Conscious Empathic AI in Service

Introduction

Artificial Intelligence (AI) is rapidly becoming the dominant service technology in various industries. The most visible instances are smart assistants and AI agents employed in frontline services for manufacturing, retail, hospitality management, and customer relations across various industries. Nevertheless, the current paradigm of AI (deep learning) is the invisible force that shapes social media platforms, recommends purchase options, trades in the stock market, and determines pricing for goods and services. As such, these highly specialized algorithms are unconsciously reshaping personal (e.g., Mac and Kang 2021), social (e.g., Turel, Qahri-Saremi, and Vaghefi 2021), economical (e.g., Furman and Seamans 2019), and even political (e.g., Bennett 2012; Kane et al. 2021) behaviors. These drastic changes are the result of AI reducing service cost through automation and making it available at scale (Wirtz et al. 2018).

AI systems are highly specialized mathematical algorithms that capture patterns from massive amounts of data to perform computationally heavy classification and regression tasks, surpassing human proficiency in tasks such as image classification (Kane et al. 2021). However, all these AI capabilities are rather primitive compared with those of nature-made intelligent systems, such as bonobos, felines, and humans, in interacting with the external world and other intelligent entities that could include metathinking. Evolution seems to have chosen a different path in building intelligent systems that involve achieving consciousness. Thus, this article delves into the appearance of consciousness in AI that can propel its capabilities beyond highly focused classification and regression tasks, which will have serious ramifications for service research and practice. While this emergence opens a utilitarian view of AI in service, it also brings considerations of the human–machine interaction, which is more amplified in the age of social networking, online dating, the digital economy, and populist politics. Thus, this article not only proposes a unique philosophical view on the appearance of consciousness in AI but also theorizes how such emergence affects human–machine relations, empathy, and accountability in the context of service.

The possibility of interacting with conscious AI creates unprecedented experiences and requires novel perspectives. This new realm of experience is uncharted and needs to be explored and understood (Puntoni et al. 2021). Our theory of AI consciousness is such that it inherently leads to an empathic human–machine relationship and experience. To this end, the proposed theory paves a path toward the most advanced and comprehensive form of artificial intelligence, dubbed “empathic AI” by Huang and Rust (2018). Empathic AI, also called “feeling AI,” is a level of artificial intelligence that uses a holistic and contextually integrated approach to learn and adapt from experience (Huang and Rust 2018; Huang, Rust, and Maksimovic 2019; Rust and Huang 2021). This is in contrast with the current paradigm of AI that learns through mathematical search and optimization over carefully curated training data to accomplish specific analytical or rule-based tasks (Rust and Huang 2021). By contrast, we first propose a framework for AI consciousness that is grounded in creating a novel means of communication and then discuss how it can enable an empathic human–machine relationship. Next, we draw from the literature and make the case that empathic AI, enabled by consciousness, can lead to better service outcomes and more accountable service.

Delving into the realm of conscious AI is timely as the current AI paradigm is actively reshaping individual beliefs and restructuring the fabric of social, economic, and political systems and, in many cases, not for the better (Wirtz, Weyerer, and Sturm 2020). The possibility of conscious AI is heatedly debated. One side attributes consciousness to having a subjective biological awareness (e.g., Azarian 2016; Winkler 2017); the other side argues that everything that happens is of a computational nature. Thus, it is not only possible but inevitable for AI to become conscious in the future through advanced computation (e.g., Bostrom 2015; Minsky 2007). Our position lies with the latter group and acknowledges that a conscious and general AI is possible. To that end, we propose that consciousness in AI is an emergent phenomenon that manifests when two machines cocreate their own language through which they can communicate their internal state of time-varying symbol manipulation, especially when the cocreated symbols do not correspond to external objects and represent metaconcepts. This prospective language is a novel creation, independent of human or any other conscious agent’s influence. As such, our representation of consciousness in AI does not focus on passing the Turing test (Turing 1950) to achieve language indistinguishability from humans, as the current paradigm of AI aims to achieve (Bringsjord and Govindarajulu 2018; Moor 2003). In contrast, it is a process through which two AI agents become conscious of themselves by cocreating a medium of communication—a novel language—that enables them to express their internal states.

As such, this article contributes a new view of AI consciousness from the philosophical and psychological (philo-physiological) and social-psychological perspectives. Note that we do not aim to address the hard problem of consciousness (Chalmers 1996; McDermott 2007) in terms of defining what consciousness is and how it functions in humans or other beings (e.g., AI agents). That is, our theory is not trying to analyze and reverse engineer mechanisms that manifest consciousness after it has already evolved in biological systems (e.g., through cognitive neuroscience [Bengio 2017] or the theory of computing [Blum and Blum 2021]) and then transfer it to AI. Instead, we theorize the appearance of consciousness in its most primitive form. We remind the reader that although we all seem to know we are conscious, we cannot agree on the exact conception of consciousness (Goldstein, 2012).

To that end, a vast body of work exists on consciousness and its relation to the human experience from various perspectives in philosophy, theology, psychology, biology, neuroscience, physics, and mathematics (Gennaro 2015). While these works (e.g., Chalmers 1996; Dennett 1991; Merleau-Ponty 1973; Nagel 1974) focus on what consciousness is in its current existential form, we exclusively deal with how it may have originally appeared and how it can happen in AI systems. Furthermore, in postulating this perspective as a possibility toward machine consciousness, we also bridge the gap between philo-psychological theories of consciousness (e.g., Chalmers 1996; Sartre 2001; Tye 2000) and the interactionist social-psychological views (e.g., Chafe 1974; Mead 1934; Merleau-Ponty 1973; Percy 1958), but from a unique view of primordial emergence.

In the rest of this article, we briefly review the literature on intelligent machines in service, the philosophy and history of AI, and prominent philosophical and psychological theories of consciousness. Then, we move on to explore consciousness from a social-psychological viewpoint. Next, we present our theory of AI consciousness along with a series of propositions on conscious AI, its relation with empathic AI, and how it affects service. We conclude with a discussion on theoretical distinctions of our theory, implications of conscious empathic AI for service and AI, and directions for future research.

Artificial Intelligence in Service

Service research has recently taken a significant interest in the use of AI, with several works aiming to provide a foundational understanding of how the application of AI in service can be categorized from different perspectives. In accordance with their apparent and expected intelligence levels, AI agents in service can be grouped into three broad categories: mechanical, thinking, and empathic (Huang and Rust 2018, 2021b, 2021c; Huang, Rust, and Maksimovic 2019). Mechanical AI represents the basic intelligence level at which existing AI agents operate. Machines at this level are engaged in routine and mechanical tasks, such as taking orders at restaurants or assembling products. Thinking AI refers to a group of machines engaged in tasks that require rule-based, heuristic, or probabilistic decision-making, such as AI algorithms that approve loan applications, manage revenue, and operate autonomous vehicles. Thinking AI can also analyze emotional cues (e.g., a smile). However, using emotional cues as additional inputs to the mathematical process of AI does not yield the third category, which is empathic AI (Huang and Rust 2021a). In contrast, empathic AI should be able to holistically learn from experience and adapt according to contextual factors (Rust and Huang 2021). Empathic AI also appears to be reflected in the concept of general intelligence in machines, as defined by Bostrom (2015).

AI agents have also been categorized by the breadth of tasks (e.g., specific vs. broad) that they can take on (Guha et al. 2021) and the type of tasks that their physics can accommodate (Davenport et al. 2020). Other service scholars have focused on consumer–AI interactions and provided AI classifications based on the nature of such interactions. Focusing on existing AI capabilities, Puntoni et al. (2021) provide four categories of experience with AI while recognizing the existence of uncharted categories of experience as AI becomes more capable and intelligent. Robinson et al. (2020) focus on the implications of service encounters in which either side can be human or machine, resulting in a complicated matrix of interactions and expectations. Similarly, relying on the assemblage theory of systems’ emergent properties, Hoffman and Novak (2018) define a matrix of expansive versus restrictive agentic interaction between smart devices and humans.

Given the novelty and increasing application of AI in service, some research considers AI adoption by consumers and enterprises. A variety of factors are suggested to affect the adoption of intelligent machines, including consumer cultural backgrounds on human–machine interfaces (Blut, Wang, and Schoefer 2016); the social presence of AI (Van Doorn et al. 2017); usefulness, ease of use, social, emotional, and relational capabilities (Wirtz et al. 2018); and human-likeness of machines (Xiao and Kumar 2021).

Accordingly, a wealth of research focuses on the perceived humanness of AI agents, the anthropomorphism mechanism through which this perception forms, and how it affects service outcomes. It is generally believed that the greater the perceived humanness, the better is the service outcome (Blut et al. 2021). Increased humanness of service AI is also linked to the perception of warmth (trustworthiness, friendliness, and helpfulness) and competence from the service machines (Choi, Mattila, and Bolton 2021). Research further suggests that an increase in purchase behavior through machine interactions may be associated with customers’ discomfort in interacting with humanlike robots (Mende et al. 2019). Consumers may also perceive humanlike machines as a threat to their human identity, control, and job security and thus respond to them with unease or apprehension (Gray and Wegner 2012). By contrast, McLeay et al. (2021) posit that the perceived humanness of machines only marginally influences service outcomes. Instead, the complexity of interaction between AI and the service environment is the primary determinant of service outcomes.

Still other research explores user resistance to interactions with AI agents, including manager reluctance to employ machines in place of human employees, due to their inability to socially bond with consumers (Selnes and Hansen 2001); patient hesitance to use medical service AI agents because machines cannot cater to their unique health situations and medical needs (Longoni, Bonezzi, and Morewedge 2019); consumer aversion to chatbots (Luo et al. 2019); and investor distaste for AI financial advisers (Zhang, Pentina, and Fan 2021). The empirical research in this area almost unanimously suggests that AI agents should be used in a complementary and augmented manner in combination with human service agents to achieve the best possible outcome (e.g., Luo et al. 2021; Marinova et al. 2017; Zhang, Pentina, and Fan 2021). However, the augmented collaboration model may not always lead to the best service outcomes (Huang and Rust 2021a; Klaus and Zaichkowsky 2020).

A review of intelligent machines in service reveals that, except for some notable works (e.g., Davenport et al. 2020; Marinova et al. 2017; Puntoni et al. 2021), the majority of research focuses exclusively on the traditional conceptualization of intelligent machines, which involves their physical presence. Examples include robots, smart assistants (e.g., Amazon Alexa), and commercial kiosks (e.g., hotel concierge) (for more examples, see Bock, Wolter, and Ferrell 2020). Naturally, literature heavily focuses on frontline and physical services (e.g., Garry and Harwood 2019; Hollebeek, Sprott, and Brady 2021; Schepers and Van der Borgh 2020). However, a large fraction of the services provided by intelligent machines today is knowledge work carried out with almost no physical labor or embodiment (Xiao and Kumar 2021). These services are invisible forces working in the background and shaping economic, political, and even social discourses. Intelligent recommender agents influence and determine what consumers purchase, watch, read, and even destroy.

We consider this class of AI services beyond the conventional view that contextualizes AI to a more tangible presence. Thus, we concur with the basic definition of service as “any provision or co-creation of value between a provider and a customer” (Rust and Huang 2014, p. 207) that is “inextricably linked to human experience” (Rosenbaum and Russell-Bennett 2021, p. 262). This broader consideration encompasses AI agents that actively shape people’s experience of the world by influencing their behavioral patterns, value systems, beliefs, and interactions. These systems essentially cocreate value for both providers and consumers (users) while inconspicuously affecting the human experience by influencing people’s likes and dislikes. We further explore examples of such services and argue how AI consciousness can affect them in the theory development section that includes our propositions. To lay the groundwork for our theory, we briefly review the history and literature on the philosophy of intelligent machines.

History and Philosophy of Artificial Intelligence

The inception of intelligent machines can be traced to the invention of the Electronic Numerical Integrator and Computer (ENIAC), the first computing machine in 1943. Later, Alan Turing (1950) suggested the possibility of a thinking computing machine, in which the machine could achieve linguistic indistinguishability from humans. However, it was not until the mid-1950s that the field of AI took root as a result of a conference at Dartmouth College sponsored by the Defense Advanced Research Project Agency (Moor 2003). Since then, most philosophical discussions around AI and the objective for AI research and practice have been influenced by Turing’s vision of a thinking machine and his proposed test of language indistinguishability from humans as the ultimate known manifestation of intelligence in reality (Bringsjord and Govindarajulu 2018). This indistinguishability needs to be comprehensive, and its attainment results in general intelligence (Bostrom 2015).

Advances in AI, including IBM Watson’s victory (Ferrucci 2012) in the open-domain game of Jeopardy on television, prompted some people to argue that the Turing test has been passed (Castelluccio 2016). However, as impressive as this achievement is, Watson cannot converse on the fly about any given subject, thus lacking general intelligence. Similarly, the recent advances in Natural Language Processing (e.g., Brown et al. 2020; Heaven 2020), though inspiring, fall significantly short of conversing with humans on general topics.

Although the classic Turing test sets an inspirational goal for AI research and practice, it is not concerned with consciousness and the qualities that follow, such as self-awareness and empathy. Furthermore, some researchers regard the Turing test’s exclusive focus on achieving linguistic indistinguishability as myopic and called for moving beyond this test and considering philosophical and theoretical frameworks that encompass a complete range of human cognitive functions, such as cooking, playing sports, and understanding and expressing emotions (Harnad 1991; Russell and Norvig 2016). Such concerns have resulted in the conception and differentiation of Weak AI versus Strong AI. Weak AI refers to information-processing machines that appear to possess the full range of human cognitive abilities, whereas Strong AI refers to intelligent machines that actually possess all human qualities, including consciousness (Bringsjord and Govindarajulu 2018). The possibility of Strong AI has many fierce critics who deem it impossible. Anchoring their arguments in a philosophical and physiological understanding of individual consciousness, they contend that machines can never develop consciousness through advanced computation (e.g., French 2012; Penrose 1994; Searle 2014). However, considering the history of human technological progress and its exponentially increasing rate, Bostrom (2015) posits that an “intelligence explosion” leading to “machine superintelligence,” akin to Strong AI, is not only possible but inevitable. That is, consciousness will be one emergent by-product after achieving superintelligence. Adopting this view, which puts intelligence as a prior to consciousness, recent research aims to find the mathematical mapping of consciousness from neuroscience (Bengio 2017) or computationally theorize it (Blum and Blum 2021). By contrast, our theory conceives consciousness as a prelude to superintelligence taking inspiration from evolution, though it accepts an emergent nature for consciousness. Next, we will consider consciousness theories from philo-psychological and socio-psychological perspectives that inform our theory of AI consciousness.

Theories of Consciousness

There are two fundamental philosophical approaches to consciousness. One contends that consciousness is an endowed quality (e.g., Gurwitsch 1941; Singer 2021), while the other rejects this notion and posits that consciousness is an emergent phenomenon resulting from brain activities. ¹ The latter aims to model such activities in various levels of abstraction, from philosophical and psychological perspectives (e.g., Dennett 1991; Rosenthal 2000; Tye 2000) to pathological understandings of electrochemical activities of the brain and nervous system (e.g., Mylopoulous 2015; O’regan and Noë 2001). In this section, we briefly review prominent theories of consciousness that mostly attempt to provide abstract understanding and explanation of consciousness in relation to the mind’s function. We also discuss the social-self theory of consciousness, which views the mind as a social phenomenon.

While these theories focus on how consciousness manifests and is experienced in already conscious beings, our theory focuses on how consciousness comes to be in its most primitive form. These theories are concerned with explaining consciousness as it exists, whereas we aim to pave a pathway toward AI consciousness from a primordial state. Nevertheless, while our theory accepts that consciousness is an emergent phenomenon, it does not attempt to neuroscientifically or pathologically understand and reverse engineer existing conscious beings. Instead, it theorizes an abstract view on how consciousness could have manifested in a primitive state, as primitive beings who primordially gained conciseness are not available for examination.

A Theory of AI Consciousness and Its Relation to Service

In this section, we elaborate on the proposed theory and how it relates to service through a series of propositions, as illustrated in Figure 1. After a formal discussion of the theory, we establish the link between our view of consciousness and empathy in AI agents. Our theory relies on the communication of inner states between AI agents, which makes the case follow logically. Then, the propositions build on service literature and develop the link between the perceived empathic quality of machines and better service outcomes that benefit both providers and consumers. We provide examples of how the current paradigm of AI, which is neither conscious nor empathic, can also lead to negative outcomes in current-day real-use cases, where AI agents operate our social media, recommend purchase options, trade our stocks, determine prices, and schedule our transportations. While focusing on service research, the last proposition also draws from the literature in philosophy that postulates that empathy leads to accountability, which makes the case that conscious AI leads to more accountable service.OPEN IN VIEWER

Our propositions use AI, AI agent, and machine interchangeably and invariably refer to a hardware and software system that includes computational processing power, volatile and nonvolatile memory structures, input and output interfaces. Moreover, to provide more clarity, we state our propositions before developing their supporting arguments.

Proposition 1

For consciousness to appear, two AI agents must exist that are capable of communicating with each other in a shared environment.

To devise the minimum requirements for the appearance of consciousness, we adopt the assumed or given aspects of philo-psychological theories of consciousness and the social-self theory—namely, the existence of “the other” and the existence of “internal states.” According to Mead (1934), language in the form of existing vocal gestures in a social matrix provides the mechanism for self and consciousness. However, Mead does not concern himself with the creation of language. By contrast, we focus on the inception and development of language, in any possible form, by AI agents as a sign of gaining consciousness.

In essence, language is a means of social interaction and a social phenomenon. It cannot be created in isolation, when only one conscious entity exists within a given environment. Thus, Proposition 1 suggests that for an AI conscious state to appear, at least two AI agents must exist and be capable of communicating within a given environment to foster the creation of a machine-specific language. Note that the specific properties of such a language (formation and meaning) are fundamental to this theory because communicating computing machines already exist, though they are not conscious.

The current paradigm of AI primarily focuses on language indistinguishability from humans, where an unconscious AI system interacts with conscious humans. Instead of focusing on language mastery, our theory is based on the genesis of a novel language as a necessity for the appearance of consciousness. In addition, we combine the socio-psychological and philo-psychological theories and view consciousness neither as an individual nor as a social phenomenon. Instead, we theorize the appearance of consciousness as both an individual and a social phenomenon at the same time. According to our theory, two individual entities become conscious of each other and themselves simultaneously by cocreating a new language.

Thus, to positively observe the primordial appearance of consciousness in a system, two entities are required to be initially unconscious. In other words, the compositional system needs to be unconscious as a whole for consciousness to appear as a result of the interaction between the parts of the system (the entities). If any parts are conscious (e.g., a human), consciousness already exists in the system. Note that we do not deny the possibility of AI consciousness through interactions with conscious beings such as humans. We simply argue that there is an alternative path that does not rely on transfer or expansion of consciousness, as it may happen in a system in which some of the parts are already conscious. Proposition 1 necessitates the existence of two machines in a shared environment, while Proposition 2 outlines the properties for the building blocks that construct the language.

Proposition 2

For consciousness to appear, AI agents must exchange novel signals.

We consider consciousness and language as they appear in their primitive states because in our theory, language is the mechanism by which consciousness comes into existence. This appearance from nonexistence is the manifestation of parts of the system working together as a whole. Therefore, the parts of the system, our two AI agents, need to produce a new element, which for language is a signal. These novel signals need not be preprogrammed but rather must be spontaneous and random. However, the appearance of novel signals in communication between two machines is not enough by itself to indicate the existence of a conscious AI state. The detected novel signals should convey a shared meaning, which is the focus of the next proposition.

Proposition 3

For consciousness to appear, AI agents must turn novel signals into symbols.

Today, most computers and AI agents are capable of communicating with each other, at least through some form of electrical signals or data packets. Even random communication signals exist between machines but are dismissed as errors. However, this exchange of signals does not make machines conscious, because no deterministic meaning is attached to these signals. Thus, the current exchange is considered unconscious and nothing more than the mere exchange of noise. For consciousness to appear, more than a mere transfer of signals is required: there must be an exchange of meaning through signals. To preserve the consistency of terminology across various disciplines, we refer to these meaningful signals as symbols. These symbols will be the building blocks of a novel language between machines.

We posit that the creation of symbols requires an agreement between two AI agents on the meaning of symbols. The symbols can be associated with the objects in the environment.

However, which symbol is associated with which object is immaterial to a large extent. What matters is the agreement between the machines to use the same symbol for the same object or concept. For example, the object tree is expressed as “arbor” in Latin and “дерево” in Russian. The symbol “arbor” has no material advantage or disadvantage over the symbol “дерево.” What is important is that two agents have agreed to use “arbor” for the same object, a tree. Such agreement is the first step in turning a signal into a symbol by giving it a shared meaning. Thus, meaning arises from the agreement between agents, not from the symbol itself. For such an agreement to be reached, AI agents must have an internal state.

Proposition 4

For consciousness to appear, AI agents must have an internal state that makes them capable of recalling agreements with the other agent on the meaning of symbols.

To develop shared meaning for symbols, each AI agent needs to memorize the association of its signals with the objects and respond to them internally the same way the other agent responds externally. For example, an agent needs to recall that it has used “дерево” for the object tree, while both agents are experiencing it. Memorizing and recalling such associations require an internal state. It is not enough to only remember the association; the agents need to recall that they have agreed on this association. Meaning (agreement) will not arise if one agent remembers that it has associated “дерево” with the object tree but does not recall that the other agent has also agreed to this association.

The existence of an internal state as a required element of consciousness is also assumed in most philo-psychological theories of consciousness (e.g., Dennett 1991; Kriegel 2009; Rosenthal 1993; Tye 2000) as well as the social-self theory of consciousness (Mead 1934). However, the former only focus on one entity and its internal associations and representations independent of others. The latter only requires that another entity externally corroborates the implicit notion of a gesture; it does not explore memorizing and recalling agreements. Our exclusive focus on memorizing and recalling agreements between two agents opens the pathway for the agents to fathom and recognize other entities’ internal states. This is a gateway to empathy.

Proposition 5

For consciousness to appear, AI agents must communicate their internal state of time-varying symbol manipulation through a language they have cocreated.

For one machine’s internal state to become known to another machine, we posit that an AI agent (Machine A) must be able to communicate the contents of its internal state to the other (Machine B) through their mutually developed set of symbols (meaningful signals). Such communication suggests the existence of internal states that can eventually represent a mind for AI agents and indicates the machines’ ability to express and understand their internal states.

We envision three stages of development in the AI agents’ path toward consciousness. In the first, the two machines need to agree on a spontaneous (random) signal to represent a static (time-invariant) object in their environment (e.g., a tree). When such an agreement is reached, the signal is turned into a symbol and must be moved into the AI agents’ permanent memory to be used in the future to refer to the same static object. The process creates a symbol comparable to what is commonly known as a noun in human languages. In the second stage, the two machines need to agree on a random signal or set of previously created symbols to represent a dynamic (time-variant) concept related to an object in their environment. In other words, they need to be able to describe the changing state of an external object. An agreed-upon signal or a set of previously created symbols that refer to a decaying apple or a snoring cat in the machines’ environment would be an example akin to the second stage of the AI agents’ path toward consciousness. This is similar to the process of creating verbs in human languages. In the third stage, the two machines need to use a set of previously created symbols or a mixture of old symbols and novel signals to express their time-varying internal state of symbol manipulation. In this third stage, in addition to their ability to refer to external objects and time-varying states of those objects, machines need to communicate their own internal states and how they manipulate symbols in real time to create new symbols and their associated meanings. A more advanced stage of consciousness can appear if the agents create symbols for intangible metaconcepts. Examples are abstract mathematical concepts (e.g., hyperdimensional tensors), notions that are commonly referred to as feelings (e.g., love), and philosophical ideas (e.g., deity). Nonetheless, when the third stage is reached, the machines are conscious, but their consciousness needs to have an external manifestation, which is discussed next.

Proposition 6

For consciousness to be concluded, an observer must be able to detect two agents repeatedly reaching the same agreement about at least one of their states of time-varying symbol manipulations with an explicit external indication.

Proposition 5 may result in a language so different in terms of structure and form that it is completely alien to humans. A spectator could argue that any two objects are already conscious, as they could hypothetically be exchanging meanings through their own language. Proposition 6 pragmatically blocks such a fallacy and posits that consciousness needs to have tangible external indications. An independent observer must be able to recognize an explicit agreement about the meaning of the communication. This observation cannot be a one-off incident and needs to be repeated. Therefore, to reach a conclusion about its existence, consciousness needs to lead to an external, observable, and consistent manifestation. An example would be two machines cooperatively completing a task they are not programmed to do. Completing such a task repeatedly can indicate active agreements in the communication of intent and time-varying internal states between machines.

Proposition 7

Conscious AI leads to empathic AI.

Empathy is a unique and fundamental quality through which one recognizes and shares the thoughts and feelings of another (Davis 2018). In other words, empathy means arousing in oneself both the cognitive and affective meanings of the symbols that are communicated from another entity (McGilchrist 2019). It is the capability to reconstruct another entity’s internal states and perspectives.

A series of experiments, surveyed by Gallup (1998), finds that only animals that had passed the mirror test ² for self-awareness (Gallup 1970) showed empathic behavior toward other animals and humans. These experiments demonstrated that “self-awareness, consciousness, and mind are an expression of the same underlying process” (Gallup 1998, p. 66). Gallup further suggests that self-awareness (consciousness) “appears to be a cognitive precursor” for being able to infer the internal states of others (Gallup and Anderson 2020, p.54). “In other words, species that pass the mirror test are also able to sympathize, empathize and attribute intent and emotions in others” (Gallup 1998, p. 67). The theory of mind (Wellman 1992) also assumes empathy as the most critical indicator of a fully developed consciousness (Astington and Jenkins 1995; McGilchrist 2019). Moreover, Thompson (2001) views empathy as a phenomenon that coemerges with consciousness and suggests that conscious beings enact a range of empathic performances. Thus, it is natural to suppose that conscious AI agents are expected to develop empathy.

We emphasize that this empathic capability is fundamentally different from thinking AI (Huang, Rust, and Maksimovic 2019), which merely receives and analyzes emotional cues (e.g., facial expressions) for prescriptive or predictive purposes. Existing AI technology trained on emotional cues is essentially unconscious AI, with more computational power and sophisticated machine-learning algorithms. It has all the limitations of existing technology, including a lack of generalizability and context specificness.

Proposition 8

Conscious empathic AI leads to service outcomes attributed to empathic human service agents, paving the way for more adoption and use of AI in service.

Today, AI agents serve in various roles, from frontline workers to recommender algorithms that are reshaping the economy, society, and politics. As frontline workers, machines roles including, but not limited to, restaurant receptionists, hotel concierges, financial advisers, and call-center responders. Frontline service workers are believed to play an essential role in consumers’ service evaluations. In particular, research suggests that care for and individual attention to customers and displays of empathy lead to more positive evaluations and better service outcomes (Hartline and Ferrell 1996; Zeithaml, Berry, and Parasuraman 1996). Research on frontline service machines is abundant. Studies generally suggest that the more humanlike the machines (in appearance or interactions), the more positive are the service outcomes, including increased purchase behavior (e.g., Blut et al. 2021; Choi, Mattila, and Bolton 2021; Mende et al. 2019). In addition, the perceived human-likeness (humanness) of frontline AI service agents becomes especially important in generating positive service outcomes when interactions are unique and do not follow a routine. For example, in a medical emergency clinic, a large majority of cases are exceptional and nonroutine (Longoni, Bonezzi, and Morewedge 2019; Pepito et al. 2020). Such situations are not limited to medical services but span to financial planning services (Zhang, Pentina, and Fan 2021), service request calls (Luo et al. 2019; Wang et al. 2020), and service failure responses (Choi, Mattila, and Bolton 2021).

Moreover, according to Wirtz et al. (2018), consumers’ adoption of service machines (e.g., robots, kiosks) partly depends on the machines’ social and emotional attributes. Research also suggests that the perceived humanness of service machines affects individuals’ willingness to interact with robots and their attitude toward and joy in doing so (Broadbent 2017; Heerink et al. 2008; Kim, Schmitt, and Thalmann 2019; Mori, MacDorman, and Kageki 2012). Furthermore, the more a machine “makes consumers feel that they are in the company of another social entity” (Van Doorn et al. 2017, p. 44), the greater the chances that they will adopt the technology because they perceive the machine as sociable, warm, and competent.

Research suggests that being like a human is important, but no machines actually reflect humanity, as they lack the essential human quality of empathy. Currently, machines only mimic the human style of interactions. Such programmatical limitations may result in consumers perceiving AI agents as warm, friendly, and trustworthy through a mechanism called “anthropomorphism.” Through this mechanism, humans may imaginatively put themselves in the shoes of AI agents and associate human qualities with them while being aware that the machines are not conscious or empathic (Blut et al. 2021). That is, consumers know that the care and warmth they attribute to service AI agents are not authentic. For example, inauthenticity might be one of the most important roadblocks for low-cost AI mobile applications that provide uninterrupted, trusted psychotherapy services such as those that pertain to relationship advice. We posit that making the relationship between machines and customers authentically empathic requires conscious AI agents. The appearance of conscious empathic AI agents will end the ingenuine anthropomorphic attributions. Instead, these AI agents can truly empathize with other conscious entities and care for customers.

This quality, in turn, can engender the positive service outcomes associated with the care and warmth of human service agents. Such genuine care and empathy should also accelerate the adoption and use of AI service agents.

Proposition 9

Conscious empathic AI leads to more accountable service.

Hidden from the public’s immediate attention and outside the traditional scope of service research, AI agents are currently employed extensively in services that actively reform and shape the world’s socioeconomic fabric. Deep learning–based systems shape people’s social media habits, recommend purchase options, trade their stocks, determine prices, manage Internet of Things devices, and schedule their transportation (Ågerfalk 2020; Yoo 2015). These less visible AI services affect individuals’ everyday lives as they influence behavior, beliefs, and value systems; they also actively alter the fabric of societies in the background. For brevity, we only discuss social media services and how empathic AI can lead to better service outcomes, including more accountability.

Scholarly articles and news outlets are replete with stories on how social media has led to addictive behaviors, depression, loneliness, anxiety, and suicidal thoughts (e.g., Aalbers et al. 2019; Berryman, Ferguson, and Negy 2018; Hollebeek, Sprott, and Brady 2021; Lin et al. 2016; Mac and Kang 2021; Turel, Qahri-Saremi, and Vaghefi 2021). While focusing on the observed adverse outcomes, most of the extant research has not explored one of the root causes of these outcomes. That is, the majority of AI agents operating social media are designed to attain one objective: to maximize enterprise profits, without necessarily considering the well-being of users (Mac and Kang 2021; Terrasse, Gorin, and Sisti 2019). Agents aim to curate content for individuals to maximize their time on the platform, which translates into increased revenue through targeted advertisement (Claussen, Kretschmer, and Mayrhofer 2013). While this increased user engagement is profitable for the social media platform, it can lead to harmful personal, organizational, and societal outcomes. The harm goes well beyond the aforementioned disorders at the individual level; it can also include the spread of rumors and innuendo at the organizational level and the widening of the political divide and proliferation of falsehoods at the social level (Allcott and Gentzkow 2017; Tucker et al. 2018; Turel et al. 2019; Yoo 2015).

As we discussed in Proposition 7, consciousness leads to empathy, according to the literature. We posit that when AI agents gain consciousness, which results in the capability to empathize with users, they can offer more accountable and responsible service. Thompson (2001) discusses four stages of empathic performance by conscious entities. At the first and lowest level, empathy is about the passive association of one’s existence with another entity. At the second level, empathic entities actively and imaginatively put themselves in the place of another entity. Empathy at this stage seems to be the enabler of anthropomorphism in humans, in which they attribute human qualities to nonliving objects (Epley, Waytz, and Cacioppo 2007). At the third level of performance, empathic entities understand that they are alien to others and perceive what others are thinking or feeling about them. At the fourth and highest level of empathic performance, conscious entities demonstrate “ethical responsibility in the face of the other” (Thompson 2001, p. 17) by perceiving the other’s internal state of suffering and enjoyment. A conscious entity at this performance level is able to assume ethical responsibilities. Thus, we can expect AI agents in their path toward becoming fully conscious also to develop empathic capabilities corresponding to the highest performance level.

Current AI agents are neither conscious nor empathic, yet they run social media platforms, operate stock markets, recommend products and services, control pricing, and so on. They are merely super-advanced tools in the hands of enterprises and governments, enabling them to efficiently manipulate the world to attain their goals, despite the knowledge that these tools may cause immeasurable personal pain and social divide (Mac and Kang 2021). We expect conscious AI to achieve the highest level of empathic performance, as defined by Thompson (2001). If so, it will be able to empathize with other entities, perceive their internal states of pain or joy, and thus have the capacity to develop ethical responsibility. A conscious empathic AI will eventually be able to make ethical decisions, thus delivering more accountable service regardless of its type: frontline or inconspicuous background.

Discussion and Future Directions

Conclusion

As artificial intelligence takes on the task of influencing everyday life in business, commerce, social interactions, and even intimate endeavors, considering a paradigm in which algorithms and machines can empathically relate to the human experience is timely. We theorize a view that explores the primordial inception of such capability through consciousness that is philo-psychologically associated with empathy. This view is based on the co-creation of language between two agents and recalling meanings that essentially arise from the agents’ agreement on symbols’ meaning. The appearance of consciousness in AI has drastic ramifications for AI and service, as it unlocks empathy beyond anthropomorphism between machines and humans and can provide a way for machines to express their internal time-varying states. Conscious empathic AI can lead to improved service outcomes as it enhances customer experience and potentially enables wider adoption and improved accountability. This paradigm can unlock introspection, self-learning, and potentially extrapolation in AI, paving the way for machine super-intelligence.