaboration", Industrial Management & Data Systems, Vol. 123 No. 11, pp. 2771-2802. https://doi.org/10.1108/IMDS-03-2022-0152

1. Introduction

Artificial intelligence (AI) denotes the “ability of a machine to perform cognitive functions that we associate with human minds, such as perceiving, reasoning, learning, interacting with the environment, problem solving, decision-making, and even demonstrating creativity” (Rai et al., 2019, p. 1). Advances in AI have given rise to its growing application in daily life in the likes of personal assistants and self-driving systems. According to a report from the McKinsey Global Institute, it is estimated that around 70% of enterprises would embrace at least one kind of AI technology by the year 2030 [1]. According to market research conducted by Tractica, a marketing research firm, the revenue of the global AI software market is expected to increase from around US $10.1 billion in 2018 to around US $126.0 billion by 2025 [2].

Among AI-infused systems, one prominent application is context-aware systems, designed to deliver tailored and timely services based on users' context (Hong et al., 2009). Notably, the fusion of AI and context awareness has given birth to intelligent AI-powered context-aware systems. For example, with the aid of AI, smartphones are becoming increasingly proficient in delivering tailored recommendations to their users about the applications they might require based on the latter's immediate context and usage habits (Sarker et al., 2020). But at the same time, even though AI possesses the capability to comprehend contextual information and formulate appropriate recommendations, its complexity and unpredictability could, at times, culminate in “behaviors that can be disruptive, confusing, offensive, and even dangerous” (Amershi et al., 2019, p. 1). Because it is challenging to assess the reliability of judgments from AI-powered context-aware systems (Kaur et al., 2021), scholars have alluded to the importance of incorporating human–AI collaboration in AI-powered context-aware systems to bolster their accuracy and reduce the damage caused by such systems (Kaur et al., 2021). Human–AI collaboration describes the process of developing a socio-technical system in which AI and humans can collaborate in a mutually beneficial fashion by exploiting individual strengths and coevolving through complementary improvement (Loske and Klumpp, 2021). Individually, human–AI collaboration might eliminate uncertainty when context-aware systems collect and interpret contextual data, culminating in enhanced performance of AI-powered context-aware services.

Despite past studies acknowledging the significance of human–AI collaboration (Jarrahi, 2018; Puranam, 2021; Ramchurn et al., 2021), related research is still in its infancy. To address the knowledge gap, we conducted a thorough literature review on human–AI collaboration and extracted 51 relevant articles from the international academic database: EBSCO Host Research Databases (see Appendix). Our review of published articles on human–AI collaboration points to two research streams: human–AI collaboration cultivation and human–AI collaboration outcomes. Whereas the research stream on human–AI collaboration cultivation concentrates on deciphering the conditions under which human–AI collaboration emerges and can be facilitated, its counterpart on human–AI collaboration outcomes places emphasis on the consequences of human–AI collaboration. But at the same time, prior research on human–AI collaboration cultivation either remains at a conceptual level or is confined to a narrow set of human–AI collaborative tasks such as cocreation of drawings (Oh et al., 2018), thereby generating limited insights into actionable design guidelines for human–AI collaboration in AI-powered services. Likewise, published articles on human–AI collaboration outcomes are rather fragmented, rendering it difficult to arrive at a holistic picture of the consequences arising from such collaborative actions.

In light of the current state of research on human–AI collaboration, we endeavor to bridge the two abovementioned research streams by delineating system facilitators across distinct phases of AI-powered context-aware services and producing a holistic view of human–AI collaboration outcomes. Specifically, we strive to answer two research questions (RQs):

RQ1.

How does the design of AI-powered context-aware services influence human-AI collaboration?

RQ2.

What are the impacts of human-AI collaboration?

To answer the two RQs, we delve into the role of human–AI collaboration to advance a future research agenda from two directions, namely, the effects of AI-powered context-aware services design on human–AI collaboration, as well as the impact of human–AI collaboration on individuals and organizations. To do so, we delineate AI-powered context-aware services into three stages (i.e. context acquisition, context interpretation and context application) before elucidating the factors to be considered in each stage following which, the consequences of human–AI collaboration together with future research directions are then discussed accordingly.

2. Theoretical foundation

2.1 Context-aware services

The rapid growth of mobile technologies combined with advancements in the “Internet of Things” has enabled context-aware services to become one of the emerging technologies in recent years (Perera et al., 2014). A context-aware system is a system that uses context to provide relevant information and/or services to the users, where relevancy depends on the users' task (Dey et al., 2001). Consistent with Dey et al. (2001), we embraced the commonly recognized definition of “context” that states that it is “any information that can characterize the situation to an entity that is considered relevant to the interaction between the user and the application. An entity is a person, place, or object that is considered relevant to the interaction between user and the application, including the user and applications themselves (p. 100)” (Dey et al., 2001). In a nutshell, context-aware services recognize and interpret users' aspects of their context (e.g. users' location, current time, the type of device and weather) and deliver services intelligently based on users' needs (Liu et al., 2016). Table 1 outlines prior studies that attempted to explain the process of developing context-aware services (Bernardos et al., 2008; Fischer, 2012), as well as our partition.

For instance, Bernardos et al. (2008) identified three phases in a context management system: context acquisition, information processing, and reasoning and decision. Similarly, Fischer (2012) proposed three steps when they developed a framework for context-aware systems: context discovery and acquisition, context representation and management, and context utilization. Although there are some nuances, the core phases of delivering context-aware services lie in three steps: context acquisition, context interpretation and context application, as depicted in Figure 1 below:

To elaborate on three stages for AI-powered context-aware services, we provide the following scenario as a running example in the remainder of the paper:

Sebastian is a 28-year-old man who lives alone and enjoys an intelligent home system in his daily life. He always gets up at 7:30 am and then spends around 15 minutes in the bathroom. The alarm goes off a few minutes before he gets out of bed automatically. After a while, the smart coffee maker would brew a cup of coffee with his favorite flavor once it detected that he had come out of the bathroom. Before Sebastian heads out the door, the smart closet could recommend the proper outfits based on his preference and the weather outside. The system could also turn on the air conditioner in advance when it concludes that Sebastian is coming home from work based on the GPS data. On weekends, if there is no urgent meeting derived from the calendar, the system would not set an alarm. Apart from managing his house more efficiently, the intelligent system also has the ability to help Sebastian stay healthy. When he falls asleep, the bed with sensors could monitor his breathing, heart rate, body temperature, and pressure level, thus adjusting the mattress firmness and temperature of the air conditioner.

Context acquisition refers to the process of acquiring the required contextual information. To provide the right services at the right time, in the right place, in the right way, collecting contextual information from physical, logical or virtual sensors is essential for context-aware systems (Yurur et al., 2016). Mobile devices are equipped with various physical sensors like temperature sensors, GPS, accelerometers, pressure sensors and microphones, so context like location, speed and temperature is easily accessible for users. Virtual sensors provide opportunities to collect higher-level contextual information such as calendar details and data derived from social media. In addition, logical sensors fused data from physical sensors and virtual sensors to infer complicated user statuses like activity recognition (Yurur et al., 2016). Let us consider the scenario we mentioned earlier. To ascertain Sebastian's exact location, physical sensors like GPS and accelerometers are employed in the context acquisition stage to judge whether there is motion and his exact location. Information from his calendar is also helpful for setting the alarm after checking his schedule.

Once the required contextual information is collected, the next step for a system is context interpretation, in which the system models the context and derives insights from statistical approaches. After gathering context derived from various sources, the system needs to reconcile them and produce a coherent representation for information processing in the next step. Bernardos et al. (2008) contented that this stage aimed at “preparing the informational picture to provide the system with enough information to reason about” (p. 612). In Sebastian's case, a large number of sensors are built into the intelligent home system. The core of the system lies in how to interpret data from multiple sources. For example, the system receives context like “his location is the kitchen”, “current time is 8:00 am” and “there is no motion in the kitchen”. These contextual details are then represented as factors and parameters for the reasoning algorithms. Within the extant literature, there are several context modeling techniques, such as key-value modeling and ontology-based modeling (Sezer et al., 2018). The intelligent system may infer Sebastian's desire to have breakfast based on observations like “Sebastian is static in the kitchen at 8:00 am”. Alternatively, he may read online news in the kitchen instead. Combined with motion recognition technologies from the camera, the system could gain improved confidence in identifying Sebastian's statuses and intentions, thus delivering personalized services.

The final stage is context application, which we define as the process of discovering, adapting and personalizing context-aware services for users based on contextual data (Grifoni et al., 2018). At this stage, a context-aware system fulfills its mission by developing personalized services for end users in terms of what services to deliver, when to deliver and how to deliver them. Considering the alarm setting in Sebastian's scenario, when Sebastian encounters unexpected things, he may set the alarm by himself, and the system will deliver service reactively. Yet, if Sebastian forgets to set an alarm and the system detects daily routines from his calendar, the alarm may go off proactively. On top of that, a context-aware system also needs interpretation of user feedback to foster the evolution of context awareness. For instance, when Sebastian ignores recommendations, the ignoring behavior could be interpreted as negative feedback. Following feedback, reinforcement learning could modify user profiles and even modify parameters in the context acquisition stage in the next round (Augusto et al., 2017).

2.2 AI-powered context-aware services

Context-aware services are increasingly fueled by AI. For instance, Sarker et al. (2020) built a smartphone apps’ prediction model based on one of the unsupervised machine learning techniques, namely, principal component analysis that could reduce multiple contextual dimensions. In this vein, smartphones could provide recommendations on what apps are needed based on a user's current context and usage history proactively. Ample research has alleged that AI plays an instrumental role in promoting efficiency and performance for context-aware systems (del Carmen Rodríguez-Hernández and Ilarri, 2021; Unger et al., 2020). Nonetheless, only a small amount of preceding research has gone into detail on the specific role that AI plays in context-aware systems. This research thus attempts to outline how AI affects three phases of context acquisition, context interpretation and context application for context-aware systems, respectively.

On the stage of context acquisition, contextual factors that are taken into account by context-aware systems are growing more sophisticated and go much beyond simple considerations like geographic location or time (van Engelenburg et al., 2019). AI techniques have the capability to broaden the sources of users' contextual information. For example, it is plausible that visual AI technologies could translate photos and videos into insights and extract contextual information (Sarker, 2022). Apple has acquired Emotient, Inc., a business that analyzes facial expressions to understand an individual's emotions using AI technologies (Bastug et al., 2017). Apart from getting raw data from sensors and user input directly, loosely related information could be fed into machine learning algorithms to infer uses' context. In Sebastian's scenario, a smart kitchen with AI capabilities could infer his health conditions by collecting loosely related information like how much heat he generates in the kitchen at certain times of the day, the types of containers in the refrigerator and types of kitchen waste. Furthermore, technological development on the Internet of Things allied with the incorporation of AI has led to the coordination of a large-scale network of devices (Arsénio et al., 2014) that which might broaden the collective level context. In sum, AI capabilities could broaden the sources of contextual information available to the system at the phase of context acquisition.

It is widely acknowledged that AI techniques are suitable for rule mining and context inference when it comes to the phase of the process known as context interpretation (Almagrabi and Al-Otaibi, 2020; del Carmen Rodríguez-Hernández and Ilarri, 2021). For example, to locate similar users, the context-aware recommender system in the study by Amoretti et al. (2017) utilized k-means clustering algorithms. Besides, Coppola et al. (2009) developed an architecture for context-aware services, in which AI techniques (e.g. rule-based systems, Bayesian networks and ontologies) were utilized to infer more abstract context from more concrete data. Similarly, as time evolves, users' habitual patterns could be recognized through the architecture embedded with AI. In Sebastian's case, the fact that he consistently wakes up at 7:30 a.m. does not indicate that the alarm clock rule is unbreakable. Through continuous learning of his behaviors, the system may discover that he rises 30 min earlier each Monday due to the weekly meetings.

On the final stage of context application, AI techniques not only provide users with personalized context-aware information/services, but also optimize the timing and manner of delivering context-aware information/services. Two aspects are highlighted in Motti et al. (2012)'s road map for how machine learning algorithms support the process of context-aware adaptation: context inference and presentation of context-aware services. If a user consistently resizes the interface of a recommender system, for instance, the system may evolve its presentation style and adjust the default size correspondingly. In the same vein, Sebastian's smart house with AI tracks and detects his behavioral patterns and lifestyles in the long term, hence presenting distinct services for weekdays and weekends. On weekdays, the coffee maker brews a cup of coffee automatically, but the homeowner prefers to prepare breakfast himself on weekends.

2.3 Human considerations for AI-powered context-aware services

The incorporation of AI capabilities into context-aware systems has been shown to be helpful (Coppola et al., 2009; Unger et al., 2020), yet context-aware systems still face a number of obstacles (Kaminskas and Ricci, 2012; Verbert et al., 2012). To elaborate on potential challenges faced by AI-powered context-aware services, we examined three stages of context-aware service delivery (i.e. context acquisition, context interpretation and context application), respectively.

Since AI capabilities allow the extraction of many sources of user context during the context acquisition phase, mismatches between distinct sources may threaten the contextual information correctness. For example, a system might simultaneously detect the user's movements from a wristwatch and the static conditions from a mobile phone. If the user does not provide further information, it is difficult to reach a conclusion about users' statuses. Therefore, it would be beneficial if users were able to validate their context under certain circumstances. This would reduce high ambiguity and signal the relevance and value of contextual information.

Whereas on the stage of context interpretation, the black box nature of AI algorithms might lead to user confusion about the learning processes, such that they have no clue why they are receiving certain context-aware services. Luo et al. (2020) claimed that despite context-aware systems becoming more and more intelligent, they are prone to bugs. The gap between raw sensor data and relevant interpretation of context may impair the provision of subsequent context-aware services. In this case, allowing the user to reedit the contextual data boosts AI interpretation and provides a more customized user experience. The running mode of a music recommender, for instance, may generate playlists that correspond to your running pace. Yet, guidelines, which are intended to apply to everyone, such as “recommend fast-paced music as running speed rises”, may not be appropriate for certain users, particularly those who like listening to tranquil music when they are exercising at a high intensity. By updating the context interpretation logic and fixing the interpretation errors, it will be possible for users to supplement AI-powered context-aware systems in an effective manner. To minimize possible mistakes, prompt feedback from users is less costly and more effective for evolving the context-aware adaptation engine.

Previous research has recognized the role of human augmentation, such as providing relevant information to AI-powered systems (Andrews et al., 2022; Metcalf et al., 2019). Apart from human augmentation, it is acknowledged that both humans and AI have their own unique strengths (Ramchurn et al., 2021). AI also has the ability to augment humans in three stages for context-aware services. At the phase of context acquisition, AI methods could be utilized to add another information source (e.g. image analysis), with the goal of reducing the ambiguity about users' context. AI-based systems might develop user-specific adaption mechanisms throughout the stage of context interpretation. Besides, AI-based systems may provide users with a confidence level for outcomes in the context application stage that would assist them to accept information or services they receive. In particular, humans have the capacity to process information intuitively and with social skills, while AI excels in speed and scalability. In a similar vein, Miller (2018) highlighted that humans and machines should work symbiotically to augment each other's capabilities. Moreover, Woo (2020) asserted that human–machine cocreation is vital for the success of the AI-oriented systems that enhance their capabilities separately. Hence, as context evolves with time dynamically (Mishra et al., 2019), a hybrid of human–machine effort for human–AI collaboration is imperative in AI-powered context-aware services for personal life or organizational business. We summarized the role of human–AI collaboration in Table 2 below.

2.4 Human–AI collaboration for AI-powered services

Recent advancements in AI could make it possible for people and technology to work together and integrate seamlessly (Korteling et al., 2021). As a result, more research has started to investigate the potential for AI and humans to work together in collaborative settings (Andrews et al., 2022; Nakahashi and Yamada, 2021; Peeters et al., 2021). Metaphors such as “teammates,” “partners” and “collaborators” in related studies highlighted a high degree of cooperation, complementarity and equality for human–AI collaboration (Korteling et al., 2021). In particular, to gain a comprehensive understanding of human–AI collaboration from the extant literature, we conducted a thorough literature review on human–AI collaboration through the EBSCO database with all subjects. We discovered the initial corpus of peer-reviewed academic studies using the terms “human-AI,” “AI-human,” “Artificial Intelligence-human” and “human-Artificial Intelligence” at the EBSCO database, yielding 520 studies. We then excluded papers that did not pertain to AI-based systems based on the abstracts of publications, resulting in 252 studies. Following that, we examined the identified papers and eliminated publications that were unrelated to human–AI collaboration. In the end, we located 51 relevant articles in total (see Appendix).

Our review of human–AI collaboration reveals two research streams: human–AI collaboration cultivation and human–AI collaboration outcomes. The human–AI collaboration cultivation research stream is centered on facilitating human–AI collaboration. A significant portion of the studies in the human–AI collaboration cultivation research stream proposed conceptual premises for human–AI collaboration that range from shared mental models (Andrews et al., 2022) to collective responsibility (Cañas, 2022) and calibrated trust (Okamura and Yamada, 2020). In addition, a minority of researchers are working on developing a prototype or framework to deploy human–AI collaboration. For instance, Strobelt et al. (2022) designed a prototype, which visualizes a more comprehensive picture of the AI system and enables users to modify model constraints, with the goal of enhancing human–AI collaboration. However, the existing body of research is restricted to relatively specific tasks and fragments, or focuses more on the conceptual foundations of human–AI collaboration, thereby providing little insight into a comprehensive and actionable design of AI-powered services for human–AI collaboration. While for the human–AI collaboration outcomes research stream, it endeavors to unearth the outcomes derived from human–AI collaboration for individuals, organizations and society. For example, Sturm et al. (2021) examined how businesses may combine human learning and machine learning to improve organizational learning. Yet, the fragmented structure of this field's studies makes it difficult to produce an integrated picture of the phenomena.

3. Research directions

Conceivably, human–AI collaboration is prominent for AI-powered context-aware services. Yet, to the best of our knowledge, no research has examined the role of human–AI collaboration for context-aware systems that necessitates an actionable and holistic framework for designers seeking to combine AI and context-awareness technologies and enhance human–AI collaboration. In line with the prior literature (Peeters et al., 2021; Schoonderwoerd et al., 2022), we define human–AI collaboration as the process of developing a socio-technical system in which human intelligence and AI can effectively collaborate by exploiting their individual strengths and coevolve via mutual improvement (Loske and Klumpp, 2021). A higher degree of human–AI collaboration indicates that users are willing to promote the performance of AI by integrating their own intelligence; and the AI-enabled system keeps learning users to augment human intelligence at the same time. Following two research streams in the domain of human–AI collaboration, we advance a framework around human–AI collaboration in AI-powered context-aware services which includes two facets: (1) How does the design of AI-powered context-aware services at different stages influence human–AI collaboration? (2) What are the impacts of human–AI collaboration? We developed a theoretical framework to illustrate possible RQs as depicted in Figure 2 below.

3.1 Effects of the design of AI-powered context-aware services on human–AI collaboration

Table 3 summarizes the salient AI-related topics for the design of AI-powered context-aware services at different stages and the future RQs they may address. The RQs are identified from three perspectives: design, mechanisms and boundary. The “design” viewpoint seeks to explore challenges that are centered on the design of AI-powered context-aware services, such as “how to design explanation approaches for human-AI collaboration in AI-powered context-aware systems?” while the “mechanisms” perspective focuses on the underlying mechanisms of relationships between system design and human–AI collaboration, such as the question, “how do explanation approaches influence human-AI collaboration in AI-powered context-aware systems?” In addition, the “boundary” perspective would pay attention to the moderating effects of the relationships in terms of how system design affects human–AI collaboration. For instance, “What user, system, and task characteristics moderate the effects of explanation approaches on human-AI collaboration in AI-powered context-aware systems?” Specifically, we propose the following RQs.

3.1.1 How does the design of AI-powered context-aware services in the phase of context acquisition influence human–AI collaboration?

During the phase of context acquisition, the context-aware system tries to collect users' personal context that identifies their identity, either by detecting it implicitly or by requesting it from the user explicitly. Specifically, users' preferences and contexts may be gleaned from a wide range of unrelated data that are collected by AI-powered systems that might easily trigger privacy concerns in this regard. Simultaneously, the system might have more accurate user profiles for context-aware services if users expressly designate their context. On the one hand, the system tried to prevent bothering consumers too much. On the other hand, user disclosure helps to ensure that user profiles are accurate for customization. We thus believe that independence, which measures the extent to which tasks on an AI-powered context-aware system are dependent on the system without human intervention, plays a significant role in this phase. We figure that human–AI collaboration may be influenced by the degree of independence between humans and AI when they engage in interactions. Few studies have examined the underlying mechanisms of the impact. We hence propose the following RQs linked to the mechanisms and boundaries of the impacts of independence on human–AI collaboration:

1. How does the system independence level influence human–AI collaboration in AI-powered context-aware systems?

2. What user, system and task characteristics moderate the effect of system independence on human–AI cooperation in AI-powered context-aware systems?

AI-powered systems should determine their acceptable degree of independence in the phase of context acquisition that implies a trade-off between AI participation and user involvement. It might be argued that privacy-sensitive contextual information, such as health problems, is better off being given by the user, yet certain contextual information could be automatically gathered by the system. For example, AI-powered context-aware systems may find it challenging to identify users' affective and social cues. Because of this, it would be preferable if we tilted the balance toward user involvement and AI participation. In that case, we propose the following RQs regarding the boundaries for system independence:

1. How to identify the boundary between AI and humans during the process of context acquisition in AI-powered context-aware systems?

2. Under what circumstances is a system's collection of contexts automatically better than a user's providing their context in AI-powered context-aware systems?

Users may be reluctant to share their context in order to prevent privacy concerns or disruptive user experience during their interactions. Nevertheless, if users believed that they had a shared goal with the system, they might be inspired to cooperate with the AI-powered context-aware systems. In addition, both users and AI might contribute contextual information at the same time; thus, it is vital to study ways to coherently combine data from AI and users. As a result of this, we propose the following RQs regarding the design of system independence:

1. How to motivate users to provide context information in synergy with the AI-powered context-aware systems?

2. How to integrate contextual data from AI and users coherently for AI-powered context-aware systems?

3. What principles should be used to identify the ultimate context for AI-powered context-aware systems when user data and AI data do not match?

3.1.2 How does the design of AI-powered context-aware services in the phase of context interpretation influence human–AI collaboration?

Ample prior research has recognized the role of explainability for human–AI collaboration (Naiseh et al., 2021; Ramchurn et al., 2021; Zerilli et al., 2022). As contended by Naiseh et al. (2021), “human-AI collaborative decision-making tools are being increasingly applied in critical domains such as healthcare … An essential requirement for their success is the ability to provide explanations about themselves that are understandable and meaningful to the users (p. 1857)”. Notably, to promote the performance of AI, explainable artificial intelligence (XAI), which aims to present understandable justifications of output or procedures to end users, has been seeing explosive growth over the last several years (Holzinger et al., 2021; Shin, 2021). During the period of context interpretation, AI algorithms are always seen as “black-boxes”, so users might have no clue how their contextual information is interpreted. This may deter users from cooperating with AI-based systems. Consequently, in accordance with Naiseh et al. (2021), we propose that the degree of explainability of a context-aware system should be considered for human–AI collaboration during the context interpretation phase.

In accordance with the definition of “interpretability” stemmed from Doshi-Velez and Kim (2017), we define explainability of a context-aware system as the capability of a context-aware system to explain its processes or outcomes in an understandable way to users. Users may have faith in AI if a context-aware system provides features that make the underlying algorithms transparent to end users. We thus consider the following RQs regarding mechanisms and boundaries of explainability on human–AI collaboration:

1. How do explanation approaches influence human–AI collaboration in AI-powered context-aware systems?

2. What user, system and task characteristics moderate the effect of explanation approaches on human–AI collaboration in AI-powered context-aware systems?

Adadi and Berrada (2018) summarized four motivations for explainability: explain to justify, explain to control, explain to improve and explain to discover, showing possibilities for different explanation approaches. Similarly, Lim et al. (2009) suggested that user experience at context-aware systems would vary depending on the kind of explanation the system provided – why explanations (Why did the system do something?) or why not explanations (Why did the system not do something?). It is still unclear how to achieve human–AI collaboration on a context-aware system via the configuration of various explanation features, or how these varied explanation features facilitate human–AI cooperation in various ways. All of these questions in terms of the design of explanation features for human–AI collaboration are interesting directions for future research:

1. How to design explanation approaches for human–AI collaboration in AI-powered context-aware systems?

2. Do various explanation approaches (“why” and why not” explanations) influence human–AI collaboration differently in AI-powered context-aware systems?

3. How can the quality of the explanations be evaluated in terms of human–AI collaboration in AI-powered context-aware systems?

Scrutability of a context-aware system allows users to correct system assumptions and adjust the model of context interpretation (Pu et al., 2012). Based on Cheverst et al. (2005), we define scrutability as the ability of a context-aware system to enable “a user to interrogate her user model in order to understand the system's behavior” (p. 236). For instance, Setten et al. (2004) designed a recommender system that enables users to indicate the type of context for generating recommendations. Due to the inherent mismatch between how context-aware systems represent the physical world and how users perceive the physical world, the growing level of scrutability augments users' control toward the context-aware system, thereby fostering human–AI collaboration. However, despite scrutability features have been discussed in the prior literature (Balog et al., 2019), its role on human–AI collaboration still remains unclear. Several directions for future research remain. For example, how to design features to promote the level of scrutability at context-aware systems? Does the level of scrutability impact human–AI collaboration via control beliefs or other psychological mechanisms? We thus propose the RQs like as follows:

1. How to design scrutability features for human–AI collaboration in AI-powered context-aware system?

2. When users can tune their own user models, and how do we set and measure the responsibilities of AI and humans in AI-powered context-aware systems?

3. How do scrutability features influence human–AI collaboration in AI-powered context-aware system?

4. What user, system and task characteristics moderate the effects of scrutability features on human–AI collaboration in AI-powered context-aware systems?

3.1.3 How does the design of AI-powered context-aware services in the phase of context application influence human–AI collaboration?

Users may voice their opinions on the suitability of information and services as well as the manner in which they are presented and delivered when it comes to the phase of context application for an AI-powered context-aware system. Practically, e-commerce mobile applications such as Amazon allow users to tap the “like” or “dislike” button next to recommendations while they browse the product list. Furthermore, a growing number of studies have begun to investigate the role of interactive feedback for AI-based systems. For instance, Cai et al. (2019) have validated that interactive dynamic feedback would be able to augment human intelligence during decision-making in a deep neural network (DNN)-backed content-based image retrieval (CBIR) system. In their context, doctors employ a medical image (e.g. x-ray images) to search for similar images to support their diagnosis results from an AI-powered system. After receiving imperfect results, doctors have the ability to leverage three refinement tools to express their feedback: refine-by-region (allow users to select a region of interest), refine-by-example (allow users to set an example for retrieving) and refine-by-concept (allows users to specify a clinical concept). Such feedback not only helps users improve their mental models, but it also confirms algorithmic processes in several hypothesis-test cycles, with the potential to strengthen human–AI collaboration. Users are able to convey their varied perspectives on AI-based systems via the use of several feedback tools. For example, users may indicate the probability of their willingness to accept the system output (Shrestha et al., 2019). However, it remains unclear how to design feedback tools to elicit human–AI collaboration on AI-powered context-aware systems, as well as the mechanisms and boundaries of such impacts. We thus propose the following RQs:

1. How to design interactive feedback tools to promote human–AI collaboration in AI-powered context-aware systems?

2. How to design feedback features to promote shared control between AI and users in AI-powered context-aware systems?

3. How to design feedback features to mitigate unpredictable AI behaviors in AI-powered context-aware systems?

4. What user, system and task characteristics moderate the effect of feedback features on human–AI collaboration in AI-powered context-aware systems?

5. How do feedback features influence human–AI collaboration in AI-powered context-aware systems?

Another essential factor that should be taken into account at the context application stage is adaptability that is the capacity of an AI-powered context-aware system to help users “recognize and adapt fluidly to new and dynamic scenario” (Dubey et al., 2020, p. 4). AI-based systems with high adaptability could modify their output in a timely manner when the circumstances changed. A higher level of adaptability assists users in recognizing their shared goals and common grounds with the system. As such, the level of adaptability has a beneficial impact on human–AI collaboration that seems to be a fruitful avenue for future research. We hence propose the following RQs on the mechanisms and boundaries of such impacts:

1. How does the level of system adaptability influence human–AI collaboration in AI-powered context-aware systems?

2. What user, system and task characteristics moderate the effect of system adaptability on human–AI collaboration in AI-powered context-aware systems?

In addition, as the context changes, AI-based systems, which have high adaptability evolve their services to better suit the new circumstances. For example, the smartphone will automatically adjust the brightness of the screen depending on the time of day, and it will lower the quantity of blue light emitted when used in the evening. However, since limited research has explored how to design system adaptability for human–AI collaboration, we thus propose the following RQ:

1. How to design system adaptability for human–AI collaboration in AI-powered context-aware systems?

3.2 Positive impacts of human–AI collaboration on individuals and organizations

Huang et al. (2019) characterized three goals of colearning for human–AI collaboration: “mutual understanding,” “mutual benefits” and “mutual growth”. In line with their study, we divided the impacts of human–AI collaboration into three types. Mutual understanding refers to the ability of learning entities (human or AI) to expect others and to be expected by others. Mutual benefits describes that the system gets superior results which humans or AI cannot achieve alone. Mutual growth refers to the extent to which AI and humans feel they coevolve over time. Table 4 provides the summary of research directions in terms of the positive impacts of human–AI collaboration on individuals and organizations. Human–AI collaboration fosters users' awareness in terms of their shared control and common ground with the system. We assert that how human–AI collaboration influences users' perceptions and decision-making processes deserve further investigations. Consequently, we propose the following RQs regarding the “mutual understanding” of humans and AI:

1. How does human–AI collaboration influence users' perceived AI mental model in AI-powered context-aware systems?

2. How does human–AI collaboration influence user's perceived shared control in AI-powered context-aware systems?

To guide the design of human–machine teaming systems, Mcdermott et al. (2018) described detailed requirements for systems developers. In their research, they proposed the concept of “calibrated trust” for the purpose of autonomous system design that is salient when “technology provides indicators enabling human operators to understand when and how much to trust an automated partner in context (p. 2).” Numerous research has explored the antecedents and outcomes of trust in the information systems (IS) community (Panniello et al., 2016; Seymour et al., 2021). However, due to the characteristics of AI-powered context-aware services, trust toward AI-powered context-aware services has changed significantly. Even a perfect algorithm is unable to determine how a user interprets the physical environment; hence, AI-powered context-aware systems have a tendency to provide services that are fraught with ambiguity. In this vein, users may build varying levels of trust depending on the context in which they are using the services. We thus define calibrated trust as the extent to which users have the ability to decide when and how much to trust the system in AI-powered context-aware systems. To calibrate users' trust, Mcdermott et al. (2018) suggested three requirements, including providing information source, elaborating automation mechanisms and providing confidence level. In addition, when an AI-powered context-aware system is equipped with high transparency and explainability, users are more likely to evaluate whether the intent of the AI is perceived to be aligned with that of the human, fostering positive human–AI affective relationships. Consequently, we propose the following RQs related to “mutual benefits” of humans and AI:

1. How does human–AI collaboration influence users' trust calibration in AI-powered context-aware systems?

2. How does human–AI collaboration influence users' affective relationship with AI in AI-powered context-aware systems?

Moreover, Yang et al. (2020) claimed that creating user-system coevolution and tailored outputs were two challenges that needed to be overcome in order to enhance the user experience in human–AI interaction. Notably, AI-based systems will “grow” as their capabilities evolve, absorbing new insights from fresh data and becoming “better” than ever before. In line with Huang et al. (2019), we defined mutual growth as the extent to which AI and humans feel they coevolve over time. We thus propose a related RQ related to “mutual growth” of humans and AI:

1. How does human–AI collaboration influence users' perceptions of mutual growth with AI-powered context-aware systems?

The use of AI as an IT strategy is becoming more commonplace in businesses. A growing number of studies are beginning to explore how employees in organizations cooperate with AI for their work (Budhwar et al., 2022; McNeese et al., 2021; Pandey et al., 2022). Initially, humans and AI are two distinct entities with different mental models. They started to foster mutual understanding through increased interactions and continuous learning over time. For instance, McNeese et al. (2021) indicate that trust in team members of human–machine teaming is associated with team performance (McNeese et al., 2021). Yet, related research is still in its early stages, with the majority of studies being conceptual in nature. It is unknown how the human–AI team influences the policies, priorities and culture of a business empirically. Notably, employees may be uncertain as to whether AI should be seen as colleagues or leaders.

Moreover, the work of Robert et al. (2020) defined AI audit as “an inspection of the AI's underlying logic, decision criteria, and data sources in an attempt to validate the AI” (Robert et al., 2020, p. 25).The IT sectors in an organization may conduct large-scale computer simulations or code checking in a periodic way. Notably, how to design AI audition for an organization is still an open question. In sum, it is promising to examine these questions related to “mutual understanding” of humans and AI in organizations in the future. For example,

1. How does human–AI collaboration influence an organization's policies, priorities and culture?

2. How does human–AI collaboration influence employees' social roles in an organization?

3. How to design AI audition for human–AI collaboration in organizations?

AI has been utilized to manage employees and complete organizational tasks in various ways in the workplace (Terziyan et al., 2018). For instance, it can prescreen candidates' resumes (Budhwar et al., 2022). AI-enabled chatbots may also assist new employees in settling into their positions. In addition, physical robots are capable of cooperating with workers to complete a task jointly (Tóth et al., 2022). Yet, it is uncertain how human–AI collaboration influences organizational performance. We propose the following RQ in terms of the “mutual benefits” of humans and AI:

1. How does human–AI collaboration improve organizational performance?

Human–AI collaboration has the potential to boost mutual growth of AI-powered context-aware systems and users, allowing context-aware systems to grasp how people reason about the environment, culture and social interactions that they can incorporate into their intelligence. Kaplan and Haenlein (2019) described three kinds of AI-based systems that include cognitive intelligence, social intelligence and emotional intelligence. Specifically, cognitive, emotional and social intelligence are connected to specific cognitive, emotional and social skills that may be learnt and imitated by both humans and AI. There has been little empirical research that investigates the consequences of human–AI collaboration on the cognitive intelligence, emotional intelligence and social intelligence of AI-enabled systems. Hence, we propose the following RQ on the “mutual growth” of humans and AI in organizations:

1. How does human–AI collaboration influence AI-powered systems' cognitive intelligence, emotional intelligence and social intelligence as time evolves in organizations?

3.3 Negative impacts of human–AI collaboration on individuals and organizations

Apart from positive outcomes, it is plausible that negative impacts, such as several ethical challenges existed in AI-powered context-aware services as a result of human–AI collaboration. In particular, as Lepri et al. (2018) indicated, algorithmic results should not be biased, discriminatory or unfair. However, an AI-based resume filtering system developed by Amazon was found biased against women [3]. Due to the fact that the majority of candidates are male, the system leaned toward male candidates as time evolved. AI-based systems learn discrimination from humans, and human–AI collaboration may impose human discrimination on human–AI teams. Research regarding fairness has begun to attract attention (Ferrer et al., 2021). Consequently, we propose the following RQ: How to reduce and correct AI discrimination from AI-powered context-aware systems?

Numerous AI-enabled systems generate services automatically. Due to the system's ability to infer user context from loosely related information and deliver services promptly, users may be uninformed of the reasons why they are getting such services and powerless to stop them, resulting in a loss of control over their usage experience. We propose the RQ: How to reduce and correct perceptions of “loss of control” for AI-powered context-aware systems?

In addition, intelligent AI-powered context-aware services are expected to do complicated tasks on behalf of users. Yet, it may hinder people's cognitive and social skills when they rely too much on technologies that are enabled by AI. We thus propose that “dependence lock-in” seems an undesirable outcome of overreliance on AI-powered context-aware services (Kalimeri and Tjostheim, 2020). The “dependence lock-in” problem means that users may diminish the potential of achieving tasks individually if they rely too much on human–AI collaboration that may in turn degrade users' capacity to make decisions on their own. Hence, if AI-based systems encounter system failures, users might have difficulty in solving several tasks independently. A look at lock-in problems and how to prevent this from happening is promising. We thus propose the following related RQ: How to reduce and correct “dependence lock-in” for AI-powered context-aware systems? In sum, we summarized possible future research directions for possible negative impacts of human–AI collaboration on individuals and organizations in Table 5.

4. Conclusions

AI is of great significance for our daily life. Researchers and practitioners face challenges in designing AI-powered context-aware services. We draw attention to the process of context-aware systems and emphasize the role of human–AI collaboration. In this study, we propose future RQs and theoretical perspectives in terms of two directions: the role of AI-powered context-aware services design on human–AI collaboration and the impacts of human–AI collaboration. Through our discussions, we hope to open up avenues for both theory and practice.

Figures

Figure 1

Stages of context-aware services delivery

Figure 2

Theoretical framework for research directions

Table 1

Developmental phases of context-aware services

Source	Phases of context-awareness process
Bernardos et al. (2008)	Context acquisition	Information processing	Reasoning and decision
Fischer (2012)	Context discovery and acquisition	Context representation and management	Context utilization
Perera et al. (2014)	Context acquisition	Context modeling	Context reasoning	Context dissemination
Grifoni et al. (2018)	Context acquisition	Context representation	Context adaptation
Our partition	Context acquisition	Context interpretation	Context application

Table 2

Role of human–AI collaboration in context-aware systems

Phase	Definition	Role of artificial intelligence (AI)	Challenges	Human–AI collaboration
				Human augment AI	AI augment human
Context acquisition	The process of acquiring the required contextual information for context-aware systems	⁃ Expand context source ⁃ Context inference	⁃ Mismatch among different context sources	⁃ Confirm their context ⁃ Indicate sensibility and importance of the contextual information	Provide information source to users
Context interpretation	The process of modeling the context and deriving insights from statistical approaches for context-aware systems	⁃ Rule mining ⁃ Intention recognition ⁃ Pattern recognition	⁃ Black box nature of AI algorithms ⁃ Prone to bugs	⁃ Revise the previous interpretation logic ⁃ Correct the interpretation error regarding context	Elaborate on adaptation mechanisms to users
Context application	The process of the system delivers context-aware services based on users' needs and profiles for context-aware systems	⁃ Provide personalized information/services ⁃ Optimize timing and the format of delivering information/services	⁃ Trained models may work in narrow scope	⁃ Deliver feedback to appropriateness of the results	Provide confidence level for results to users

Table 3

Design of AI-powered context-aware services on human–AI collaboration

Stage	Research topics	Possible future research questions
		Design	Mechanisms	Boundary
Context acquisition	Independence	⁃ How to motivate users to provide context information in synergy with the AI-powered context-aware systems? ⁃ How to integrate contextual data from AI and users coherently for AI-powered context-aware systems? ⁃ What principles should be used to identify the ultimate context for AI-powered context-aware systems when user data and AI data do not match?	⁃ How does the system independence level influence human–AI collaboration in AI-powered context-aware systems?	⁃ How to identify the boundary between AI and humans during the process of context acquisition in AI-powered context-aware systems? ⁃ Under what circumstances is a system's collection of contexts automatically better than a user's providing their context in AI-powered context-aware systems? ⁃ What user, system and task characteristics moderate the effect of system independence on human–AI cooperation in AI-powered context-aware systems?
Context interpretation	Explainability	⁃ How to design explanation approaches for human–AI collaboration in AI-powered context-aware systems? ⁃ Do various explanation approaches influence human–AI collaboration differently in AI-powered context-aware systems (“why” and “why not” explanations)? ⁃ How can the quality of the explanations be evaluated in terms of human–AI collaboration in AI-powered context-aware systems?	⁃ How do explanation approaches influence human–AI collaboration in AI-powered context-aware systems? (e.g. perceived transparency, user expectations and their shared goal with AI)?	⁃ What user, system and task characteristics moderate the effects of explanation approaches on human–AI collaboration in AI-powered context-aware systems?
	Scrutability	⁃ How to design scrutability features for human–AI collaboration in AI-powered context-aware system? ⁃ When users can tune their own user models, how do we set and measure the responsibilities of AI and humans in AI-powered context-aware systems?	⁃ How do scrutability features influence human–AI collaboration in AI-powered context-aware system (e.g. shared awareness and accountability)?	⁃ What user, system and task characteristics moderate the effects of scrutability features on human–AI collaboration in AI-powered context-aware systems?
Context application	Adaptability	⁃ How to design system adaptability for human–AI collaboration in AI-powered context-aware systems?	⁃ How does the level of system adaptability influence human–AI collaboration in AI-powered context-aware systems?	⁃ What user, system and task characteristics moderate the effect of system adaptability on human–AI collaboration in AI-powered context-aware systems?
	Feedback	⁃ How to design interactive feedback tools to promote human–AI collaboration in AI-powered context-aware systems? ⁃ How to design feedback features to promote shared control between AI and users in AI-powered context-aware systems? ⁃ How to design feedback features to mitigate unpredictable AI behaviors in AI-powered context-aware systems?	⁃ How do feedback features influence human–AI collaboration in AI-powered context-aware systems (e.g. self-awareness)?	⁃ What user, system and task characteristics moderate the effect of feedback features on human–AI collaboration in AI-powered context-aware systems?

Table 4

Positive impact of human–AI collaboration on individuals and organizations

Goal of Human–AI collaboration	Possible future research questions
	Individual level	Organizational level
Mutual understanding	⁃ How does human–AI collaboration influence users' perceived AI mental model in AI-powered context-aware systems? ⁃ How does human–AI collaboration influence user's perceived shared control in AI-powered context-aware systems?	⁃ How does human–AI collaboration influence an organization's policies, priorities and culture? ⁃ How to design AI audition for human–AI collaboration in organizations? ⁃ How does human–AI collaboration influence employees' social roles in an organization?
Mutual benefit	⁃ How does human–AI collaboration influence users' affective relationship with AI in AI-powered context-aware systems? ⁃ How does human–AI collaboration influence users' trust calibration in AI-powered context-aware systems?	⁃ How does human–AI collaboration improve organizational performance?
Mutual growth	⁃ How does human–AI collaboration influence users' perceptions of mutual growth with AI-powered context-aware systems?	⁃ How does human–AI collaboration influence AI-powered systems' cognitive intelligence, emotional intelligence and social intelligence as time evolves in organizations?

Table 5

Negative impact of human–AI collaboration on individuals and organizations

Research topic	Possible future research questions
AI discrimination	How to reduce and correct AI discrimination from AI-powered context-aware systems?
Loss of control	How to reduce and correct perceptions of “loss of control” for AI-powered context-aware systems?
Lock-in	How to reduce and correct “dependence lock-in” for AI-powered context-aware systems?

Table A1

Summary of the extant literature on human–AI collaboration

Author(s)	Context of human–AI collaboration	Methods	Research streams		Insights in human–AI collaboration
			Human–AI collaboration cultivation	Human–AI collaboration outcomes
Loske and Klumpp (2021)	Human–AI collaboration in outbound logistics of grocery retailing	Fuzzy data envelopment analysis (DEA)		✓	Outbound logistics activities benefit more from human–AI collaboration than from human-only or AI-only approaches when it comes to the mentioned vehicle routing problems
Pynadath et al. (2018)	Human–machine teams in autonomous systems	Conceptual	✓		Transparency in communication could boost teammates' situation awareness, strengthen the trustworthy relationship and enhance the performance of the human–automation team
Lebovitz et al. (2022)	Human–AI augmentation for medical diagnosis	An ethnographic field study	✓		Authors distinguished between “engaged augmentation” and “unengaged augmentation” when it comes to human–AI collaboration In order to reconcile their own knowledge practices with AI's claims of knowledge, experts that use AI interrogation procedures that have reduced their overall uncertainty
Fügener et al. (2021)	Human–AI complementarity	Online experiment	✓		AI advice would influence “unique human knowledge” – the human knowledge that an AI lacks Capacity of humans to profit from AI advice is proportional to their relative competence as well as the ability to distinguish between accurate and wrong advice
Schoonderwoerd et al. (2022)	Colearning in human–AI teams	Experiment	✓		Human–AI colearning promote human–AI collaboration Specific sequences of interactions were devised for the purpose of initiating and facilitating colearning
Korteling et al. (2021)	Human–AI teaming	Conceptual	✓		This study explains the underlying traits, distinctions and idiosyncrasies that distinguish human and artificial intelligences When human team members comprehend the distinctions between human and artificial intelligence, it may encourage human–AI cooperation
Okamura and Yamada (2020)	Collaboration between human and autonomous AI agents	Online experiment with a web-based drone simulator	✓		The safety and effectiveness of human–AI cooperation is frequently dependent on how trust is calibrated between humans and AI agents The study proposes a trust calibration technique that includes a framework for identifying the improper calibration state by monitoring the user's behaviors and cognitive signals to advise the user to restart trust calibration
Chen and Krishnamurthy (2020)	Human–AI collaboration for mind mapping	Algorithm design	✓		Comparative analysis of human–AI cooperation and human–human collaboration in mind mapping
Baki Kocaballi et al. (2020)	AI documentation assistants	Co-design workshops with general practitioners	✓		AI documentation assistants for care consultations still need human supervision. To guarantee patient safety, doctor safety and quality of care, a variety of human–AI collaboration models will need to be developed and tested
van den Broek et al. (2021)	Human–AI hybrids for hiring	An ethnographic field study	✓		Human–AI hybrids are the consequence of a process of mutual learning in which intense interaction with the AI drove users to reflect on how they created knowledge
Silva de Oliveira et al. (2022)	Human–AI hybrid systems	Conceptual	✓		This study contributes to the development of cognitive systems by offering recommendations for the appropriate application of hybrid human–AI knowledge
Grover et al. (2020)	Human-in-the-loop planning	Prototype design	✓		This study proposes a decision support system for sequential planning issues that involves a human in the loop
Puranam (2021)	Human–AI collaborative decision-making	Conceptual	✓		This study proposes typologies for examining various sorts of human–algorithm division of labor and the learning configurations they are grouped in
Andrews et al. (2022)	Human–AI teams	Literature review	✓		Human–AI shared mental models are vital for team performance
Liu et al. (2022)	The human–AI scoring system	Experiment		✓	Radiologists may increase the accuracy of COVID-19 severity assessment by using the human–AI scoring system
Asan et al. (2020)	Human–AI collaboration in healthcare	Conceptual	✓		Optimal trust, which refers to humans and AI should both maintain a certain degree of skepticism about the judgments made by the other and are needed for human–AI collaboration in healthcare
Sturm et al. (2021)	Coordination of human learning and machine learning	Agent-based simulations	✓	✓	Human explorative learning can be reduced by using machine learning Human adjustments to machine learning systems are generally beneficial; however, this benefit can be harmful in some situations Knowledge generated by machine learning systems can boost organizational learning in turbulent environments
Metcalf et al. (2019)	Artificial swarm intelligence	Conceptual	✓		Instead of focusing on how artificial intelligence may replace humans, ASI encourages people to consider how it could augment human skills
Naiseh et al. (2021)	Human–AI collaborative decision-making	A multistage qualitative research method	✓		The study examines the design of interactive explanations to support trust calibration and provide design principles and methods for such explanations
Longoni and Cian (2022)	AI-based recommendations	Experiments		✓	The word-of-machine effect describes that AI recommenders are better at utilitarian tasks than humans, but worse at hedonistic ones. Human–AI collaboration would reduce such effects
Knop et al. (2022)	AI-enabled clinical decision support systems	Literature review	✓		The study offers a thorough overview of the crucial factors, such as training data quality, explainability and medical expertise which affect how medical professionals interact with and collaborate with AI-based clinical decision support systems
Siemon (2022)	AI-based systems for idea evaluation	Experiments		✓	In a collaboration scenario, using an AI-based system to evaluate ideas reduces anxiety about a future review, while human evaluation increases anxiety about an evaluation to come
Pandey et al. (2022)	Human-in-the-loop machine learning	Experiments	✓		The study develops a paradigm for human errors in stream processing and a method for mitigating human errors
Ramchurn et al. (2021)	Trustworthy human–AI partnerships	Conceptual	✓		The study discusses the necessities of trustworthy human–AI partnerships such as creating human–AI team simulations and explanation features for verifying and validating human–AI relationships
Jarrahi (2018)	Human–AI symbiosis in organizational decision-making	Conceptual	✓	✓	This article analyzes how humans and AI might work together in organizational decision-making processes that are often marked by ambiguity and complexity
Mirbabaie et al. (2022)	Human–AI collaboration at the workplace	Survey + interview		✓	This research demonstrates three indicators for AI identity threat derived from human–AI collaboration in the workplace: changes to job, loss of status position and AI identity
McNeese et al. (2021)	Human–autonomy teaming	Experiments	✓	✓	Trusting a teammate (autonomous agents or human) – is related with improved team performance, and that trust may grow over time
Jia et al. (2022)	Human–AI teaming	Prototype design	✓		A visual explainable active learning method is proposed as a means of enhancing human–AI cooperation
Burton et al. (2020)	AI-based systems	Literature review		✓	The study organizes the proposed causes and remedies for algorithm aversion into five categories. Human-in-the-loop decision-making could be one of potential solutions for algorithm aversion
Oh et al. (2018)	Human–AI cocreation	Prototype design + survey	✓	✓	The study designed a prototype for human and AI drawing pictures collaboratively. When AI provided more clear instructions, users' satisfaction would increase Users always want to take control of the work; yet, they wanted the AI to explain its objectives Users enjoy interacting with the system though they gave the AI a poor rating in terms of predictability, controllability and comprehensibility
Cai et al. (2019)	Human–AI collaborative medical decision-making	Interviews + a qualitative laboratory study	✓		In addition to case-specific justifications for AI models, physicians needed upfront knowledge of the model's fundamental, global features in order to collaborate successfully with AI
Nakahashi and Yamada (2021)	The human–agent team	Markov decision process	✓	✓	The study models a collaborative agent with implicit guidance to balance human autonomy and task performance
Cabitza et al. (2021)	Human–AI interaction in collaborative medical tasks	Simulation	✓		Effective interaction protocols may ensure superior decision-making team performance than individual agents
Maadi et al. (2021)	Human-in-the-loop machine learning for medical applications	Literature review	✓	✓	The study summarizes the role of humans in the machine learning process and how they interact with machine learning approaches
Steyvers et al. (2022)	Human–AI hybrid systems	A Bayesian combination model	✓		Differentiating mistakes made by human and machine across multiple class labels could improve hybrid human–machine performance
Sundar (2020)	Human–AI interaction in mediated communication	Conceptual	✓	✓	Various affordances are predicted to influence user engagement and experience based on how much they allow users to interact with the system, give them a sense of control, offer them tangible benefits, and provide and give them ways to improve each other
Shrestha et al. (2019)	Human and AI-based organizational decision-making	Conceptual	✓	✓	The study compares human and AI-based decision-making on different dimensions and offers a unique paradigm detailing how both types of decision-making might be merged to optimize the organizational decision-making
Zerilli et al. (2022)	Human–AI teams	Conceptual	✓		The study discusses the relevance of different forms of algorithmic transparency played in rebuilding trust when humans and AI work together
Newton et al. (2022)	Human–bot teams in the open-source software development	Experiments	✓	✓	Compare human and human–machine teams regarding their similarities and differences The effects of the usage of bots influence productivity of the open-source software development
Sowa et al. (2021)	Human–AI collaboration in the workplace	Interviews		✓	Human–AI cooperation was shown to boost productivity in this research
Peeters et al. (2021)	Hybrid collective intelligence	Conceptual	✓	✓	The impacts of AI on society are being debated from three angles: the techno-centric, human-centric and collective intelligence-centric views
Strobelt et al. (2022)	Human–AI collaboration for data-backed text generation	Prototype design	✓		This study designs a prototype called GenNI (Generation Negotiation Interface) to enhance human–AI collaboration in the context of text generation
Terziyan et al. (2018)	AI in Industry 4.0	Prototype design	✓	✓	The study proposes the Pi-Mind technology that can capture and duplicate decision models of humans. The Pi-Mind technology allows humans and artificial intelligence (AI) to share responsibility for the consequences of their decisions
Cañas (2022)	Human–AI collaboration	Conceptual	✓	✓	Cosupervision and coresponsibility for collaborators (i.e. human and AI) is required for human–AI collaboration
Wei et al. (2022)	Human–AI collaborative conversational systems	Quasi-experimental design	✓		The informational demands of users were taken into consideration while developing a hierarchical categorization method for human–AI collaborative conversations
Rajpurkar et al. (2022)	AI in health and medicine	Literature review	✓	✓	Human–AI collaboration is a promising future research direction for the medical AI domain
Veitch and Andreas Alsos (2022)	Human–AI interaction in autonomous ship systems	Literature review		✓	Human–machine collaboration could lead to better system performance than any counterpart could accomplish by operating alone
Yu and Li (2022)	Human–AI collaborative work	Online experiment	✓		Workers' trust in AI is affected by AI decision-making transparency in a human–AI collaborative work environment
Rai et al. (2019)	Human–AI hybrids in digital platforms	Conceptual	✓		Human–AI hybrids may vary from “substitution” (AI replaces people) to hybrids augmentation (humans and AI enhance each other) to assemblage (AI and humans are working as an integrated entity)
Raisch and Krakowski (2021)	AI in organizations	Literature review	✓	✓	Unlike automation, which refers to AI taking over human, augmentation refers to people collaborating with machines to complete a task. Automation and augmentation are not just mutually exclusive but also inextricably linked
Askarisichani et al. (2022)	The human–AI nexus in group decision-making	Literature review	✓		The study discusses significant factors in managing human–AI decision–making in a group environment

Notes

1. URL: https://www.mckinsey.com/featured-insights/artificial-intelligence/notes-from-the-ai-frontier-modeling-the-impact-of-ai-on-the-world-economy

2. URL: https://omdia.tech.informa.com/OM018678/Artificial-Intelligence-Software-Market-Forecasts--2Q21-Analysis

3. https://www.reuters.com/article/us-amazon-com-jobs-automation-insight/amazon-scraps-secret-ai-recruiting-tool-that-showed-bias-against-women-idUSKCN1MK08G