Recommendation Systems as Technologies of the Self: Algorithmic Control and the Formation of Music Taste

Recommendation Systems

Since recommendation systems could not have been possible without recommender algorithms, these two terms are often used interchangeably. This, however, does not mean that such systems can be reduced to a mere code but should rather be seen as a complex assemblage of information exchange consisting of multiple elements and processes. Although their design can be based on different computational paradigms, ¹ ultimately all recommender systems gather data about particular characteristics (e.g. interests, tastes, or curiosities) of an individual user and, on the basis of that data, offer information and services specifically tailored for that user.

The origin of recommender algorithms is often traced back to a practical/economic problem in online delivery systems. Offline information and product delivery has a limited scope (e.g. the number of books an offline store can carry is physically limited). In offline establishments, therefore, there is usually a pre-selection process that prioritizes the items with broad public relevance. Virtual delivery systems, in contrast, can offer an almost unlimited amount of information, which ‘forces on-line institutions to recommend items to individual users’ (Leskovec et al., 2014: 309; emphasis added). Recommender systems, therefore, can be seen as a product of ‘algorithm providers’ attempts to thoroughly know and predict their users’ (Gillespie, 2014: 168), in order to deliver more ‘personalized’ information for ‘customer satisfaction’. Thus, as is already observed in the literature on search engines (e.g. Hargittai, 2007; Halavais, 2013), such algorithms can act as critical mediators that shape the communication between users and their cultural ecosystem (Gillespie, 2014: 167), exercising power by excluding/including and organizing cultural information (Morris, 2015: 451). Furthermore, by combining this ‘mediatory power’ with algorithmic personalization, recommender systems can also function as technologies of control that guide their users’ activities and, thereby, constitute the technical basis of a system of ‘soft’ power (Beer, 2009; Cheney-Lippold, 2011; Hallinan and Striphas, 2014; Morris, 2015). Beer (2009: 997) explicates that this ‘soft’ power is not a matter of ‘someone having power over someone else, but of the software making choices and connections … in order to shape the everyday experiences of the user’.

As also noted in some of these studies, however, we know little about how users experience/utilize these systems. This question is particularly important for elucidating what Gillespie (2014: 168) calls the entanglement of algorithms with practice – a notion that has important parallels with similar, though broader, concepts in Science and Technology Studies literature such as ‘co-construction of users and technology’ (Oudshoorn and Pinch, 2003) or ‘co-emergence of subjects and objects’ (Day, 2011). What these concepts underline is that, in practice, users neither passively submit to the guidance/control of algorithms, nor remain completely unaffected; the effects that an algorithm can produce on users are co-constructed through ongoing ‘negotiations’ between the two.

If one crucial implication of this entanglement, then, is that the effects of algorithms can be reshaped through user practices, the other is that ‘users reshape their [own] practices to suit the algorithms they depend on’ (Gillespie, 2014: 168; emphases added). And, as users try to modify their own practices, they go through various processes of self-reflection and transformation. In short, a recommender system does not only mediate users’ relation to their socio-cultural environment but it might also mediate their relation to themselves, which brings us to our next topic.

Technologies of the Self

Introduced to the current literature through Foucault’s later works, this concept has been a major source of inspiration for the recent studies on digital/web technologies of the self. Here, we shall dwell on two points that are directly relevant for our concerns. ²

As is well-known, in his later works, Foucault (1990a, 1990b, 1993, 1997a, 2005) gradually shifted his attention from the question of how subjects are produced in knowledge-power networks to how human beings turn themselves into subjects, through practices of the self. ‘Practices of the self’, however, are not solipsistic endeavours but require complex assemblages comprising discourses, other people (e.g. disciples and teachers), institutions (e.g. schools or monasteries), meditative techniques as well as a whole range of tools (e.g. letters and diaries). It is such assemblages – and not just technical artefacts – that Foucault (1997b) delineates by ‘technologies of the self’ (see also Brenninkmeijer, 2010). As we shall see, recommendation systems too constitute such an assemblage.

Secondly, Foucault (1990a, 1990b) qualifies practices of the self as ethical because they are utilized to give an ‘ethos’ – i.e. a ‘character’ or ‘style’ – to one’s existence. More specifically, he conceptualizes technologies of the self as one of the four crucial dimensions of an ethical self-formation practice. Namely, a technology of the self is used by individuals: (i) to work on a particular aspect of their self (‘ethical substance’); (ii) with a view to give this aspect an ultimate form (‘telos’); and (iii) within the framework of historically variable ‘modes of subjectivation’, denoting whether such practices are carried out to conform to existing norms or to forge a ‘new aesthetics of existence’.

Foucault primarily focuses on the latter option in his later work, delineating the historical uses of self technologies in the constitution of certain areas of one’s life (e.g. diet or sex) as ‘substances’ of ethical care as well as aesthetic elaboration. As such, ‘care of the self’ does not just target what one thinks but is rather an askesis that aims to change how one lives and conducts oneself through self-disciplinary labor (Foucault, 1990b, 2005). We largely adopt this conceptual scheme here. ‘Music taste’, we shall suggest, constitutes an ‘ethical substance’ for some last.fm users, which they try to ‘improve’ with the help of the recommender system. The ‘collaborative work’ of users with the recommender system, however, does not only bring about a change in what they listen to but also in their conduct as music listening subjects. In other words, it is by mediating the users’ relation to their own conduct, and not only to musical products, that recommender systems function as self technologies.

Current studies on digital/web technologies of the self also pay close attention to the capacity of digital media to induce a self-reflective attitude in users (e.g. Abbas and Dervin, 2009; Aycock, 1995; Bakardjieva and Gaden, 2012; Haider, 2015; Rettberg, 2014; Sauter, 2013; Siles, 2012). So far, however, these pioneering studies have investigated this capacity in reference to personalized uses of social media, focusing predominantly on social networking sites, weblogs, micro-blogs, news groups and forums, and ‘virtual worlds’. These types of social media, it is argued, enable users to experiment with various forms of ‘self-writing’ or ‘self-publishing’, giving way to practices of self-discovery, self-mastery and self-care (Aycock, 1995; Bakardjieva and Gaden, 2012; Boellstorff, 2008; Haider, 2015; Rasmussen, 2014; Siles, 2012). The key to the use of such media as self technologies lies in their capacity to facilitate novel types of encounters between the self and others, ‘be they companions in virtual worlds …, readers (for example, of a blog), listeners (Podcasts), or viewers (YouTube, Dailymotion)’ (Abbas and Dervin, 2009: 2).

The insights of these studies, however, are not directly applicable to recommender systems, which are not primarily designed to facilitate self-other encounters in virtual space. Although, especially in the case of collaborative filtering, recommendations offered to a particular user are based on the data gathered from others (Jannach et al., 2010; Leskovec et al., 2014), this ‘mediation’ does not entail any actual interaction between the users. Indeed, other users become almost completely indiscernible in this process as their data are ‘assimilated’ by the algorithmic system, which acts as their spokesperson. Unlike in personalized social media, then, it is the self-algorithm interaction that matters most in this case. To grasp how recommender systems function as self technologies, therefore, we need to direct our attention to this interaction.

How Can a Recommendation System Function as a Technology of the Self?

A recommender system collects data about users, organizes that data into a meaningful form and makes recommendations. A closer look at these three operations is essential for understanding how such a system can function as a technology of the self. First, although some applications ask users to manually enter data about themselves, this is usually considered a tedious and unreliable method (Jannach et al., 2010). In most cases, a recommender algorithm ‘learns’ about its users by accessing the data about their previous online or offline activities. This means that a recommendation system does not record an already delineated characteristic of the user but, rather, it is through the operations of the system that this ‘characteristic’ is constituted and brought to the attention of users (see, for example: Hallinan and Striphas, 2014: 8–10).

Secondly, recommender systems can also incite users to reflect on this constructed characteristic because they respond to every activity/preference of users with a new set of recommendations. As Beer and Burrows (2013: 60) point out, this can be seen as ‘recursive feedback of data’, which ‘has a constitutive affect on people’s lifeworlds’. This ‘affect’, we shall suggest, is both exhilarating and exciting because the flow of recommendations demands endless self-reflection from users, while offering them ever new possibilities for transforming their characteristics. And, lastly, since users can influence their ‘profiles’ by modifying their activities, such systems also accompany users in this self-transformation process and mediate their relation to what appears to them as an ‘objectified’ aspect of themselves.

To recapitulate, we can imagine a hypothetical user who is initially engaged in sporadic online or offline activities. Once the data about these activities are loaded into a recommender system, however, they assume a specific form, signifying a ‘characteristic’ of the user. Next, recommendations begin to pour in, alluding to new activities/items that might be of interest to the user, and thereby invite the user to revise his/her current choices/characteristics. Finally, the user can go back to the starting point and attempt to modify his/her activities in light of these revisions. And, as this cycle keeps on repeating, users develop new ideas about themselves as well as new practices.

Let us, however, immediately note that recommender systems are employed in numerous different domains, where the above-mentioned properties often have no practical significance. Many commercial websites, for example, are designed to make one-time recommendations. In many others, user data and ‘profiles’ are not fully accessible to users. Nevertheless, some systems in current use do exhibit all the properties discussed above, as is the case with the empirical focus of this study.

The Scrobbler and the Library: Giving ‘Music Taste’ a Visible Form

The first stages of users’ interaction with last.fm as a recommendation system can be depicted in terms of a basic demand and a corresponding offer. ‘Tell us about your music taste’, last.fm says, ‘and we’ll give you great recommendations’.

But how does one ‘tell about’ one’s ‘music taste’? In practice, last.fm’s demand is fulfilled by users by downloading the program named ‘scrobbler’ to their digital music playing devices. The scrobbler is designed to transmit the data about the music listening activities of users to their ‘libraries’ in their last.fm profiles. In effect, it is this individualized music library that is supposed to signify the ‘music taste’ of a user.

It would, however, be quite misleading to think that one can ‘tell’ about one’s music taste simply by scrobbling since the information recorded by the scrobbler goes through a series of transformations: the songs ‘scrobbled’ at diverse points by a user are first arranged neatly in the form of a library, which is then interpreted as the representation of the user’s music taste. The library, therefore, is not the expression of a pre-given ‘music taste’. Rather, this ‘taste’ gains a visible form when the myriad of scrobbles are combined in the library. Contrary to what the last.fm message – ‘tell us about your music taste’ – seems to suggest, therefore, it is the library that tells users about their music taste.

Turning to the experience of users, the first point we should note is that the scrobbler is akin to a wearable technology that accompanies users and, as such, its ‘presence’ is amply felt. This finds one of its most vivid expressions in a group titled, People who have an Obsessive Scrobbling Syndrom [sic.], with over 2650 members and numerous discussion threads. ⁴ In one of those threads, users share their experience when there is a technical problem with the scrobbler such that their lists are not updated properly. Here are some highlights: ⁵

I start feeling like an incomplete person. And I get really depressive. (remco)

I begin to panic. These songs HAVE to be in my charts. I feel incomplete. … There’s just something that makes me have to scrobble EVERY song, really creepy. When I was on holiday a few weeks ago I could not take my iPod with me, because it was broken. The whole holiday I listened to CDs on my old discman … But! These songs would not be scrobbled!!! So when I came home, I, immediately turned on my iTunes and played all night songs which I listened to on my holiday … That’s how far my obsession goes. (Taralita)

This value, we should stress, does not simply derive from the fact that users can ‘present’ their libraries in a public platform – this is not a social media effect. Rather, the preservation of the scrobbling ‘data’ remains a matter of concern for many users because they themselves can ‘look at’ it and discern something there about their own ‘taste’:

[When last.fm fails to update properly,] I lose my will to listen to music because what’s the point really, if you can’t obsessively look at the stats afterwards. (pecusita) ⁸

Moreover, the library also includes information about the kinds of music/artists listened to the ‘top lists’ of the user, and the diversity or the ‘coherence’ of the overall collection. This gives the library a symbolic character, which calls for a rather intricate interpretive process. Here are some examples of how users elaborate on what gives ‘unity’ to the otherwise hard-to-interpret patterns in their libraries:

If you look at [my library in different phases of my life], there will always be something uniting. For example, I don’t know, when I was a kid, … there would always be this kind of ‘rebel’ towards society element in it. … But then [more recently], I would always look at something that calms me down. So, the music that I listen to right now, … is usually something more joyful … (Intw#3)

I’d like to listen to stuff that has refined instrumental parts and are relatively demanding in terms of rhythm and melody. I mean … even if they are from very different genres, I can listen to songs that involve unusual [rhythmic] measures. (Intw#4)

Beyond the Library: Recommendations and ‘Music Taste’ as an Object of Care

In a recommendation system, the information in users’ libraries is not only scrutinized by them. It is also fed into an algorithm that continuously recommends new items to the user. As one user puts it, it is as if the algorithm says, ‘if you really like this [song], I’ll give you others that you might like even more’ (Intw#2).

In effect, the recommendation algorithm on last.fm offers an interpretation of the library, which does not just specify what the present music taste of the user is, but also what it can be. Thus, last.fm recommendations perpetually hint at a ‘world’ beyond the existing boundaries of a user’s library. There is considerable evidence that users employ this ‘guidance’ not only to discover new music but also to revise and modify their current ‘music taste’.

One obvious effect of recommendations is to sensitize users to the ‘quantitative’ aspects of their ‘music taste’. Thus, typically, one user answers the thread, ‘What makes you happy?’ ¹¹ with ‘finally reaching 1000 artists in my library’ (Wothsthedeal), while another ‘confesses’ perpetually sampling ‘boring’ music. More importantly, new recommendations often motivate users to change their listening practices to give their ‘music taste’ a new form. One user, for example, stressed that when he noticed he listened to certain musicians ‘excessively’, he forced himself to listen to other artists (Intw#4). Similarly, another noted:

I become attentive to what I listen to too little and what genres escape from my sight [and] in order to keep my musical perception wide, I turn to these [new kinds of music]. (Intw#5)

But what exactly is this ‘desirable form’? Does last.fm provide a ‘telos’ for this process? Clearly, recommendations do not specify what ‘good’ music is in general, but what could be good for the individual user. Nevertheless, the condition of constantly receiving recommendations creates a tendency in the users to care about the diversity of the music they listen to:

Last.fm has changed me. Made me too self-conscious of my listening habits. Before, I’d play the same artist for days and days, but now I constantly struggle to diversify. I recently made a playlist called ‘diversify!’, it contains a lot of artist[s] that are close to my top 50, and I’d like to see them on my profile page. (pecusita) ¹²

Strictly speaking, however, diversification is anything but an unambiguous ‘telos’. The perpetual flow of recommendations ensures that diversification is never completed but remains an endless pursuit. As one forum contributor asks: ‘How do I get it so that my supposed “taste” in music becomes more diverse (as in I want to listen to EVERYTHING)?’ (Alkajak). ¹⁴ For, indeed, how does one build a perfectly diverse library if not by scrobbling all music ever created?

We shall therefore suggest that the pursuit of diversity, accompanied by the constant flow of recommendations, calls for a virtually endless care and cultivation of ‘music taste’. In fact, some users perceive ‘indeterminacy’ as the trademark of their ‘music taste’:

I do not have a definite music taste, I have an indefinite music taste (giggling) … Frankly, I have a still changing music taste. That’s the sense in which there is a general ambiguity [in my library]. (Intw#5)

Last.fm, then, functions as a technology of the self, not by urging users to interiorize a definite musical ‘content’, but by inviting them to modify their listening habits through recommendations. User comments indicate that in the absence of the recommender system, one might enjoy music without much concern about, say, how often one listens to the same song or how diverse the artists one listens to are. It is, in other words, by encouraging a change in this ‘careless’ conduct of the user as a music-listening subject that the recommender system plays a role in self-transformation. This, moreover, implies an askesis because ‘music taste’ is now perceived as the outcome of an activity – besides enjoying music one also listens to it with a view to improve one’s ‘music taste’. Thus, many users comment that they feel caught up in an endeavour that requires constant effort, which they nevertheless find exciting. The following remarks can give us a good glimpse of the peculiar ‘lifeworld’ ¹⁵ of these users:

I’ve been trying to diversify my musical taste … you get into the trap of finding more and more genres, and then you get overwhelmed. But it’s awesome! (BossHossV8Bikes) ¹⁶

More generally, one might ask how much novelty a recommender system can offer. While most current systems propose items similar to users’ previous choices, attempts to develop algorithms that can offer substantial novelty and serendipity already exist (Morozov, 2011; Shapira et al., 2011). In any case, our findings suggest that, even when users are incorporated into an algorithmic control circuit, this does not necessarily stop them from getting excited about cultivating their characteristics (like ‘music taste’) in new ways. We shall discuss the broader implications of this observation in the conclusion.

Recommendation Systems and Social Media: A Complex Co-existence

Many users, then, employ last.fm to reflect on and cultivate their music taste, which attests to the potential of recommendation systems to function as technologies of the self. However, we do not claim that this is the only use of last.fm. The forums indicate much variation in user practices, one major reason for which is that last.fm is not only a recommendation system but also has a social media component allowing users to publicly share their libraries (see Baym and Ledbetter, 2009). One might expect these two components to reinforce each other since, as is widely observed, visibility in social media makes internet users self-reflective. Such a simple correlation, however, is disputed in recent studies that underline the complex effects that algorithms have on social aggregation/disaggregation (Totaro and Nino, 2016). In any case, it would be quite misleading to frame last.fm merely as a platform for ‘presentation of the self’s music taste’.

Indeed, the availability of social media channels on last.fm often leads users to employ the tools on the website for quite different ends than cultivation of music taste. Most notably, some users try to impress others by scrobbling songs that they do not actually listen to. The occurrence of such ‘fake scrobbling’ is common knowledge among last.fm users and openly debated in the forums of two rival groups: ‘true listeners’, ¹⁷ who only scrobble when they listen to music, versus ‘untrue listeners’, ¹⁸ who say that they are ‘proud’ of doing the opposite. These debates reveal a striking tension between users who employ the recommendation system for self-cultivation purposes and those who employ the social media component for self-presentation – though these are not necessarily mutually exclusive tendencies. It is, for example, quite difficult to say whether the following user is actually ‘manipulating’ the system to impress others or genuinely modifying his listening practices:

sometimes i feel like I’m cheating on my top artists, for instance when I’ve discovered a new (to me) band and listen to them 24/7 so they’ll be in my top 20 in no time and kick some other artist out that i might have been listening to for ten years. (las_cruces_jail) ¹⁹

However, equally emotional appeals in the forums can also give way to the formation of identities around quite peculiar musical orientations. One forum group, for instance, invites users not so much to ‘enrich’ their library but to embrace a ‘random and questionable musical taste’. ²¹ This means social media channels enable users to invent various criteria for evaluating ‘music taste’, which differ considerably from the criteria used by the recommender algorithm. Finally, whereas a recommender algorithm allows users to interact with itself only as individuals, social media channels render these users visible as a public and give them a voice vis-à-vis the platform. In last.fm’s forums, for example, users critically discuss the characteristics of the platform itself and even demand modifications. ²² Thus, last.fm’s social media channels can give way to a ‘co-constitution’ process between the platform and its public – a trend that Van Dijck (2013) has already observed in other contexts.

All this, however, does not mean that the contents produced in social media are immune to algorithmic processing. As Gerlitz and Helmond (2013) note, for example, through the current integration of social media to the ‘Like economy’, the emotional content of user comments tends to be translated into ‘machine-readable’ quantitative data. Thus, rather than positing a sharp contrast between ‘humanly’ social media versus ‘cold-blooded’ algorithms, we propose to see this as a complex co-existence.

As a corollary, it is worth noting that whereas social media channels assume a person-to-person communication, which traditionally has been the main mode of sharing cultural information (Kayahara and Wellman, 2007; Williams, 2006), the design motto of recommender algorithms is often formulated as providing ‘better recommendations than humans’ (Jannach et al., 2010: xiv). But what exactly might ‘better’ mean here? Granted, the algorithmic system has access to ‘bigger’ data than human actors. But suggesting that this is sufficient for determining, say, the ‘taste’ of a flesh-and-blood user amounts to assuming that ‘information’ as ‘some mysterious entity is responsible for imbuing people and objects with shape, quality or character’ (Striphas, 2015: 407). We do not think that the term ‘better’ is warranted in this sense. What we can say, however, is that, in conjunction with their utilization as self technologies, recommender systems also offer a distinct type of ‘companionship’ to their users, which we consider next.