Servants and masters: An activity theory investigation of human-AI roles in the performance of work

Organizations considering AI adoption must be mindful of media that portrays dysto-pian future scenarios. While machine sentience remains philosophically and ethically moot, the future implications of AI adoption are unclear. The issues that surround AI adoption need to be examined but there are a lack of implementations cases around which empirical research can be undertaken and practical experience can be gained. AI adoption needs to be considered from multiple viewpoints including, but not necessarily limited to the social, ethical and legal issues, and not merely be reduced to questions of financial return or organizational efficiency.

Alle n, Ro b e r t A., W hi t e, G a r e t h R.T, Cl e m e n t , Cl ai r e E., Alex a n d er, P a ul a n d S a m u el, Ant h o ny ORCID: h t t p s ://o r ci d.o r g/ 0 0 0 0-0 0 0 2-4 4 8 3-4 6 0 0 2 0 2 2. S e rv a n t s a n d m a s t e r s : a n a c tivity t h e o ry inv e s ti g a tio n of h u m a n-AI r ol e s in t h e p e rfo r m a n c e of w o r k. S t r a t e gi c C h a n g e 3 1 (6) , p p. 5 8 1-5 9 0. 1 0. 1 0 0 2/js c. Thi s v e r sio n is b ei n g m a d e a v ail a bl e in a c c o r d a n c e wit h p u blis h e r p olici e s. S e e h t t p://o r c a . cf. a c. u k/ p olici e s. h t ml fo r u s a g e p olici e s. Co py ri g h t a n d m o r al ri g h t s fo r p u blic a tio n s m a d e a v ail a bl e in ORCA a r e r e t ai n e d by t h e c o py ri g h t h ol d e r s .

| INTRODUCTION
The path of human development can be marked by the emergence of tools and technologies that have enabled further technological progress and disrupted existing social norms (Kittel, 1967;Lewandowsky, 2016;White, 2017). The digital revolution is perhaps the most significant sociotechnical change that has occurred in the last two centuries, and its impact has not yet been concluded (Agarwal, 2015;Feldman, 2002). Among the many technologies this phase of rapid technological development has fostered, Artificial Intelligence (AI) promises (or threatens) to deliver change unlike any seen before. Whereas new technologies have previously afforded the means of replacing skilled labour with machines, or enabled seamless communications to take place over global distances, AI is anticipated to revolutionize every aspect of work and society and even potentially remove human physical involvement from both (Tegmark, 2017).
This research was undertaken in an organization that has begun to trial and implement AI solutions and is considering how it should formulate its future AI strategy. Amidst a wealth of subjective media content but a dearth of empirical literature, the organization commissioned the study that aims to understand the employees' perceptions of how AI systems have affected, and are expected to continue to affect, the performance of work.
The paper is organized thus: first, an overview of AI systems a n dc h a l l e n g e si sp r e s e n t e dt h a tconceptualizes humans and AI agents as elements of systems of work. The Activity Theory literature is then reviewed before the study's methodology is detailed.
Next, the thematic analysis of the interview data is presented and a discussion of the key findings is made. The paper closes with statements of contribution, limitations and suggestions for future research.

| ARTIFICIAL INTELLIGENCES
Artificial Intelligence (AI) has undergone a lengthy gestation period, arguably much longer than any other technological development.
From the earliest notions of "programs with common sense" in 1959 (Sabanovic et al., 2012;McCarthy, 1989: p. 99), it is only in the last few years that human-like intelligent systems have become a technical possibility, if not a practical reality. In 2017, the US introduced the 'Future of Artificial Intelligence Act' (Cantwell, 2019), and in 2018 the UK's House of Lords issued a report on AI (Lords, 2019). Both of these identified the lack of a universally accepted definition of AI and adopted the concepts of 'Narrow' and 'General' AI. Narrow AI (AIn) pertains to those systems that can demonstrate human-like or higher, levels of cognition within a limited set of functions. General AI (AIg) refers to systems that can replicate human-like, or higher levels of cognition across all domains and are thereby indistinguishable from humans (Moor, 2003).
Due to the lack of universally accepted definitions of AI there is considerable debate over whether such systems already exist: adopting one definition over another leads to inconsistencies in classification (Kaplan and Haenlein, 2019). Jarrahi (2018) and Gartner (2019) of AIg are much more difficult to discern and even "may never exist" (Gartner, 2019, p. 14). AIg scenarios, at present, tend to be hypothetical a n da r eo f t e nl i n k e dw i t h'doomsday prophecies' (Atkinson, 2016;Buntinx, 2019;Collins, 2018;LaSane,2018).
The development and use of AI systems affects how work is done and requires new skills to develop further implementations (KPMG, 2019;SIA, 2019). Whether the adoption of AI systems will ultimately result in the permanent loss of jobs or a change in the types of jobs that are required in the future remains a moot point (Choudhury, 2019;Rees, 2019). Whatever the outcome, such an emerging technology's legal and ethical implications will influence and shape its impact on the workplace (Stahl et al., 2017).

| ACTIVITY THEORY
Activity Theory has been widely used as a framework for academic research (White et al., 2019). It is "essentially a learning theory" (Jarzabkowski, 2003, p. 27) that enables the understanding of processes such as expansive learning that occur over relatively long periods (Engestrom, 1987(Engestrom, , 2000aEngestrom et al., 2005). Bedny et al. (2001: 414) concur, stating "action is the basic unit of learning activity." Thus, the processes of mental cognition and work behavior can be interpreted as the processes of human learning, since "through activity a per-son…[obtains] knowledge" (Bedny & Karwowski, 2004:151). Blackler (1995) and Benson and Whitworth (2007) maintain that in the study of work activities, it is the identification of stresses or contradictions or 'disturbances' that are sought. The continual forging, relaxing and reforging of relationships between actors and artifacts, termed 'knotworking' by Engestrom et al. (1999), become the focus of attention during Activity Theory enquiry. Kain and Wardle (2005: 122) note the value of Activity Theory in identifying these conflicts and contradictions in work-based systems that "interfere with the realisation of individuals' and communities' goals." Ardichvili (2003) details the component elements of Activity Theory shown in Figure 1. Tools are those implements that Subjects use to perform an Activity in the pursuit of the Object. An Object may be the "focus of study of some discipline (e.g., general accounting rules in financial accounting)" (Ardichvili, 2003:9 ) .B e d n ye ta l .( 2004) define a further type of Object that is artificial, created by individuals to regulate their actions and termed Artifacts. The process of performing an Activity, or 'doing work', is influenced by several organizational factors. Rules are those conditions in the workplace that govern how work is performed.
These may comprise governing regulations, standards or procedures.

| RESEARCH CONTEXT
Established as the Anglo-Persian Oil Company in 1909 (Ferrier, 1982), BP is a global energy business with operations in Europe, North and South America, Australasia, Asia and Africa. The company has 74,000 employees and operates in 70 countries worldwide. It is one of the world's "supermajors' (Davis, 2006), recording a turnover in 2017 of $244.582B (USD) (and has around 20,000 barrels of oil equivalents in its reserves).
BP has committed to digital transformation, and AI is regarded as a key near-term source of value creation both inside and outside the company's boundaries, as demonstrated by the investment of $5m (USD) in Belmont Technology for their cloud-based geoscience platform (Ali, 2019;B P ,2019a). Increasing global demand for energy, particularly in China, India and throughout Africa is driving the requirement for significant investment (BP, 2019b). Across the industry, there has been an increasing push to utilize emerging technology (Jacobs, 2019). Areas, where the deployment of AI is envisaged to be able to deliver benefits, include data analytics, customer service solutions, and complex documentation assessment at both upstream and downstream points to optimize production, manufacture and sales (IEA, 2017). Figure 2 presents the Activity Theory framework populated with AI systems as Tools. That is, they are devices that are utilized to aid in the performance of work. The Subjects of the work system are the 'Employees' that perform work to achieve some organizational or individual Object or outcome. The Rules of the work system, the Community and the Division of Labour are elements that influence work performance. In accord with Thompson (2004) and Engestrom's (2000b) assertions, it is the individuals that perform the work that is the focus of the investigation of systems of work, and the Employees are, therefore, the sources of data that are utilized in this study.

| Theoretical framework
Through exploring individual insight into the utilization of AI, and the Activity System surrounding their implementation, this study discovers the 'knotworking' that occurs in the performance of AI-facilitated work. These tensions are examined and provide indications about the challenges that the further adoption of AI present and thereby aid in guiding the organization's strategic planning.

| METHODOLOGY
Semi-structured interviews were used in this study to gain deep insight into participants' positions (Denscombe, 2010;Fox, 2009).
The questions were operationalized from the literature and broadly conformed to the structure of Activity Theory: that is, questions were based upon Tools, Subjects, Objects, Work, Rules, Community and the Division of Labour (Charmaz, 2006). Openended questions were used to elicit broad responses and guide the development of further probing questions around emergent and interesting subjects (Halcomb & Davidson, 2006;Lynch, 2000;Schwartz & Schwartz, 1955;Strauss & Corbin, 1998). The cyclic development and refinement of interview questions is a crucial approach to improving the reliability of interpretive research (Becker, 1958;Bositis, 1988;M i l e s ,1979;S a n d a y ,1979; Schwartz & Schwartz, 1955).
Interviews were conducted with 11 key stakeholders of the host organization. Each participant had at least 2 years of experience within the field, and their gender profile reflects that of the company in general (see Table 1 for details). The interviews were captured using a digital voice recorder and accompanied by field notes of pertinent issues and comments (Paolisso & Hames, 2010). The lead researcher transcribed the interviews, and two researchers independently analyzed the data using thematic analysis (Guest et al., 2012). The analyses were cross-compared with reach a consensus and then member-validated (Sandelowski, 1993).
The interviewees were invited to participate in the study with the option to withdraw at any time. The participants have been anon- that is considering future AI adoption is the respondents' perceptions that it would affect the majority of current roles but that their own role would not be impacted.

| DISCUSSION
This section considers the challenges that AI presents to the adopting organization that were identified in the analysis. In keeping with an Activity Theory approach, the focus of the discussion is those tensions that affect the nature of the performance of work. Termed 'knotworking' by Engestrom et al. (1999) and 'disturbances' by Blackler (1995), these tensions affect the achievement of individual and collective goals.

| Concepts of AI
As discussed in the literature review, AIn applications are relatively easy to visualize. This capability may be due to there being more instances where AIn has already been implemented, and people can therefore draw upon concrete examples. It may also be due to the notion that AIn is thought to largely comprise methods of automating existing forms of work. The empirical evidence certainly supports the idea that many individuals perceive AIn as an approach that could undertake the mundane, repetitive elements of their own jobs.
Contrastingly, AIg applications are considerably more difficult to appreciate. In accord with the literature review, many individuals pictured future or advanced forms of AI as precursors to some dystopian scenario (Sections 6.1 and 6.6). This viewpoint may be due to their lack of exposure to AIg and may also be due to bias through being exposed to media-generated doomsday prophecies.
This study suggests an alternative interpretation of the impact of AIg, that is, individuals consider AIn applications as presenting less of a threat to their own job security than AIg. For instance, many of the participants volunteered examples of AI as an 'augmenting technology' that would improve their own performance of work, mainly through undertaking the routine operations (Section 6.1, 6.2 and 6.3).
However, when they were prompted to discuss how future AI could be utilized in the organization, they all identified that it could entirely replace existing jobs (Section 6.7). This perspective could simply be a 'fear of the unknown' or a fear that AIg is a threat to their job security, however, since all participants identified that future AI would be a threat to everyone else's jobs but not their own, we interpret it as an indication of the sense that AIg is perceived as a threat to their job security.
Whether our interpretation that AIg presents a threat to job secu- inhibit the desire to explore further opportunities for AI adoption.
Consequently, the organization risks lagging behind the industry and its competitors. The organization may therefore be caught in a 'catch 22' situation whereby the lack of AIg case examples precludes the ready adoption of more advanced AI, which, in turn, further embeds the reasons to be fearful of future AI. AI developers and potential adopters may need to justify the expenditure on AI systems not merely by fiscal measures but also in knowledge acquisition and through improving the company's readiness for change.

| Intellectual property rights
The issue of IPR ownership featured highly in all of the discussions with the respondents (Sections 6.5 and 6.6). While the emphasis placed upon IPR may be due in part to the nature of the adopting organization, which relies heavily upon the development of new insight and technologies, it is not an issue that is unique to this company or sector. Consequently, IPR within AI scenarios is likely to be an issue that affects a broad range of organizations and industries.
The empirical evidence indicates that the IPR that AIn systems may generate would be perceived to belong to the organization. The arguments for this viewpoint ranged from comparison to human workers whose IPR belonged to the employing organization and the notion that anything that arose from the analysis of the company's data would similarly belong to that organization.
However, the perspective of ownership changed when consider- The discussion around the ownership of IPR became more complicated when the respondents considered AIg systems that were also 'sentient' or 'human-like' (Section 6.6). At this juncture, many raised the (potential) problem of motivating an AI system-to 'hand over' any IPR or intelligence it may be capable of generating. Activity Theory considers the subject of the work system, the human worker, to be the agent that possesses motivation to carry out work.
However, AIg potentially introduces the need for the dimension of motivation to be considered in the Tool within the work system, which is discussed further in the following section.

| Disruption of activity systems
The potential for both AIn and AIg to radically alter the ways of working is indisputable. Even if AIg systems fail to become a reality, AIn affords the means of transforming the nature of work through the automation of routinised activities and analyzing data sets that are beyond human capabilities. However, it would be remiss not to consider how future AI systems that approach or become AIg may impact the ways of working. To this end, Activity Theory affords an apposite lens for this study, and its use indicates that AIg adoption has significant implications for the performance of work in the future.
The previous section raised the notion of sentient AI systems posing a unique set of problems for their adopting organizations regarding IPR ownership. In particular, it suggested that AIg systems require the dimension of motivation to be recognized within Tools in the Activity Theory framework. The literature review and empirical evidence also point to further development that Activity Theory may require. For instance, the respondents discussed how their current roles could be augmented by using AI systems, primarily through the automation of repetitive tasks (Sections 6.1 and 6.2). They also identified the likelihood of future AI systems replacing many current jobs, although interestingly, not their own (Section 6.7). Additionally, they recognized the potential for future AI systems to undertake missioncritical work such as strategic decision-making and directing the activities of human workers. The literature review identified instances where human workers are already employed as data sources for AI systems that make 'intelligent' decisions based on mental models derived from human inputs.
Collectively, we posit that such developments would require the structure of the systems of work, as depicted by Activity Theory, to be reformulated. Figure 3 presents our reconfiguration of Activity FIGURE 3 AI as subjects in the performance of work is likely to become one of society's 'wicked problems'. Only relatively recently have higher order primates been granted rights that recognize them as intelligent 'non-human persons' (Barnes, 2015;Mazie, 2015;Sentience Politics, 2018;Sommer, 2017). Primates have also been argued to possess a sense of property (Brosnan, 2011).
One can question at what point similar concepts may be applied to sentient AI systems.
This question has, in part, already been addressed. For instance, the AI robot 'Sophia' has been granted citizenship in Saudi Arabia (Galeon, 2017), and citizenship tests for future systems are being developed (Independent, 2019). In the U.S., corporations have been granted certain rights, and there are arguments that AI systems should be granted similar (Yampolskiy, 2018). However, there are also arguments against such a move (Conversation, 2017), but these seem to be based upon 'trust' and lack of understanding rather than being the- envisage. Dialogues around the future potential of advanced AI systems tend to reduce to generic anti-utopian scenarios, but discussion of their application in work-based environs appears enlightening.
There is a unanimous and fearful appreciation that advanced AI systems will replace the majority of roles, however, most interestingly, nobody considers that their own function is at stake. Second, for this organization, the question of who owns the intellectual property generated by an AI system is important. Generally, this problem is seen as a commercial question to be addressed in any contracts for the procurement of AI systems. However, it becomes highly problematic when considering that such future AI systems may be sentient. Third, current changes in legislation have granted AI systems citizenship, and this trend may develop to encompass other inalienable rights. Collectively, these pose significant problems for society at large as well as for organizations that adopt or develop the technologies. The emergence of more advanced forms of AI, along with the current utilization of human workers as data sources for intelligent systems, suggests that future systems of work may substitute human and nonhuman workers within the framework of Activity Theory.
Through adopting systems of work as the lens through which the impact of AI has been examined this study proffers a contribution to Activity Theory. Traditional Activity Theory identified human workers as the subject of the systems of work, that is, they possess the motivation to perform work. The extant literature and the empirical evidence suggest that intelligent AI systems may become the Subject of systems of work in which human workers become the Tools. While this scenario is redolent of a dystopian future, we posit that it is not necessarily the case. Contrastingly, the broader application of AI systems may enable human workers to focus on the creative elements of organizational work or afford them the time and opportunity to pursue more enriching social activities.
The focus of this study has been the perceptions of key stakeholders within a single company. While their perceptions ought to be indicative of the issues that face many 'non-tech' companies it must be recognized as a limitation of the generalizability of the findings.