Abstract
In this article, we evaluate the politics of recommendation engines by focusing on an indispensible feature of their operation: training. We use the notion of training as a key word which helps us link three bodies of knowledge: data science, the history of automation, and aesthetic and political theory. Training is a staple in the operation of algorithmic systems, and artificial intelligence more generally; it is a practical methodology by which these systems become intelligent. Training is also a key feature of how workers throughout history came to perform their labor, and how, during the 20th century, machines came to acquire this human ability, that is, automation. And lastly, drawing on Immanuel Kant’s theory of aesthetic judgment, Hannah Arendt offers a political theory where training is key to political judgment. We trace the meaning and significance of ‘training’ in these three fields in order to draw conclusions from one field to another.
[T]hough understanding is capable of being instructed, and of being equipped with rules, judgment is a peculiar talent which can be practiced only, and cannot be taught (Kant, Critique of Pure Reason, 177–78)
Introduction
If ‘artifacts have politics’, as Langdon Winner (1986) postulated, what can we say about the politics of algorithms? How can we say anything about an artifact so opaque? Understanding the digitally mediated world we inhabit is informed by our assumptions concerning what these media are: how they operate and what they do. The ontology of the medium and its representations allow us to think critically about their politics. But that is not quite that simple in the case of recommendation engines – algorithmically driven media that lack representation. As users of digital platforms, we are faced with the output of recommendation engines. We might assume what their input is, or hypothesize about it. Yet, as regular users we are ignorant of the algorithms underlying recommendation engines (Bucher, 2018; Panagia, 2020). Algorithms are not a medium of representation; they do not represent objects and they do not represent an appearance or a reality; their modality is non-representational (Borgmann, 2000). Following Taina Bucher’s suggestion to resist our tendency to see algorithms as a black box (Bucher, 2018), we attempt a kind of hypothetical reverse engineering. In this article, we evaluate the politics of recommendation engines by focusing on a key feature of their operation without which they cannot operate: training.
We use the notion of training as a key word which helps us link three bodies of knowledge: data science, the history of automation, and aesthetic and political theory. Training is a staple in the operation of algorithmic systems, and artificial intelligence more generally; it is a practical methodology by which these systems become intelligent. Training is also a key feature of how workers throughout history came to perform their labor, and how, during the 20th century, machines came to acquire this human ability, that is, automation. And lastly, drawing on Immanuel Kant’s theory of aesthetic judgment, Hannah Arendt offers a political theory where training is key to political judgment. We trace the meaning and significance of ‘training’ in these three fields in order to draw conclusions from one field to another.
Training is not merely a concept of professional, historical, and theoretical importance, but also part of our everyday language (e.g., training school, on-the-job training, potty training, strength training.). Training is a multilayered, complex practice that entails acquiring practical and tacit knowledge, knowledge that comes from the practical experience of doing a certain action in certain circumstances. Training pertains to a kind of knowledge which needs to be experienced, thus, distinguishing it from learning. One needs to hold a kite and feel the wind to understand how to fly a kite; one needs to ride a bike to be able to balance on two wheels and move forward. The aim of training is to perform a task – which we have already learned, or know how to perform – better than the day before, and it entails doing an action repeatedly. Hence, one can learn tactics and strategies for how to improve running but one cannot train from watching films or reading books about running. One needs to run in order to train for running. The Oxford Dictionary defines training as ‘the action of teaching a person or animal a particular skill or type of behavior’ (Oxford Languages, 2023). We want to suggest that training for judgment – a key human practice – now involves also machines. Although machines cannot make judgments, they are increasingly performing an action that masquerades as judgment and in doing so, they are de-training humans in their ability of making judgments, which as we argue with Arendt, requires training. This process has a far-reaching political implication which we wish to lay bare.
We proceed in four steps. We start in the following section “Training in algorithms” with the field of data science and present how training has become a centerpiece of algorithmic machines, focusing on our case-study: recommendation engines in Amazon and Netflix. Training is a concept that increasingly appears in the context of algorithmic culture in general and algorithmic recommendation engines in particular, for the task of personalizing recommendations and making them relevant to specific users, in order to keep users engaged with the platforms. We lay out the postulation that the more algorithms train, and the better they become at ‘making aesthetic judgments the more users rely on them, a process which in fact undermines their own training capacity for making judgments’.
In “Technology and training: a historical view,” we take an historical approach mobilizing critical studies of production technology – a second body of knowledge where training is key – to show this dialectical move of training between humans and machines. Next, we move to the third body of knowledge which engages training as an important practice for exercising judgment. In this section, “Training for aesthetic judgment,” we first present Kant’s understanding of aesthetic judgment as a unique human faculty, different from pure reason and practical reason, and as a central pillar for creating a community. Lastly, in “Linking aesthetic and political judgment,” we follow Arendt in tying together aesthetic judgment and political judgment, and highlighting the significance of training – as an individual and communal practice – for political life. We end in the conclusion: “De-training of judgment” with some reflections about the political ramifications of recommendation engines and their effects of de-training. Specifically, we seek to explain the emerging dominance of recommendation engines in cultural digital platforms in terms of a long historical trend whereby human skills, acquired and maintained by a painstaking process of training, are transferred to machines. Building on the historical precedent of physical and embodied skills, which have been transferred to technology during the industrial age, we now portray the process whereby a human skill of aesthetic judgment – which requires training – is transferred to algorithmic technology.
Training in algorithms
One of the most distinctive features of digital media is arguably personalization. Personalization is accounted for by theories of media production, media consumption, and the political economy of the media, to name a few. It is dependent on platformization, a socio-technical environment which has matured in the last two decades (Poell et al., 2021). Personalization requires digital media platforms to collect vast amounts and varied types of users’ data, process them algorithmically, and render them into knowledge about these users. This knowledge is not only sold to third parties (e.g., advertisements) but is also integral to the operation of these platforms. For example, in the case of Amazon and Netflix, this knowledge pertains to users’ tastes, wants, and interests, and it is used in order to match relevant content and present it to them.
This simplified image, however, plasters over an important feature of a much more complex process, which we seek to lay bare and make the centerpiece of this article: training. In order for media platforms to supply relevant, quality, personalized recommendations for users, their algorithms need to be constantly trained. 1 Not only do they need a constant flow of raw data in order to remain relevant and up-to-date, that is, reflect a user’s current interests and tastes; more than that, algorithms are often custom-made for each user, and change over time and with the introduction of new data. In other words, new data changes not only the results of the algorithms but also the composition of the algorithms which yield these results. Recommendation engines – which serve as a case-study for algorithmic machines in this article – do not merely learn once who we are, but need to be trained day-in-day-out in order to keep adapting to new circumstances and maintaining their performance satisfactory.
But how do algorithms train? And who trains them? For sure there is quite a lot of labor involved in engineering these algorithms, a labor done predominantly by Amazon’s and Netflix’s employees. But all this work would be futile without the flow of enormous quantities of users-created data and continuous feedback. The algorithms underlying these platforms are not merely technical and mathematical devices; in order to produce results (i.e., recommendations for books and movies) which are meaningful and relevant for users, they need to think like humans, to be immersed in human ways of thinking, cognitively and culturally. And this can only be achieved by tapping into the human mind, which in this case means mining human-made data. The indispensability of mobilizing the cognitive abilities and cultural schemas of humans to train algorithms is perhaps no more epitomized than in the case of Amazon’s Mechanical Turk. Amazon is using its own crowdsourcing platform to hire workers to perform a kind of labor which its intelligent machines cannot yet perform well enough. This allows the intensification of exploitation and its broadening to new realms (Aytes, 2012; Bergvall-Kåreborn and Howcroft, 2014; Kwek, 2020). By dividing work into minute, simple tasks Amazon not only gets these tasks done but also mobilizes humans to train its own algorithms (Stephens, 2023).
The bulk of this project of training algorithms, then, takes place through turning users into trainers. Training, we’d like to intimate, is not merely an abstract and metaphorical way of describing how algorithms come to know and characterize users on digital media. Instead, it is a key practice in the operation of algorithmic systems. Indeed, it is an emic term, part of the technical discourse in the academic and professional field of data science (Crawford and Paglen, 2021). In this discourse, training refers to the methodology by which supervised learning is carried out, entailing the use of training sets and training data to train algorithms. In these sets, data is collected and labeled, so that algorithms learn to anticipate output from input. In the context of Netflix, such labeled data might be data related to the film, such as genre, actor/director, date of release or original language, and can also be related to users’ explicit ratings (thumbs-up/down). So, for example, if a certain user has recently watched Sci-Fi films, Netflix will most likely suggest them new releases of Sci-Fi films.
Yet, Amazon and Netflix likely use not only supervised machine learning but also, and increasingly so, unsupervised machine learning algorithms in order to find clusters (groupings in data) and patterns by association. But here too, training plays a part: When using unsupervised learning, you have input data (X) and no output variables. In contrast to predictive models that forecast a target of interest, no single feature in a descriptive model is more essential than the others. Unsupervised learning describes the process of training a descriptive model because there is no aim to learn (Chakraborty et al., 2022: 28).
In this model the training does not take place through input data or a training set but through the performance of an action: finding a cluster or finding a pattern. In the context of Netflix, unsupervised machine learning takes the form of collaborative filtering that analyzes patterns in users’ behavior, such as their viewing history, ratings, and interactions with the platform. Based on these patterns the algorithm identifies users who have similar tastes and recommends them content that those similar users have enjoyed. So, for example, if the algorithm finds that users who have rated Sci-Fi films highly also enjoyed action thrillers with female leading characters, it will recommend action thrillers with female leading characters to other Sci-Fi enthusiastic viewers.
Training underlies both supervised machine learning and unsupervised machine learning, albeit in different forms. While training in supervised learning takes place on the label of the input data (through training sets), training in unsupervised learning takes place in the performance of the action: recognizing relationships and forming clusters, which far from being a neutral technological maneuver is a methodology that has a history, is embedded with ideology and requires training as Wendy Chun (2021) shows in her discussion of homophily.
What makes training, rather than learning, the key practice in creating and maintaining algorithmic decision-making machines? The answer lies in two particular features of the epistemology of algorithms, which underlie how recommendation engines know any specific user, and which can be contrasted to previous epistemologies for gauging the audience, or consumers’ characteristics. Presumably, algorithmic analysis of big data is dictated solely by universal, objective, mathematical assumptions (Fisher, 2022). However, the social study of really existing algorithms (those integrated into everyday cultural devices) has revealed that they are rife with sociological, anthropological, and psychological theories about human beings, and embody preconceived human ontology. The way these assumptions are integrated into the operation of algorithms is through human training, through their constant engagement with human actions.
Having no theoretical or ontological assumptions about human behavior, algorithmic analysis makes no hypotheses about the link between variables (e.g., what links ‘gender’ to ‘favorite film genre’). It therefore seeks to collect as many variables as possible, however preposterous their link to the dependent variable may seem; its appetite for variables is omnivorous and knows no limits (Mayer-Schönber and Cukier, 2013). So, in order to gauge users’ ‘favorite movie/book genre’, algorithms may collect and process variables such as ‘type of digital device used’, ‘time of day’, ‘outside temperature at time of use’, and so forth. Algorithmic knowledge, then, is purely deductive; it is derived from experience and hence requires training. In data science, this feature is commonly referred to as volume (lots of data) and variety (different kinds of data).
The second feature of algorithmic knowledge pertains to its dynamic character, which again explains its dependence on training (or on a constant cycle of learning, de-learning, and re-leaning). Algorithms assume no fixed human ontology. Indeed, they do not really see ‘a human being’, ‘a woman’, or ‘a New Yorker’ as an essence but as an ad-hoc constellation of data points, a constellation which changes as new data is extracted. In more simple terms of our case study, a user’s cultural taste (such as ‘favorite film genre’) is not constant but variable. Deborah Lupton accurately describes this as ‘lively data’, data which is always on the move and changing (Lupton, 2016: 42ff). Training the algorithm, then, happens not once; recommendation engines do not learn what a specific person wants, or what kind of a person they are. Rather, they need to be trained constantly. In fact, since there is no end point to that process, and since new data is always coming in and changing the algorithms, we can say that much like Schrodinger’s cat, the algorithm knows only when it is asked to deliver a recommendation (see: Cetina and Bruegger, 2002).
Training the algorithms of recommendation engines requires two conditions which are met in the ecology of digital platforms. First, that as much of human activity as possible will be performed on digital platforms, so that it leaves a high volume and variety of digital footprints. That explains the emergence and hegemony of platforms on the internet: everything takes place within these walled gardens, and (virtually) everything which takes place within them produces data (Fuchs, 2012; Zuboff, 2020). The second condition is that users are not merely caught within platforms but that these platforms are designed as to create as much activity of users which leaves data traces (Fisher, 2012, 2015a). This is done through engagement and interactivity practices, which are not merely end-products of digital platforms but its conditions of possibility. For example, personalized recommendations for books on Amazon or movies on Netflix can hardly be performed prior to users (as a social category, and as discrete individuals) engaging with these platforms, thus producing relevant data.
We have pointed out that one of the epistemic features of big data and algorithms is their absolute reliance on objective data – any event which can be registered digitally, and none which cannot. Hence, we can say that algorithms exclude the reflective and critical processes which take place within individuals, that is, their subjectivity (Fisher, 2022). The knowledge that algorithms produce is brought forth to users in the form of recommendations. These recommendations are somewhat of a hollowed-out reflection of users, representing only those aspects of themselves which can be registered as data. Notwithstanding this exclusion or omission, we cannot underestimate their effects: recommendation engines are increasingly determining the cultural diet of their users. Eighty percent of movies and TV shows watched on Netflix are based on personalized recommendations made by these engines (Chhabra, 2017). If users help to train algorithms to determine their tastes and interests, we would like to suggest that users are in fact also contributing to a process of de-training themselves. The process we seek to highlight is therefore dialectical: as users are training algorithms, providing them with a constant flow of personal data, and as this training allows algorithms to make recommendations which are deemed reliable, as a proxy to aesthetic judgment, users are also being de-trained from the process of reaching their own decisions, or their own aesthetic judgments (more on that in what follows).
Technology and training: A historical view
Algorithmic training, then, can be understood as a precondition for the process of automating aesthetic judgment. As such, we might think about that as the latest iteration in a history of training as a connecting tissue between humans and machines. Training (as well as de-training) has been part and parcel of the increasingly more intimate relations between humans and machines, particularly in the context of production technologies throughout the 20th century. Since training plays such a key role in labor, that is, in the preparation of humans to make their own world, it is little surprise that the notion of training as a connecting tissue between humans and technology has been thoroughly studied in the Marxist analysis of technology. In Americanism and Fordism (1971), Antonio Gramsci thinks of training in terms of regulation. Taking his cue from the rationalization of production at the beginning of 20th-century American industrialism – epitomized by the Ford Motor Company – Gramsci remarks that this endeavor ‘has determined the need to elaborate a new type of man suited to the new type of work and productive process’ (Gramsci, 1971). He merely alludes to the nature of training required in order to create a new type of human being by describing it as a ‘psycho-physical adaptation to the new industrial structure’ (Gramsci, 1971).
Living and functioning within an industrial order requires a continual training not only at work, but of the whole life cycle. At the time of writing, Gramsci assesses the success of the project to be partial: This struggle is imposed from outside, and the results to date, though they have great immediate practical value, are to a large extent purely mechanical: the new habits have not yet become ‘second nature’. But has not every new way of life, in the period in which it was forced to struggle against the old, always been for a certain time a result of mechanical repression? Even the instincts which have to be overcome today because they are too ‘animal’ are really a considerable advance on earlier, even more primitive instincts (Gramsci, 1971).
We can infer three tenets of training in industrial society from Gramsci. First, training is contra-natural; it is required to make a particular type of human, which would not have emerged otherwise. Second, training is processual and continual; as the material and technical environment changes it also requires a new phase of training and adaptation. And third, training is not merely individual, but societal and civilizational as well. Gramsci’s analysis elucidates a critical point for our understanding of training: humans are constructed and transformed through the productive process and through the technology that dominates it. It is transformed in such a way that renders humans malleable to and compatible with the new productive process and the broader social arrangements it entails.
But what precisely is the nature of training in the context of industrialism, which Gramsci merely outlines very briefly? A key tenet of industrialization – particularly automation, the assembly line, and Fordism – renders tacit, cognitive, and embodied human knowledge into information (Robins and Webster, 1999). This process entails extracting abilities and skills located within human bodies and minds and translating them into formal linguistic or mathematical information. Unlike digital platforms, where data is more readily available in a digital form, extracting tacit knowledge during industrialization required a hard labor of recording and analyzing manual labor and formalizing it into information. This endeavor was epitomized by the work of Frederick Taylor, who developed the academic and practical field of scientific management, a.k.a Taylorism (Taylor 1967). His endeavor entailed translating manual skills, for which massive stakes of individual training had to be carried out, into discrete units of information that were then arranged in an easily accessible form (a manual, a chart, etc.). The labor process could then be spliced down to minute fragments which could be taught much more quickly and with less painstaking efforts to unskilled laborers. Taylorism in fact facilitated the transformation from highly skilled manual workers – craftsmen – trained through apprenticeship, to an unskilled workforce, which could pick up simple tasks with ease. With time, these simple tasks could be transferred from humans to machines, a process commonly referred to as automation (Robins and Webster, 1999: 87–108; Mattelart, 2003: 37–40).
Two seminal works on automation can help us deconstruct the components of training in more detail. Training is a keyword in both Harry Braverman’s Labor and Monopoly Capital (1974), and in Anson Rabinbach’s The Human Motor (1992). Both works are important for our discussion on algorithms in so far as they position training at the nexus of humans and technology. Braverman analyzes Taylorism and the fragmentation of the labor process in terms of deskilling, the dwindling down of skills gained through a long process of training. This changes the relations of social power as skills are the most precious resource that workers have vis-à-vis capital: as unskilled workers, and then machines, learned how to do what hitherto only skilled workers could, such skillfulness was devalued and in fact disappeared. Embedding human skills and knowledge in machines drained workers of their skills as these were no longer required. With the introduction of scientific management – with its fragmentation of the labor process, its simplification of tasks, and the integration of automatic machinery – human labor itself became fragmented and simplified.
Anson Rabinbach (1992) explains the changing conceptions of labor in industrialism, resulting from the dominant scientific and technological transformations of the time. Helmholtz’s discovery of the laws of energy as the underlying operational rules of the universe and their construction as universal and abstract entities were central to a revolution in the perception of humans as machines for the extraction of labor power. This new discourse had led to the emergence of a new science of labor – the most renowned practitioner of which was Taylor – that in turn led to new bodily practices involved in the labor process.
The emerging discourse on the industrial human at the beginning of the twentieth century was intertwined with practices that in effect created a new type of human, of which Gramsci talked. Rabinbach recounts how ‘the emergence of a physiological approach to labor coincided with important changes in work during Europe’s second industrial revolution’, which involved the introduction of ‘electric power, of steel and chemical production, and of the rise of industries producing heavy machinery’ (Rabinbach, 1992: 122). The new factory introduced new types of workers (unskilled immigrants, leading to rapid turnover) and a new labor process that eliminated craftsmanship and was determined by new technologies (Rabinbach, 1992: 123; see also Noble, 1984).
We have used the history of training in industry to link technology with training. But if we consider digital platforms as a production technology, and users as workers (Fisher, 2012, 2015a, 2015b; Fuchs, 2013), we can unearth a similar process happening now with recommendation engines. Our argument about the training of algorithms on digital platforms follows a similar path but shifts the critique from a Marxist critique focusing on labor to a political, Arendtian critique focusing on judgment. Training the algorithms to ‘make aesthetic judgment’ entails a corollary and dialectical process of de-training users; as users train the algorithms, they become de-trained.
It could be argued that recommendation engines do not aim at making aesthetic judgments at all but just at sustaining attention, since the aim of the platforms is to keep users engaged. Although a true argument, we believe it does not change recommendations engines’ social and cultural effects. As aforementioned, eighty percent of watched content on Netflix is based on recommendations users receive from the engine. (Chhabra, 2017). Therefore, we assume most users see these recommendations as valuable. Our argument points not to what recommendations aim to do but to how users engage with them. By accepting the recommendations generated by the platform, users are de facto treating the recommendations as personalized, that is, as good-enough. We argue, then, that most users treat the recommendations as a good approximation of their judgment, thereby delegating their ability for judgment to the platform. By doing so, they further distance themselves from engaging in judgment as a historical, social and political faculty, essential for developing critical thinking. In the last section, we explain the political significance of this delegation, linking aesthetic and political judgment.
But let’s remain for another brief moment with the historical comparison and delve on the justifications for the training of machines and the concurrent de-training of humans. Granted, the process of automation and deskilling was motivated by the attempt of capital to increase its control over labor, and could be achieved due to differential power relations between them (Noble, 1984). But ideology also played a central part in automating the labor process. Such a deep transformation of how labor and industry are carried out, which affected people’s material, cognitive, and emotional well-being, was carried over the wings of an ideology of modernization, describing the rationalization of the labor process as universal, objective, and benevolent. At the heart of this ideology were unquestionable, depoliticized ‘goods’ such as efficiency, productivity, and technology-as-progress (Noble, 1999); even scientific management was claimed to represent objectivity and universal rationality (Robins and Webster, 1999).
The process of training algorithms to make aesthetic judgment and de-training people is likewise based on power relations and ideology. Training the algorithms is not a voluntary process: if an individual wishes to participate in the digital world – which at this point in the Western world, would be almost impossible not to – they have to consent for their data being monitored and rendered into knowledge. By using platforms such as Amazon and Netflix, users are also inadvertently training the algorithms, helping them improve with each use, that is, create better knowledge. Where personalization is involved (which is increasingly the case in many digital platforms), user-generated data is also used to train personalized algorithms. This is a dimension of power relations between digital platforms and individual users which is a structural precondition for the training of algorithms, which require a plethora of data.
But here, too, there is another dimension, crucial for the training of algorithms and for the corollary process of de-training users: ideology. At the heart of personalization through data is a promise to help us ease The Burden of Choice (Cohn, 2019), without undercutting our individuality. We may be at ease letting others direct our cultural diet as long as we feel that it is not forced on us from the outside but actually reflects our true tastes and preferences. Recommendation engines presumably do not undercut individualism but instead strengthen it, giving it powerful technological tools. This promise should be understood against a century and a half of mass media which targeted entire social groups and populations according to large scale categories, providing them with essentially homogeneous content and turning them into homogeneous target groups and eventually audiences. The homogenizing character of the mass media has already been criticized through and through. Particularly, it has been shown to construct audiences (rather than merely cater to existing ones) and to impose particular tastes on them (Ang, 1991; Bermejo, 2009). Recommendation engines, however, operate differently, using personalized parameters tailored for each user. The changing mode of operation requires new tools for critical analysis. Thus, in this article we seek to critically analyze personalized media applying an Arendtian approach for the evaluation of recommendation engines and show that contrary to the ideology of personalization, recommendation engines do not empower our individuality but in fact undercut an important component of it: our capacity to reach aesthetic judgment on our own, a capacity which is dependent on the constant training of individuals and communities.
Training for aesthetic judgment
The recommendation engines of Netflix and Amazon are trained to recommend to their users the best cultural object to fit their wants. To perform this task, recommendation engines need to ‘know’ what kind of books or films a specific user would like to read or watch at a certain time of the day on a certain day of the week, as well as ‘choose’ from the massive pool of books and films those few that would appear under the category: ‘top picks for you’ (Hallinan and Striphas, 2016). The process of selection that aims to match a user’s taste, that is, deciding ‘this is a good film to watch’, ‘this is a good book to read’ entails performing an aesthetic judgment. Every time we choose to watch a film or read a book, we assume that a specific cultural artifact would be a better choice than others, that is, we make an aesthetic judgment. Training, as we have shown in previous sections, is a powerful feature of algorithmic recommendation engines, which seeks to suggest personalized recommendations to users. Training, as we will show in the next section, is essential for the human capacity of making aesthetic judgments. The centrality of training as a key practice of both humans and machines underscores our analysis. To more fully understand the relationship between the training of humans and machines we explore in this section the nexus between taste as a feature of human aesthetic judgment and taste as it is constructed in Netflix’s ‘taste communities’.
But first, a few words about how we understand judgment. Judgment is the faculty that helps us organize and understand our realities, it ‘gives coherence and meaning to human experience’ (Zerilli, 2005). Aesthetic judgment is the essential building block in what we call culture. Judgment, according to Arendt (1981: 216) is a separate faculty of the mind, which has its own ‘modus operandi, its own way of proceeding’; it is a distinct capacity which has nothing to do with scientific reason, logical deduction or induction. In her reading of Kant, Arendt sees judgment as the ability to think of particulars as contained under a universal. Alas, particulars are not mimetic to the universal, since the universal is not a given but must be discovered out of particulars (D’Entrèves, 2006). Thus, we cannot say ‘we like roses; hence, we like this rose’ because we may like roses but we may find this specific rose ugly. We may have a vast knowledge about roses, we may know the ‘truth’ about their origin in 18th century Asia and their evolution into the cultivated popular rose of our time, and indeed this knowledge may help us appreciate the rose we see in our garden. But the knowledge, or truth, we hold regarding roses cannot explain the beauty of the particular rose we now see. There is nothing necessary in the beauty of a specific rose; to decree it as beautiful we need to exercise judgment. Judgment cannot supply a practical reason or a logical reason for declaring ‘this rose is beautiful’ or ‘this rose is ugly’. We cannot ‘prove’ its beauty or ugliness by reference to its ontology. The claim about the beauty or the ugliness of the rose is not grounded in a property of the rose, its history, or its ‘truth’, but in the way we look at the rose. Thus, beauty is not a quality of the rose but an expression of the pleasure we feel while looking at the rose (Zerilli, 2005). And therefore, training (and not learning about roses, or analyzing them) is key for developing an aesthetic judgment about the rose and judge it as beautiful or not.
If aesthetic judgment is not logical, inductive, or deductive, what is it then? Kant notes three conditions for aesthetic judgment: subjectivity, universality and communicability. A judgment of taste, Kant argues, is based on subjective pleasure, different people experience pleasure in different ways, from different objects, and independently from other people. This is the reason that aesthetic judgment requires freedom, reflexivity, and autonomy. At the same time, while aesthetic judgment is a personal and subjective practice, it is also a practice that has a communal aspect: it requires universal validity and communicability (i.e., means to make this validity public). In order to form a judgment of taste, individuals need to appeal to the judgment of others. The validity of their judgments rests on the consent they can elicit from a community of different subjects about specific objects: this ability needs to be communicated and tested in the public realm. It is in the public realm, through communicability, that the beauty of the artistic works of Picasso, received universal validity. The public realm is where different individuals can freely express their opinions on particular matters, make an argument and see whether they agree or disagree with others. When we judge, Arendt (1992: 72) maintains, we must always judge as members of a community, guided by what we all have in common, because it is through this common that we belong to a human community. Forming an aesthetic judgment is not a solitary act but a process that requires a genuine encounter with others – a community; and in turn, aesthetic judgment is also what forms communities. Judgment then, is an ongoing process that requires a constant evaluation of ideas, it cannot be performed by deduction or induction or by following a set of rules, in such a case, also an automaton or algorithmic engine could judge. Judgment, Kant (1999) emphasizes, ‘can be practiced only, and cannot be taught’, and for this reason it is a unique performative action, a practice that requires a constant analysis of specific objects constrained by changing circumstances, and therefore a constant training.
Reading Kant, Arendt highlights a seemingly paradoxical feature of judgment: it is subjective and personal, but it needs a community within which it can be validated and universalized. For Arendt, the community, or the voices of others are present also in judgments that we make on our own. For example, when one says ‘this is beautiful’ one does not mean ‘this pleases me’ (in the same way that we might say that dark chocolate pleases us, acknowledging that it might not please others who like milk chocolate). Rather, One Implicitly demands agreement from others, because in every judgment we make we ‘take into account’ the judgment of others and hence hope that our judgments ‘will carry a certain general, though perhaps not universal validity. The validity will reach as far as the community of which my common sense makes me a member…’ (Arendt, 2003: 140). Put differently, when we make a judgment of taste, even if we do not communicate it to others, the perspectives and voices of others are part of our internal dialogue. But of course, this does not mean that we are not capable of making judgments on our own, or that when making them we need to align ourselves with the judgments of others. Nevertheless, we cannot claim that our judgment is a completely individual judgment, detached from that of others. In Arendt’s words: ‘while I take into account others when judging, this does not mean that I conform in my judgement to theirs. I still speak with my own voice… But my judgement is no longer subjective either, in the sense that I arrive at my conclusions by taking only myself into account’ (Arendt, 2003: 141). Training for judgment depends on a community because when one judges one does so as part of a human community. Algorithmic engines, at the service of corporations as Netflix and Amazon, cannot create communities. Their aim is to maintain attention, keep users engaged with the platform, not generate deliberation. Removing the community from the ability of making judgments erases the ability to express disagreements and the ability to convince each other about the value of a certain cultural object. The removal of human communities from the process generates recommendations that masquerade as judgment, de-training users to deliberate and eventually make their own judgments.
For Arendt, the deliberative process that takes place in one’s mind cannot replace the actual deliberation with others: the open communication and exchange of ideas in the public sphere, and the effort to convince others about our opinions. Communication, Arendt (1992: 70) reminds us, should not be confused with expression. While communication is a dialogical experience, and therefore requires an interlocutor, expression does not (sounds can be a form of expression while for communication we need words). Communication originates in the Latin verb communico (communicate, share) from the Latin adjective communis (common, commonplace) which in turn serves as the root for the word community. The relationship between community and communication, evident in their shared etymology, has for Arendt a distinct political implication which she again ties to judgments of taste. The process of forming an aesthetic judgment requires a sincere encounter with different opinions. To form a judgment of taste one needs to appeal to the judgments and opinions of a community, that is, one needs to communicate with others who may have different opinions and engage in a deliberative process of exchanging arguments. The validity of these judgments rests on the agreement one can elicit from a community and is in turn what forms a community. It is precisely this back and forth of communication between individuals and their communities, a central pillar of aesthetic judgment, that requires a constant training.
Interestingly, community is also a key term in the discourse of Amazon and Netflix. Amazon has a detailed community statement that sets the limits of what a community is, who the community members are, and what can or cannot be said by community members; Netflix speaks about ‘taste communities’. Presumably, their use of the term presents a narrow meaning of ‘community’: it reduces community members to customers, a far cry from the meaning Arendt assigns to it. Yet, when the term community is used by these companies in the context of recommendation engines, things get more complex. Netflix recommendation engines, observes commentator Johan Steyn: are able to cluster people who have the same viewing habits, using machine learning that their resulting predictive algorithms use to create ‘taste communities’. The Netflix recommendation engine filters more than 3,000 show titles and 1,300 recommendation clusters at a time for about 195-million users in more than 190 countries. This makes it easier and quicker for customers to locate the shows they desire to watch (Steyn, 2022).
‘Taste communities’ might be thought of as the new communities of digital culture, since they allow people around the world to find a shared interest in specific cultural objects. Moreover, platforms, such as Amazon and Netflix, have long introduced features that allow users to give feedbacks to specific cultural objects in the form of stars (Amazon) and thumbs up or down (Netflix). Yet, members of these communities cannot deliberate about these objects or their shared interests nor can they exchange opinions on the platform; they can, therefore, not train for making aesthetic judgments. The feedback feature may be understood as expressing one’s taste, rather than communicating it. In platforms, such as Amazon or Netflix, users’ judgments of taste, that is, their choices on the platform, are recorded by the recommendation engine algorithm, which processes them as data and renders them as future recommendations, forming echo chambers or loops of ‘taste communities’ which are communities that exist without communication.
Netflix might talk about ‘taste communities’, but in fact what they construct are aggregates of discrete individuals. From the point of view of individuals, they do not see themselves as members of a community, since they cannot know who the other members of that community are, or what they think about the recommendations they are served. Indeed, they cannot even imagine them to be part of the same community (Anderson, 2006). But Netflix can. Johan Steyn explains in his opinion column: Netflix is able to comprehend the psychology of its clients, thanks to the data it collects. It can thus modify its customers’ experiences by employing landing cards: images or video trailers customized to what the individual clients would likely click on (Steyn, 2022).
Landing cards, video trailers, artwork, the floating rows of tiles that arrange the suggestions into categories, or even the film that starts playing automatically when Netflix is turned on, are different for different users: for each user, Netflix creates a slightly different cultural object, according to their personal inclinations and the company’s personalization strategy (Eklund, 2022; Pajkovic, 2021). Thus, while strengthening personalization as a key feature of their marketing strategy, Netflix weakens the common aspects of cultural objects. The hyper-personalization that Netflix propels by their customized user experience deteriorates that which ‘we’ – humans – have in common and can potentially make us into a community: the common experience of a common world. This communality, based on aesthetic judgment, and entailing subjectivity, universality, and communicability, cannot materialize without training. Training is essential for making aesthetic and political judgments because training allows us to be attuned to the changing circumstances of the world around us and to our own changing subjectivity, it directs us to pay attention to the judgments of others while making our own judgments and engage in a deliberative process of communication about subjective taste and feeling, as part of a community.
Linking aesthetic and political judgment
We have thus far focused on the aesthetic judgment of films and books. But as aforementioned, we wish to draw political implications from our engagement with the cultural field. In this we once again follow Arendt. In her effort to form a theory of political judgment, Arendt turns to Kant’s theory of aesthetic judgment. She explains this rather peculiar theoretical maneuver as follows: The reason why I believe so much in Kant’s Critique of Judgment is not because I am interested in aesthetics but because I believe that the way in which we say ‘that is right, that is wrong’ is not very different from the way in which we say ‘this is beautiful, this is ugly’. (Arendt, 2018: 382)
For Arendt, the mechanisms of decision-making regarding aesthetic judgments – which involve subjective taste and feeling, the persuasion of others, and the activation of a sensus communis – are very similar to the mechanisms used to make political judgments. Zerilli explains: [...] Judging is political, not because it is about political objects that are prior and external to it, but because it proceeds by taking into account the perspective of others and does not rely on an algorithmic decision procedure or the mechanical subsumption of particulars under known rules (Zerilli, 2016: 10)
Politics, then, is procedural, not ontological. And the process of judging an aesthetic object is similar to the process of judging a political situation. Not because these are similar objects or similar experiences but because the method of forming a judgment about these very different objects is similar. Thus, aesthetic judgments, which we make when we choose which film to see or book to read, provide a form of training for political judgment. Indeed, historically, aesthetic critique preceded political critique in the public realm, and in fact allowed the constitution of a public sphere (Habermas, 1989: 51ff).
Aesthetic judgments, as political judgments, require subjectivity, reflectivity, communicability, and freedom, as well as taking into account a community sense. This form of reflective judgment is what Arendt calls representative thinking. In her essay ‘Truth and Politics’ (2006b), Arendt explains representative thinking as the capacity to make present those who are absent and to represent the standpoint of others with the help of the imagination. Representative thinking is for Arendt (2006b: 221) ‘one of the fundamental abilities of man as a political being insofar as it enables him (sic.) to orient himself in the public realm, in the common world’. We are never alone while forming an opinion because also in our minds, we are in a continuous conversation with others. This conversation with others is a specific way for training the ability to make judgments and it requires, in turn, training the imagination, which is the quality that allows us to have as many perspectives present in mind when making judgments.
By taking into consideration as many perspectives as possible we enlarge our mentality and train a representative judgment which is for Arendt (2006b: 237) essential for political thought: Political thought is representative. I form an opinion by considering a given issue from different viewpoints, by making present to my mind the standpoints of those who are absent; that is, I represent them … The more people’s standpoints I have present in my mind while I am pondering a given issue, and the better I can imagine how I would feel and think if I were in their place, the stronger will be my capacity for representative thinking and the more valid my final conclusions, my opinion.
The more perspectives one holds, and the more conversations one maintains the more valid one’s opinion will be. Validity, here, should not be read as truth, logic, or as a scientific standard but as an opinion that embraces the common sense. A single standard, such as a scientific standard, or an algorithmic based recommendation, as in our case, would push away the need to persuade others of the merit of our own opinion. A single scientific standard for judgment would be also dangerous because it would erase the plurality, central to the human condition, and would mean that individuals are no longer required to train their imagination to go visiting before making judgments, since they would no longer be required to envision other perspectives, deliberate, and persuade each other of the validity of their opinion.
Imagination is a central force to the faculty of judgment, according to Arendt, because it allows us to position ourselves in other places where we are not and ‘go visiting’ other perspectives before making judgments. This process of deliberation between one and oneself, as Bilsky (1996) notes, enables us to distance ourselves from our particular circumstances and familiarize with the circumstances of others with the help of the imagination. Through the imagination we can visit other people and other places, and thus experience how we would feel in their shoes – not being them but being in their situation, this is what Arendt calls following Kant an ‘enlarged mentality’.
But how to mobilize this force of the imagination vital for making judgments? Arendt (1992: 43) answers, by emphasizing the centrality of training in this process: [T]he force of imagination makes the others present and thus moves in a space that is potentially public, open to all sides; in other words, it adopts the position of Kant’s world citizen. To think with an enlarged mentality means that one trains one’s imagination to go visiting [italics added].
Training is the action that generates and brings about the imagination, it is the action that allows the imagination to enlarge our mentality. Training is needed because the imagination is not a given, or something that just appears out of thin air. Judgment, much like imagination, is not something we are born with, but, instead, requires training, much like running, biking, or maintaining an ongoing deliberation between one and oneself and between us and our communities.
Conclusion: De-training of judgment
Arguing, as we are, that the more recommendation engines train in deciphering our projected aesthetic judgment, the more human-users are de-trained in that same practice, should not be read as an a-historical, and a-sociological statement. It is meant to underline the kind of force which recommendation engines generate upon users. But that is not the end of the story and certainly does not render humans complete cultural dopes, or objects at the hands of algorithms. Indeed, a plethora of research has underscored the active role of users vis-à-vis algorithmic machine. Even though the particular operations of algorithms remain largely under the guise of a ‘black box’ (Pasquale, 2015), users nevertheless form some knowledge about these operations. This can span from mere awareness that algorithms are responsible for making their recommendations (Gruber et al., 2021; Siles et al., 2022), to devising more substantial folk theories about how algorithms work (Siles et al., 2020; Ytre-Arne and Moe, 2021), to being skilled with them (Gruber and Hargittai, 2023).
A productive framework to think about this range is Taina Bucher’s general notion of ‘algorithmic imaginaries’ (Bucher, 2017), pertaining to the assumptions users make about how algorithms work, and under which they operate. These forms of algorithmic awareness, literacy, or consciousness, then, create a feedback loop: users take into account (assumed) algorithmic operations when fashioning their own actions, which in turn changes the algorithmic output. This process has been studied under the notion of domestication, by which users practically integrate media technology into their lives, and at the same time negotiate their operations and thus their effects (Berker et al., 2006). For example, Siles st al. (2020) have shown how the assumed workings Netflix algorithmic recommendation engine influence users’ use of them. When the engine seems to be out of tune (e.g., if another person has been watching Netflix under your account), users try to tweak it back by playing more suitable titles (not even bothering to watch them) thus ‘manually’ training the algorithm. As this example implies, at the end of the domestication spectrum, users might even resist algorithmic governance (Witzenberger, 2018).
Focusing on algorithmic recommendation engines, and leaving aside the question of imaginaries, domestication and resistance, as we do, has a double rationale. First, in this article we wanted to point out a particular (qualitative) type of dynamic between humans and technology (along the axes of training and aesthetic judgment), leaving the question of their actual (quantitative) influence aside for further research. And second, notions such as imaginaries, domestication and so forth, are all underpinned by the idea that users have folk theories about algorithms which inform their engagement with them. Our purpose is not to engage with these theories but actually say something (positivistic) about how they operate. Put differently, we assume that de-training is not phenomenologically experienced by users; instead, it is our analytical contribution to the discourse on algorithms.
In this article we sought to explain the emerging dominance of recommendation engines in cultural digital platforms in terms of a long historical trend whereby the particular human skill of aesthetic judgment, acquired and maintained by a painstaking process of training, is transferred to machines, which on their part cannot perform it, but act as if they could. In this context, it is crucial to remember: what recommendation engines do is
One thing that aesthetic and political judgment have in common, according to Arendt, is their non-coercive character, that is, that ‘they can only appeal to but never force the agreement of others’ (Arendt, 2006a: 222). In her essays ‘The Crisis in Culture’ (2006a) and ‘Truth and Politics’ (2006b), judgment appears as a faculty that allows political actors to decide what actions to undertake in the political realm. Training, as we have shown, is essential for aesthetic judgment, which, in turn, is essential for political judgment. In The Origins of Totalitarianism, Arendt (1973) shows how totalitarian ideologies benefit from a preceding degradation of sensus communis in societies where they take hold, because people can no longer tell fact from fiction. Common sense, she writes, is related to our sense of the real and its operation is opposed to totalitarian ideological thinking. When totalitarianism rises to power it destroys the remains of the sensus communis. By disabling discourse, totalitarianism isolates and encloses human beings in their own minds. Ultimately, isolation undermines one’s sense of self, which is formed in relation to others, and one’s sense of reality.
Our sense of reality is not only at danger under totalitarian regimes but more broadly – it could be argued – by the ‘degradation of culture’ into entertainment. Indeed, Arendt is among those who contributed to this critical discourse. Culture and entertainment are, for Arendt, two different things. Culture is what societies choose to cultivate, what endures. Cultural artifacts are those which ‘comprehend and give testimony to, the entire recorded past of countries, nations and ultimately humankind’ (Arendt, 2006a: 199). Entertainment, on the other hand, ‘like labor and sleep is irrevocably part of the biological life process. And biological life is always, whether laboring or at rest, whether engaged in consumption or in the passive receptions of amusement a metabolism feeding on things by devouring them’ (Arendt 2006a: 202). Thus, while culture lasts, entertainment has to be constantly remade, therefore, ‘culture relates to objects and is a phenomenon of the world; entertainment relates to people and is a phenomenon of life’ (Arendt 2006a: 204). In this paper, our aim is not to make a judgment regarding the nature of what Netflix or Amazon recommend their users. Part of it can fall under what Arendt calls entertainment and part of it can fall under what she calls culture. Our argument is that recommendation engines erase the difference between culture and entertainment; they erase the distance between what pleases us and what we judge to be good. In this way we see the ‘crisis’ in our current digital culture. Not in the degradation of culture into entertainment but in the erasure of the distance between them through the platformization and personalization of recommendation engines.
The personalized marketing strategy of platforms, such as Netflix and Amazon, carried over an ideology of personal empowerment, undermines users’ training for aesthetic judgment and by extension, their training for political judgment. Not because the cultural products that these platforms suggest (and increasingly create) deteriorate their judgment, but instead, because the process of making judgments, which entails representative thinking, community and communication, is being deteriorated. The personalization that recommendations engines promote isolates and encloses human beings in their own minds. Their judgments of taste no longer need to be communicated but merely expressed, by watching a suggested film or not; and judgments no longer need to be negotiated as part of a community. Thus, while algorithmic recommendation engines are mastering their training for ‘making aesthetic judgments’ for particular users, the human training for making aesthetic, and in turn political judgments, is being deteriorated. This training, which requires to evaluate the specific under the universal, to cultivate a representative thinking, and to keep in mind the sensus communis that bonds us into a community, is essential for cultivating the ability to form judgments about new political circumstances and contested futures.
Footnotes
Acknowledgements
We are thankful to the reviewers of this article for their valuable observations.
Declaration of conflicting interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) received no financial support for the research, authorship, and/or publication of this article.
