Abstract
In this short paper, we reflect on ‘algorithmic management learning’ – a phenomenon that dates back to the early twentieth century but has gained fresh impetus in the dawning age of artificial intelligence. In particular, we suggest that management learning has today become a human-machine hybrid. This form of management learning is not only increasingly non-reflexive, it is also impeding the human capacity to be reflexive and to learn reflexively. We conclude by outlining the consequences of algorithmic management learning for the future of this journal.
Introduction
How do managers learn about the task of managing organizations? This is a question that goes to the heart of Management Learning. On the surface, the answer is self-evident. Managers learn by taking courses in business administration in a university; they learn by developing their skills on professional development programmes; and they learn by on-the-job experience in their daily work. Perhaps they even learn by reading this journal.
On another level, the answer has become more complicated in recent years. Managers now also learn by interacting with machines, such as AI-powered knowledge assistants and data-driven people analytics. As we will argue, these algorithmically-mediated forms of learning do not leave the fundamentals of human-centred management learning untouched. This paper sketches out what we might call ‘algorithmic management learning’ – a phenomenon that dates back to the early twentieth century but has gained fresh impetus in the dawning age of artificial intelligence.
We suggest that the (re-)emergence of algorithmic management learning poses new, important questions for the future direction of this journal – most fundamentally, (1) how is machine learning different from human learning and (2) how is the capacity for humans to learn impacted by the increasing prevalence of machine learning? We begin by locating these questions in the history of this journal and the broader field of management learning before examining the prospects for (algorithmic) management learning.
The development of management learning
Tracing the origins of management learning is anything but a straightforward task, but it is hard to imagine that Frederick Taylor’s (1911) Principles of Scientific Management would be missing from such a history. For Taylor, the practice of management is based on knowledge that is first gained from shopfloor workers and then systematized by efficiency engineers (Braverman, 1974). In this way, the old ‘rule of thumb’ mode of production – controlled by workers without input from supervisors or factory owners – is replaced by a ‘scientific’ mode of production. Work now becomes an engineering puzzle, one that can be solved only by a trained cadre of managers. In this context, ‘management learning’ involves both (a) learning the ‘one best way’ to undertake a given task in a company through observation and measurement and (b) disseminating the principles of scientific management to other managers in other industries, as Taylor did in his day – and as management textbooks for business students do in our own time.
Scientific management marks the birth of management learning, in the sense that Taylor was one of the first to establish a set of standardized guidelines for managing staff in any organizational context. On this basis, managers can be taught how to manage effectively – regardless of their area of expertise. But scientific management also marks the birth of algorithmic management learning, in the sense that Taylor sought to extract knowledge from workers and then reduce this knowledge to a series of detailed step-by-step instructions that are directed towards a predetermined goal. Scientific management, in other words, follows a strictly computational logic in its pursuit of efficiency – a logic that, as we will see, has only accelerated in recent years with the rise of machine learning in data-driven organizations.
In the decades following Taylor, management would continue to develop as a field of knowledge. Indeed, it now claimed the title of ‘science’. During the Cold War, management science took its cue from related disciplines such as operations research, logistics, cybernetics, information systems, and economic game theory. What these areas shared was a focus on statistical modelling and mathematical problem-solving in pursuit of optimal decision-making in organizational contexts. The idea of management as a science, based on rigorous analytical methods, played a central role in allowing managers to pursue a project of professionalization throughout the twentieth century (Khurana, 2007). Key to this professionalization project was the assumption that management did not only involve on-the-job experience; it also involved skills that could only be gained in institutions of higher education, such as basic competencies in economics, statistics, and behavioural sciences.
Academic journals in the field of management studies bear the imprint of this professionalization project. It is no surprise that, in the years following the Second World War, a number of influential academic journals were founded, not least Management Science (est. 1954), Administrative Science Quarterly (est. 1956), and Academy of Management Journal (est. 1958). Instrumental in establishing ‘science-based professionalism’ (Khurana, 2007: 291) in the field of management, such journals were meant to provide a forum for research about management but also – initially at least – for managers.
How does Management Learning fit into this history of management learning? The journal was founded in 1970 under its original title Management Education and Development. Early issues of the journal located it squarely within the project of management professionalization that had been underway in the United States for the past two decades. As Derek Pugh (1970: 1) wrote in the inaugural issue, the aim of the journal is to help ‘management teachers to contribute to the policies and practices of the profession [of management]’. In particular, Management Education and Development focused on three categories of management learning: management education in higher education, management training on professional development programmes, and practical experience in organizations (Wexley and Baldwin, 1986). In all three categories, the individual manager took centre-stage as the learning subject.
In its initial years of publication, Management Education and Development maintained a close connection to management science. Tellingly, the first-ever issue of the journal saw it publish several articles on operations research and economics (e.g. Cook, 1970; Fisher, 1970). Other issues from this era of the journal include contributions from computer science and other fields in which the machinic elements of management came to the fore (e.g. Powell, 1978; Wheelwright, 1972), although the individual manager would retain their status as the ultimate subject of learning – that is, the human at the heart of the organization who uses science to inform their executive decisions.
This view of the manager, as a trained professional tasked with running organizations, came under sustained attack during the 1970s and 1980s. During this time, the neoliberal vision of investor capitalism recast management as an impediment to shareholder value rather than its main driver. Managers (and the business schools that trained and socialized them) could no longer be trusted to deliver dividends for the owners of company stock. This is why, according to neoliberal economists like Michael Jensen and William Meckling, managers needed to be brought into line by large shareholders such as pension funds, insurance companies, and commercial banks (Useem, 1993). The manager now started to be seen as a tool of other institutional actors rather than as a professional decision-maker in their own right. Management learning, if it now meant anything at all, was to be directed towards increasing shareholder value on behalf of investors.
Needless to say, Management Education and Development did not embrace this shift in the managerial role. Most of the articles published in the first two decades of the journal stayed true to the ideal of management as a profession based on education (in university settings) and development (on training courses). Yet the journal did undergo significant transformation in the 1990s. During this decade, a cleft opened up between ‘mainstream’ management studies and ‘critical’ management studies, a division that was partly methodological and partly political. Such a polarization had been brewing since the publication of Burrell and Morgan’s landmark study Sociological Paradigms and Organizational Analysis in 1979, but reached its clearest articulation in Alvesson and Willmott’s edited collection Critical Management Studies in 1992. The rift between functionalist and critical approaches to management set the scene for Management Education and Development to relaunch itself as Management Learning in 1994. In the words of the editors at the time, Management Learning now took an explicitly ‘critical view of the processes of management and organization’ (James and Snell, 1994: 7). It was a transformation that let the journal distance itself, once and for all, from the remnants of management science and fully embrace the human-centred tradition.
As Management Learning emerged in the form we know it today, the meaning of the word ‘learning’ also started to change. The types of ‘learning’ that came to dominate the field became less mechanistic and more humanistic. Management science disappeared from view, squeezed out by more qualitative and interpretivist forms of inquiry. In particular, the humanities – literature, history, philosophy – came to occupy a prominent place in the journal (Johnsen et al., 2021; McAulay and Sims, 2009). As a result, Management Learning became a broad platform for discussions about power, ethics, neo-colonialism, and various forms of exclusion – issues that are central in critical management education (Contu, 2009; Grey, 2004; Zulfiqar and Prasad, 2021). But it is perhaps the concept of ‘reflexive learning’ that best exemplifies the journal’s shift away from management science – a key topic in the journal over the last few years (e.g. Beech et al., 2021; Parker et al., 2020).
In their contribution to the 50th anniversary issue of Management Learning, Durepos et al. (2020) argue that exercising reflexivity is an ontological act. As they put it, it is through ‘continuous and self-conscious reflection on past decisions and actions’ that one gains the insights needed to make an active choice in ‘the composing of oneself’ (Durepos et al., 2020: 8). Reflexive learning, in other words, involves a deliberate relating to one’s past, a form of relating that could lead to a break with that past. This view aligns with Hannah Arendt’s (1958) understanding of ‘action’, a type of human activity that involves bringing something radically new into the world. Human beings, she says, are born with the capacity for new beginnings and it is the exercise of this capacity that differentiates us from other living beings. And from machines. This is why reflexive learning – the ability to make something new out of oneself by reflecting on our own past – is a concept that helps to illuminate the guiding mission of Management Learning today: namely, to ‘open up existing ways of thinking to scrutiny’ in order to ‘promote new perspectives and interpretations’ (https://journals.sagepub.com/description/MLQ).
As critical management scholars ourselves, we fully endorse the goals of Management Learning in its current form. But we also feel that the journal’s 55th anniversary provides an occasion to reflect on Management Learning’s potential blind spots. In this spirit, we suggest that the journal – or, more specifically, the field it represents – has embraced humanistic forms of management learning at the risk of overlooking more machinic forms of management learning. Since the 1990s, Management Learning has rightly distanced itself from the engineering mentality that dominated the field when it was first published – a mentality that continues to dominate many US-centric management journals. But in so doing, Management Learning has missed the opportunity – with some exceptions (e.g. Barros et al., 2023) – to reflect on other trends in management learning, particularly the rise of machine learning and AI-assisted tools in organizations.
In what follows, we reflect on machine learning and human learning in relation to the concept of reflexivity. In so doing, we hope to point to new directions of research under the umbrella of ‘algorithmic management learning’.
Machine learning and human learning
As we mentioned earlier, scientific management is profoundly algorithmic. However, the machine learning algorithms that are increasingly prevalent in data-driven companies today are qualitatively different from the basic optimization algorithms that date back to Frederick Taylor’s time. Under Taylorism, the process of learning comes to a halt when the ‘one best way’ (as Taylor famously called it) is revealed – this is the moment at which the ‘science’ part of scientific management ends and the actual work begins. For machine learning algorithms, by contrast, there is no ‘one best way’; the algorithm continues to develop and improve, in a recursive loop, on the basis of data it receives. For example, a machine learning algorithm – trained on historical data about employee turnover – predicts, on an ongoing basis, which members of staff are likely to leave the organization. And its future predictions will take into account which employees actually left the organization. The point is that machine learning algorithms learn from themselves, that they actively recode themselves in light of mistakes or misjudgments they made in the past.
Machine learning algorithms have a relation to the past that is captured in the etymological origin of ‘learning’. The term stems from the Old English leornian, with the meaning of ‘to follow or find the track’. Algorithmic learning is rooted in the tracing of the past, especially the historical ‘traces’ that we all leave – as employees, consumers, citizens – in online records of our behaviour. Such traces accumulate in data pools that can be used to predict and modify individual and collective future behaviour (Zuboff, 2019). The power of algorithmic learning lies in its ability to forecast and optimize decisions by making inferences from the data that it is trained on. In other words, the past directs the future. This is notably different from reflexive learning. As we noted above, reflexivity also involves a relation to the past, but it is an active relation (or a re-relating) – a form of learning that does not merely ‘find’ the optimal path but rather opens up space for changing paths.
We often talk about the ‘decisions’ that algorithms make, such as a recruitment algorithm that ‘decides’ who to hire or a targeted advertising algorithm that ‘decides’ what product to recommend. However, we may reasonably ask whether algorithms are truly capable of making decisions at all.
In his 1961 book Decision Order and Time in Human Affairs, economist GLS Shackle argues that a ‘decision’ that follows rationally from past knowledge is not a real decision. A true decision, he says, produces something radically new or unexpected. As Shackle (1961) puts it, a true decision is ‘more than mere response to circumstances and contains an element which we may call inspiration, which brings essential novelty into the historical sequence of states of affairs’ (p. ix). Such inspired decision-making finds its possibility in the human freedom to imagine alternative futures. From this perspective, algorithms, while they might be said to make choices, cannot make decisions in Shackle’s sense of the word: they are unable to stop, reflect, and start again. The Netflix recommendation algorithm, for example, cannot have second thoughts about recommending a certain show or movie; it simply does or does not. So, while we might want to claim that algorithms can ‘think’, they certainly cannot be said to ‘think twice’ or have ‘second thoughts’ about a matter. Each algorithm must necessarily follow its own path even when this path changes as a result of its learning loop.
But algorithmic decision-making (or, rather, algorithmic choice-making) does not lead to a world in which machinic predictability replaces ‘the new’. On the contrary, algorithmic interactions produce novel modes of human thought, action, and relations – just think about how Uber drivers strive to maintain an average customer rating of over 4.6 in order to remain visible in the system (Rosenblat, 2018). Such behaviour would be unthinkable without the algorithmic nudges that prompt it. Yet algorithmic learning introduces these new ways of relating (to oneself and to the world) without ever breaking with the past. There is only continuity in algorithmic processes, even when they go awry. Consider racist chatbots or sexist recruitment algorithms. Even in such cases of apparent deviation, the results necessarily follow from how the algorithm is programmed to compute the inputs it receives. The so-called ‘madness of algorithms’, in other words, is madness that follows from the algorithm’s own logic (Amoore, 2020).
The difference between human learning (of the reflexive kind) and machine learning (of the algorithmic kind) can also be characterized in terms of a special kind of speed or, more accurately, a type of slowness. Algorithms are not merely faster in their capacity for processing data much more rapidly than humans. Their speed is in part due to the fact that they do not get side-tracked. This capacity for singlemindedness is a critical source of efficiency, but it is also, in another sense, an incapacity. The ability to slow down, to drift, to wonder, and to wander is a necessary condition for reflexivity, and it is a condition that machines necessarily lack.
The increasing prominence of machinic thinking holds out the possibility of this path-dependent singlemindedness affecting the way humans think as well. As our lives, and the paths we are on, become more intertwined with algorithmic learning, so we ourselves are in danger of losing the capacity to reflect. In other words, algorithmic learning does not merely assist us in diagnosing problems, making better decisions, and improving our efficiency; it also impacts on our ability to learn. As philosopher Mark C. Taylor (2013: 276) puts it, ‘there is a recurrent feedback loop in which human beings produce technologies that turn back on them and recreate their creators in their (that is, the technologies’) own image’. In other words, as humans interact with machines, so we become more machinic in our thought and behaviour.
As we have said, algorithms are incapable of reflexivity, of ‘thinking twice’. But they also make reflexivity more difficult for those who are targeted (or assisted, or directed) by algorithms. Not only is Netflix incapable of pondering what movie to recommend, it also seeks to take away the need for the user to ponder. This is why sociologist Eran Fisher (2022) has recently argued that algorithms produce a new form of subjectivity. Because algorithms free us from the burden of choice, whether frivolous choices about which movie to watch or more weighty matters that involve ethical judgement, they create ‘a form of knowledge about the self, which ultimately excludes the self from the process of learning and knowing about the self’ (Fisher, 2022: 12). From this perspective, algorithmic learning interrupts the process of reflection – on one’s self and on one’s place in the world in relation to others – that is at the heart of human learning.
To return to Durepos et al. (2020), we agree with their suggestion that reflexivity involves a revaluation of the past, which requires a pause in the running of processes. However, we take issue with their claim that ‘reflexivity is intrinsic to management learning per se’ (Durepos et al., 2020: 8). What we see today is that management learning has become a human-machine hybrid – a kind of ‘algorithmic management learning’. Such a form of management learning is not only increasingly non-reflexive, it is also impeding the human capacity to be reflexive and to learn reflexively.
The future of (algorithmic) management learning
In the bestseller The Master Algorithm, computer scientist Pedro Domingos (2015) offers a utopian account of a future ‘society of models’ in which we are all doubled by our ‘digital half’, a kind of virtual doppelgänger that interacts with other digital halves. As a result, we no longer need to engage ourselves with cumbersome activities like applying for a job, buying a car or even finding a partner. The upshot, of course, is that we will be able to lead more fulfilling lives: ‘Tomorrow’s cyberspace will be a vast parallel world that selects only the most promising things to try out in the real one’ (Domingos, 2015: 270).
This vision is appealing but also deceptive. Contemporary developments in algorithmic learning do not point to a world in which our digital ‘halves’ interact with each other to improve our real lives. When we shop on Amazon, for example, we do not interact with the digital half of Jeff Bezos. Instead, we deal with a much larger system of artificial intelligence in which our ‘interests’ only play a role as input in an ongoing process of optimization. We are not even ‘half’ a person in these interactions, but far, far less – a mere data point swimming in a pool with billions of other data points. The question is, in a world increasingly governed by data-driven analytics, what happens to our subjectivity – and our reflexivity? This question is relevant to ask of all forms of interaction between humans and algorithms. But it is particularly relevant to ask of the field of management learning, a field that foregrounds the figure of the human manager as a reflexive learner.
Since the mid-1990s, Management Learning has come to understand ‘learning’ as a quintessentially human affair, defined by capacities such as reflexivity or ethical judgement that machines do not possess. But such a narrow understanding of learning does not do justice to recent technological advances in organizational contexts in which human learning becomes just one part of the puzzle, rather than the puzzle itself. Given the increasing dominance of machinic forms of learning, we suggest that it is important for the field of management learning to de-centre the figure of the human manager as the sole learning subject. ‘Management learning’ should no longer be treated as an essentially human affair in a time in which algorithmically-driven management systems are playing an increasingly prominent role in organizations – not only in logistics and operations but also in recruitment and HR. Management Learning is not solely responsible for deciding the kinds of submission it receives from scholars. But we would certainly encourage the journal to widen its scope to include more machinic forms of management learning while maintaining its critical, reflexive ethos. The key question for critical scholars to address is this: how can organizational actors maintain their reflexivity when so much of what (ought to) count as management learning today seeks to circumvent or tame that very reflexivity? It is a question that requires us to think alongside, against, and beyond algorithmic forms of management learning.
