Exploring users’ algorithmic knowledge and reflexivity in a music streaming context: A critical realist approach

Abstract

Digital platforms such as Spotify have specific characteristics and properties that influence, to some extent, how the platform is used. However, users develop their own interpretations of these properties as well as unique ways to engage with the platform. This study applies a critical realist framework to explore how reflexivity modes are practiced in the context of Spotify as an example of algorithmic recommendation systems. From this perspective, reflexivity is a person’s capacity to reflect on their contexts, data, previous experiences, and knowledge, among other elements, before deciding how to act. Findings from interviews with Spotify users suggest that participants practice multiple reflexivity modes when interpreting Spotify’s recommendations and deciding what to listen to. These modes depend on each participant’s concerns and algorithmic knowledge.

Article

Music streaming platforms use algorithmic systems to recommend music to their users based on their behavioural data (Prey, 2019). In the same way that music industry actors reflect on the user data they collect, users can also reflect on the data they give, the recommendations they receive, the platform’s characteristics, and their own potential actions based on their concerns, goals, experiences, and knowledge. Research has shown that media users’ understanding of algorithms and streaming platforms is diverse and that users use different knowledge sources to develop conceptions of technology that lead to diverse reactions to the recommendations (Bucher, 2018; Cotter and Reisdorf, 2020; De Vito et al., 2018; Seaver, 2019; Siles et al., 2020; Swart, 2021). This previous research has examined how people think about technology and how algorithms shape users’ engagement with and reactions to content in digital platforms. Many of these studies focus on how people know algorithms. Bucher (2017) explores how encounters with algorithms in different situations generate an idea about how algorithms (should) work and what can be done with them in the form of algorithmic imaginaries. Similarly, research on folk theories about algorithms finds that users develop an understanding of how the algorithmic systems work based on their use with them (Colbjørnsen, 2018; De Vito et al., 2018; Siles et al., 2020; Ytre-Arne and Moe, 2020). Other authors have worked with the concepts of encoding and decoding, explaining how users develop different knowledges about algorithms based on information collected through their own experiences, mainstream media, or an educational setting (Lomborg and Kapsch, 2020; Shaw, 2017). Using these folk theories and knowledges to decode algorithms, users develop different ways of using and reacting to the platform. More recently, Kapsch (2022) applies the concept of ‘small acts of engagement’ to understand how users identify and respond to algorithms in their daily media use.

These studies hint at how reflexivity mediates the influence of a structure and the actions taken by users, as well as how users develop an understanding of algorithmic systems. Through this knowledge, users take an informed action towards algorithmic recommendations. By doing this, these studies address reflexivity in an indirect way by describing how people understand algorithms or by recognising reflexivity in a general way, often missing the different forms reflexivity can take. Thus, I argue that reflexivity is a broader process that can be used to explain how people think about data and algorithms, and I offer an interpretation of reflexivity as a mediator between structures, technology, and agency.

Recent work suggests that sociology, media, and communication research has not paid enough attention to reflexivity and how it is affected by media technologies, especially from a critical realist perspective (Ansaldo, 2022; Chernilo, 2017). This article studies user’s reflexivity when using music streaming platforms from a critical realist framework. Critical realist social theory argues that structural and cultural properties condition but do not determine people’s actions, with reflexivity being a mediator between structure and action (Archer, 1995, 2003, 2007, 2012). I follow Archer’s (2007) view of reflexivity to explore how users practice different reflexivity modes (autonomous, meta, communicative, and fractured) to think and react to algorithms. Reflexivity, as seen here, precedes and produces social practices and is considered as the capacity or mental ability people have to process information, experiences, emotions, and concerns, and consider their contexts before acting (Ansaldo, 2022; Archer, 2007; Wilson, 2006). In this way, the (imperfect) knowledge, imaginaries, theories, and previous experiences people have developed are part of this reflexivity process. Here, I examine how Archer’s reflexivity modes are practiced in a digital context to understand how reflexivity mediates the influence of technology and its affordances over media uses and practices.

The current study explores reflexivity in users of music streaming platforms and advances existing work on reflexivity in a media entertainment context by specifically referring to algorithmic recommendation systems. This is done by asking how users reflect on algorithms and how this thinking process influences how people use music streaming platforms. In the next section, the critical realist framework of reflexivity is explained and compared with other work in this area. The interviews are then analysed to examine different reflexivity modes in music streaming and how they relate to user agency in this context.

Theoretical framework

Within the critical realist philosophical framework, Margaret Archer (1995) developed the morphogenetic approach to explain the interplay between social structures, agency, and culture. From her perspective, social and cultural structures provide a context that conditions but does not determine action. Here, action has the potential to change or reproduce these structures. She argues that a person will develop interests that constrain action. In this way, human action is built on a person’s past and ability to plan and strategise for the future. To explain this, Archer explicates the concept of reflexivity, a key component of her model that acts as a mediator between social and cultural contexts and sociocultural actions (Archer, 2000, 2003, 2007; Caetano, 2015; Sayer, 2009).

Archer (2007) defines reflexivity as a person’s mental ability to consider themselves in relation to their (social) contexts, and vice versa. This is expressed as an internal or silent conversation people have with themselves, in which they plan, rehearse, decide, and contemplate things, among other mental activities (Archer, 2003, 2007; Caetano, 2015; Sayer, 2009). Thus, reflexivity mediates between structure and action, ultimately transforming a person’s interests and concerns into projects, goals, and potential actions to reach the goals (Archer, 2007). Thus, internal conversations are more than mere introspection and have a practical intent. Structural and cultural contexts shape events and provide constraints and enablements that are then interpreted subjectively by people. The effects these events might have will depend on how they are interpreted, as each person will go through reflexive deliberation to decide how to act accordingly.

Archer theorises reflexivity as a continuous activity and moves away from the reflexive and non-reflexive duality used in other theories to consider that people can be reflexive without realising it. She does this by arguing for multiple reflexivity modes that she uses to categorise people’s reflexivity practices, depending on how internal conversations are shared externally in action (Archer, 2007; Caetano, 2015; Chernilo, 2017). Although a person can practice any of the reflexivity modes and in various combinations, people tend to have a dominant mode (Archer, 2003, 2007, 2012). People with communicative reflexivity need to validate their internal conversation with others before acting. In other words, they value their relationships and tend to rely on others’ judgements when deciding how to act. Although they are not passive or habitual in their actions and decisions, communicative reflexives often rely on long-term plans and resist change. Autonomous reflexives, conversely, do not need interaction with or validation from others before acting and tend to focus on planning and relying on their own goals, leading to self-contained internal conversations and goal-based agency (Archer, 2007; Carrigan, 2017). Thus, they rely on their own internal dialogues and, if necessary, prefer expert advice or independent research, such as looking for information online. Similarly, people with meta-reflexivity do not rely on others and focus on their own internal conversations. Meta-reflexives reflect and become critical of their previous internal conversations, rely on their own concerns, and practise more self-monitoring (Archer, 2007). This leads to reflecting about what goals to achieve and how to prioritise them, instead of focussing on how to achieve their goals. Thus, meta-reflexivity is linked to concern-based agency (Carrigan, 2017). Finally, people with a fractured reflexivity have limited or incomplete internal conversations, leading to disorientation rather than allowing them to act with a clear purpose.

Archer (2007) hints towards the potential impact of technology, and the possibility of autonomous reflexivity becoming the dominant mode in this context. By looking at Archer’s work in the context of technology in organisations, Mutch (2010b) refers to reflexivity as the particular conceptions of technology that are shaped by engagement with technology and particular contexts. In terms of data-intensive technologies in organisations, Mutch (2010a) discusses how new applications disrupt existing patterns and force employees to become aware of the data, eventually leading to people practising autonomous reflexivity as a dominant mode. In this way, structures influence action not only through reflexivity but also through technology, meaning that structure and action are mediated by both reflexivity and technology and that they also influence each other.

Reflexivity in critical realism and other theories

Critical realism offers a way to rethink the relationship between agency and structure, all of which are central concepts in classic sociological work. Multiple scholars, such as Giddens, Beck, and Bourdieu, have addressed the structure-agency discussion and explained how and why people act as they do by conceptualising reflexivity from their perspectives. Giddens, for instance, focuses on the agent as the object of reflexivity, and how agents revise their self-narratives and the stories of who they are (Bagguley, 2003). Both Giddens and Beck note that, in modern societies, traditions and routines are replaced by reflexivity due to the increase in choice, development of identity, and individualisation of the social (Haugseth and Smeplass, 2023). Bourdieu, by contrast, focuses on habits and routines to conceptualise reflexivity and argues that human action can be coherent without conscious deliberation (Elder-Vass, 2007).

These views on reflexivity have some limitations, as they overfocus on one pole of the agency-structure binary, conflate agency and structure, or fail to consider variations in reflexivity and their relations to concerns and goals (Archer, 2007; Bagguley, 2003; Carrigan, 2017; Elder-Vass, 2007; King, 2010). Bagguley (2003) argues that Giddens’ view on reflexivity is somewhat incompatible with his views on agency based on the duality of structure he theorises in his work. Further, although some of these views complement each other, they are based on a reflexive versus non-reflexive duality and apply reflexivity in a homogeneous way throughout society (Carrigan, 2017; Chernilo, 2017; Haugseth and Smeplass, 2023). Archer (2007), however, argues that there are multiple ways to be reflexive, as seen in the different reflexivity modes outlined above. Thus, although action can be part of a routine or habit or linked to a social group, Archer takes a more active view of reflexivity. Finally, it is also relevant to note that while these theories (including Archer’s) include emotions to different extents, none of them consider them an important element of reflexivity (Burkitt, 2012). Thus far, reflexivity has often been theorised as a rational and individualistic concept.

Technology, awareness, knowledge, and reflexivity

As part of datafication processes, structures, users, and technology shape each other and are part of a situated process that requires paying attention to contextual conditions and focussing on different levels of effects (Siles et al., 2023). This highlights the loops that are formed through the dependence on and circulation of data between media institutions, their audiences, and the digital platforms used and how they continuously influence and domesticate each other’s practices and understandings (Mathieu and Pruulmann Vengerfeldt, 2020; Siles et al., 2023). Discussions on how users and algorithms influence each other are inevitably related to how users think about and give meaning to technologies and practices. As algorithmic platforms become more common, algorithmic technologies gain more power to perform everyday tasks, but users also gain more power and resources to interpret and reflect on these technologies (Willson, 2017). Thus, a reflexivity process takes centre stage. This process influences and is influenced by a user’s contexts, experiences, and understandings, such as a person’s algorithmic knowledge, imaginaries, and engagements with the platform.

Users gain algorithmic knowledge and develop strategies to engage with the algorithms (Bucher, 2017; Lomborg and Kapsch, 2020). Still, due to the algorithmic system’s dynamic nature and its tendency to keep the technical aspects hidden, it is not possible to precisely state what there is to know about algorithms. However, algorithms can still be known and understood in different ways (Bucher, 2018). Seaver (2019) suggests that, in a simple way, knowing algorithms is a matter of revealing them. However, as he then explains, knowing algorithms in this way is a partial, temporary, and unpredictable understanding of them. Although there is no established definition or measurement of algorithmic knowledge, previous work suggests that there are different levels of algorithmic knowledge centred around the relationship between simple awareness and a more complex understanding of what is going on (Cotter and Reisdorf, 2020; Hargittai et al., 2020). That is, awareness as the first step towards more complex forms of knowledge. Therefore, algorithmic knowledge can be defined as the awareness and understanding that there is a system that collects and processes one’s information and behaviour to automatically tailor and recommend personalised content online, in this case, music. Still, potential actions do not emerge until this knowledge becomes part of a reflexivity process.

In line with Archer’s approach (1995; see also Mutch, 2010b), where structures and technology influence reflexivity and action, previous research tends to posit that algorithmic knowledge can be obtained outside the platform through education, the media, or talking to other people but that it is more commonly obtained through experience within the platform (Bishop, 2019; Cotter and Reisdorf, 2020; De Vito et al., 2018; Lomborg and Kapsch, 2020). In this way, algorithmic knowledge can be seen as a more technical or traditional form of knowledge that people might learn or read about, while experience-based knowledge is often labelled algorithmic imaginaries (Bucher, 2018) or folk theories (De Vito et al., 2018; Siles et al., 2020). Both imaginaries and folk theories are considered ways of thinking about algorithms that explain what they are, how to use them, and what effects they might have. In general, they explain something that is usually abstract and unknown. Although more technical algorithmic knowledge might influence a person’s algorithmic imaginary or folk theory, it is not necessary to have technical knowledge of algorithms to think about them. In the case of music platforms, algorithms are perceived in terms of the quality of their recommendations and how intense, broken, useful, or magical these recommendations are perceived (Colbjørnsen, 2018; Siles et al., 2020).

Archer argues that relationships with objects and non-human creatures shape who we become, and these relationships, as well as practical knowledge, engagement with the world, and human abilities and skills, work through reflexivity (Archer, 2000, 2007; Chernilo, 2017). Considering reflexivity as the way we decide to interact and relate with objects, it is possible to consider how users interact with technology through this lens: people reflect, contemplate, plan, and decide how to interact with technology, based on, among other things, their knowledge. Users interpret the material features of a technology and the experiences they have had with similar technologies in the past (Mutch, 2010b). These interpretations, such as algorithmic imaginaries, happen through an internal conversation that all social agents have, where they reflexively deliberate their beliefs, goals, attitudes, concerns, and social environments in general (Archer, 2003). As people have life experiences, or, for example, use technology, they learn from these experiences and develop awareness and then, knowledge. This (imperfect) knowledge will also influence a person’s actions over time (Archer, 2007; Wilson, 2006). Overall, I argue that algorithmic knowledge, both in the sense of technical understanding or experiential imaginings, is a central element in the reflexive process that mediates how social contexts, technology, and agency influence each other. In other words, users become aware of algorithms and develop algorithmic knowledge, that becomes part of their reflexive process. Thus, through reflexivity, users’ understanding of these technologies allows them to assign meaning to them, engage with them, delegate tasks to them, and interpret how they affect their everyday lives.

Methods

Interviews with Spotify users were conducted in 2022 and 2023 to explore their perceptions and uses of music recommendations and algorithms. The sample was formed by 20 Spotify users between 21 and 65 years old (34 average; 8 men, 12 women; none of the participants identified as non-binary or trans). Participants have 5 different nationalities (Norwegian, Costa Rican, Chinese, French, and German) and live in Norway (17) and Costa Rica (3). The participants shared an interest in music and had comparable experiences with Spotify, regardless of the country which they came from or lived. The sample is highly educated, ranging from currently enrolled bachelor’s students to PhD graduates studying or working in a diverse range of fields, including tourism, gymnastics, urbanism, video game development, healthcare, and academic research.

The participants were recruited through snowball sampling by asking colleagues, friends, and family to invite and suggest potential participants. Invitations were sent through social media posts, recruitment posters, and flyers at a university campus, a public library, and record stores. The interviews lasted 44 min on average and followed a non-evaluative conversational approach (Smith and Elger, 2014). In the interviews, participants were asked to describe their use and engagement with Spotify, their music taste, and what they know about the recommendation system. Participants were not directly asked to describe their internal conversations as Archer (e.g. 2003) did in her studies, but the interviews included think-aloud elements to explore similar thought processes and mental activities (Siles et al., 2020; Swart, 2021). In the think-aloud component, participants were asked to open their Spotify account to show how they usually used the platform and comment on the recommendations they received, how they interpreted them, and how they engaged with them. Questions about their perceptions and engagement were more specific to the content they received in their Spotify. This was useful for seeing how they used the platform and the recommendations in a live context. Although reflexivity at that moment is not the same as when they are on their own, this offers a picture of how they think with and about the platform and the recommendations.

The interviews were transcribed, anonymised, and analysed. Participants are referred to in this article using pseudonyms. The analysis was done in NVivo and followed a combination of critical realist thematic analysis and other categorisation and connecting strategies to identify patterns and reflexivity modes (Fryer, 2022; Maxwell, 2012, 2013; Rubin and Rubin, 2005). The interviews were initially coded using a data-led approach, in which codes came from the data and were not pre-established categories. These initial codes were then grouped, recoded, or edited as necessary, based on Archer’s main reflexivity modes, their link to knowledge, agency, and reflexivity and how reflexivity is used in digital contexts.

Analysis

Reflexivity, from Archer’s perspective, is seen as a mental capacity and an internal conversation that each person has in their minds before acting. This refers to mulling over things, planning, deciding, and having conversations with yourself, among other behaviours, in relation to particular contexts and previous experiences. While participants were not asked to describe their internal conversations, all participants talked about some form of reflexive process regarding their use of Spotify and the music recommendations. As Archer (2007) theorises, individuals reflect on the elements they encounter before acting based on their concerns and contexts. Thus, I argue that in addition to their social environments and personal goals and motivations, a person’s algorithmic knowledge and imaginary also becomes part of their reflexive process.

In the first part of the analysis, the algorithmic knowledge and imaginaries of the participants are described and analysed to explore the relationship between them. The second part focuses on the reflexivity modes users practice to identify how they think about music and algorithms. This follows Archer’s (2007, 2012) reflexivity modes, which are also adapted by Carrigan (2017): autonomous, meta-reflexive, communicative, and fractured. In addition to the relation of these reflexive modes to the participant’s use of algorithmic recommendations and music listening habits and contexts, the relationship between reflexivity and algorithmic knowledge is discussed. It should be noted that individuals are always reflexive, and although an individual might have a dominant reflexivity mode, the same person may also practise other modes (Archer, 2007). The goal here is not to categorise each participant into a reflexivity mode but to understand which reflexivity modes are developed in digital contexts and on different occasions.

Knowing algorithms

All the participants are aware of algorithms but show different levels of algorithmic knowledge. Considering algorithmic knowledge as a more technical understanding of the platform, their algorithmic knowledge is closer to an imaginary or folk theory because it tends to be based on experiences. Regarding whether they know what an algorithm is, the participants typically refer to what they think the algorithms do. They explain that an algorithm is what collects information and data, such as what one listens to or skips, and gives recommendations based on that. For example, Olivia explains:

I’m not sure what it is, but I have heard the word. I think it’s like… there aren’t people sitting there and choosing for me what I should like, of course, there’s an algorithm, maybe, that selects different tags and stuff to their songs and puts them together.

Definitions and explanations of Spotify’s algorithmic system are often subjective, even for participants who work with or study technology-related fields and give definitions using technical language. Still, they all explain that they can only guess how Spotify works, as they have only worked or studied with similar recommendation systems and have never seen Spotify’s code. These explanations are based both on technical knowledge and the users’ experiences, because unless they have access to Spotify’s code, they can only explain how they notice recommendations similar to what they listen to or how they are surprised by certain recommendations that do not fit their taste. For instance, Noah explains, ‘I can imagine [how Spotify works]. I work with machine learning, and there are thousands of ways they could do it. I would guess that they have some features for each song’.

Knowledge and awareness are related to how people engage with music platforms and recommendations. For most participants, ‘just using’ the platform is the main source of algorithmic knowledge and awareness, and in response to whether knowing more about algorithms may change the way they use Spotify, most say it would not. They explain that they already get the music they like, and knowing more about how the recommendations work would not change that. Sofie, for example, explains that she thinks that knowing more about Spotify would change the way she uses it, but she is not interested in making a change because she already gets what she needs from the platform. Thus, participants judge their satisfaction with and knowledge about Spotify’s recommendation system when deciding how to act about it. She explains:

I guess it would help me, and I would know how to influence, how to change, and how to use it more productively, but I’m not sure I would put a lot of time into learning it. […] I mean, I don’t really, fully, care about how Spotify works. I think it would help me, but I think I would be too lazy, too. […] I mean, I don’t feel the need to put some time and knowledge into improving [Spotify]. I am satisfied with my Spotify.

This link between Sofie’s interests, algorithmic knowledge, and how she uses the platform shows the existence of a broader reflexivity process mediating the platform and the actions taken by the user. Archer (1995, 2003, 2007) argues that social and cultural structures and contexts influence agency. This influence depends on people’s interpretations of their contexts. Mutch (2010b) adds that these structures are inscribed in technology, which is also reflexively interpreted and used. Hence, participants such as Sofie in the example above reflexively engage with the platform and its recommendations. These reflexive practices depend on a person’s motivations, interests, goals, and previous experiences, which lead to interpretations of the recommendations and eventual decisions on what song or playlist to play. In addition to their interests and motivations, in this case, their algorithmic knowledge has a potential influence on the reflexive process. As seen in Sofie’s case, she acknowledges that knowing more would impact how she uses the recommendations, but that she does not need to know more. Thus, users reflexively use the platforms and recommendations, regardless of their algorithmic knowledge or imaginary. Nevertheless, their understanding of the algorithm, regardless of the level, might impact how they engage with these technologies.

The relationship between algorithmic knowledge and reflexivity can also be related to a person’s education and the social origins of reflexivity. Although Archer does not focus on the origins of reflexivity but rather on its consequences for social mobility, other authors have shown how different social backgrounds might influence reflexivity modes, concluding that people with lower educational backgrounds tend to have a more fractured reflexivity, while people with higher educational backgrounds tend to practice more autonomous and meta-reflexivity (Caetano, 2015; Golob and Makarovič, 2019). Similarly, research shows how demographic backgrounds are associated with algorithmic knowledge, specifically how higher education and higher algorithm awareness levels are related (Gran et al., 2020).

In the current study, participants have multiple nationalities but still share similar interests and experiences when using Spotify. To some extent, their context influences their choice of music; for example, some listen to local music to experience the country in which they live or to certain music due to nostalgia or homesickness. However, the focus here is not on what music they play but on how they interpret recommendations and algorithms. The participants show high levels of awareness and understanding of Spotify’s algorithmic system. This is not entirely a surprise, considering that they are highly educated. This suggests that they are more likely to be more critical and practice more autonomous and meta-reflexivity modes (Archer, 2007; Golob and Makarovič, 2019). Notably, as will be discussed next, the participants practise all modes of reflexivity.

How users think about music and algorithms

As described earlier, all the participants seem to practice all reflexivity modes, even though they tend to have a dominant mode. The aim here is not to identify each participant’s dominant mode but to explore how each mode develops in a music streaming context. Most participants show a degree of autonomous reflexivity. This reflexivity mode refers to decision-making processes that do not require validation from others before acting (Archer, 2007). Thus, a person might prefer a song or agree with a recommendation at a particular time, because it will work better for a particular goal.

This is the case, for example, among multiple participants who explain that they often think about and are very careful about what music they listen to depending on their mood, context, or if someone else will be around. Noah explains that he likes to use playlists, such as Discover Weekly, when he wants to find new music. However, in other contexts, they might decide to use their archived music. For example, Sofie carefully selects songs that she knows will motivate her in the morning or at work, while Noah mentions that when he is feeling sad, he might play happier music to liven the mood. This relates to the role affects and emotions also play in reflexivity and the internal conversation leading to the use or selection of a song recommendation. The reflexive process in which they autonomously consider these different elements in light of their goals and motivation is highlighted in Sofie’s quote below. She indicates playing specific songs to manage her mood in different situations, for example, to motivate herself in the morning, at work, or when she is facing her fears, overall claiming responsibility for playing the right song at the right time:

It changes all the time. Sometimes I want to listen to a music, but I force myself not to because I know it’s gonna put me in a mood that is not productive or that is not what I need at the moment. So, sometimes, I’m like, ‘OK, you want that, but it’s not good for you right now, so don’t listen to it’. So, I don’t just listen to feeling; I also try to think rationally and ask what you need right now to be productive and to do what you want to do? […] For me, it’s really important to choose the right music. […] I feel I have a responsibility during the day to have the good music at the good moment, but it definitely changes.

Listening to music to regulate a particular mood or activity is a key aspect of listening to music in general and in the context of Spotify (Hesmondhalgh, 2013; Sachs et al., 2015; Siles et al., 2019). Although previous studies have not focused explicitly on reflexivity, they provide some hints on the role of emotion in internal conversations and reflexivity’s position between context and action. Hesmondhalgh (2013) shows how people choose music in specific contexts for private and personal use. More recently, Siles et al. (2019, 2020) focus on how users think about algorithms and the playlists they use. They explain how users consciously reflect on the names of playlists, as well as on how algorithms work and how they can be used. Participants practicing autonomous reflexivity do not seem to be critical of algorithms or the use of their data. As their goal-oriented agency suggests, participants explain that they want to find the right music for the right time. While they usually have an artist or playlist in mind, this can be algorithmically recommended. Still, they prefer a playlist they know they can trust and do not leave their music to chance.

In line with this, algorithmic knowledge plays a smaller role than the actual recommendations. Among participants who show that they practice autonomous reflexivity, Sofie can be considered to have a more basic understanding of Spotify but is not interested in learning how it works because she already gets what she needs from it, while Noah has studied and worked with information technologies like algorithms but does not think this influences how he uses the platform. However, their decisions are goal-oriented and do not necessarily need to consider how everything in the platform works. They are more concerned about the music to which they will listen. Algorithmic knowledge and imaginaries take more relevance in the next reflexivity modes.

Meta-reflexivity, the second mode identified here, is closely related to the first mode, as they both lead to using the streaming platform in an active way. However, participants practicing autonomous reflexivity tend to listen to the music they know. Those practising meta-reflexivity try to explore and expand their collection and are more critical of the algorithmic recommendations. Meta-reflexivity leads to people being more critical about their own internal conversation (Archer, 2007) and, in this case, more critical about their music tastes, listening behaviours and the effect they may have on music recommendations. For instance, Sofie explains how she sometimes forces herself to try new music. Ingrid, who hobbies as a DJ, explains how she feels about Spotify’s recommendations:

It’s like gold digging or… crate digging. I’m quite curious, and that’s a way I like to use Spotify. Because they recommend playlists for you depending on what you’re listening to, and that’s the way I find a lot of new music that I love as well. And it doesn’t need to be new music from our year. So, I love using Spotify that way as well to crate dig, you know.

The self-criticism shown by those practising meta-reflexivity is also transferred to how they perceive algorithms and data. Although they sometimes accept and use the recommendations, they also argue that the algorithmic recommendations trap them into a bubble. Jakob, Nora, and William, among other participants, explain that they do not like algorithms because they just get the same songs that they already like. William explains:

I don’t think that Spotify knows me very well. They’re trying to guess who I am. So, the Release Radar and Discover Weekly don’t really work because they don’t know what I know from before, really. And even though I’ve put so much information into the system.

Meta-reflexivity, in this case, leads to being more critical of how Spotify works. Regardless of their technical knowledge, participants practising this mode have a clear idea of how Spotify’s algorithmic system should work for them, and they notice when it does not work this way. Thus, meta-reflexives rely, to some extent, on their previous algorithmic imaginaries and what they expect from the recommendations. This can be seen, for instance, in William’s quote above, and he later explains, ‘For me, algorithms are supposed to help me, care for me, know me. That’s what they’re supposed to do [but they don’t]’. Thus, meta-reflexivity leads to a more negative view of the recommendations, seen as repetitive or not fitting as well as they should.

Although music is social and shared, it is also very personal and often used alone (Hesmondhalgh, 2013), for example, when working or on public transport. Hence, it is not surprising that autonomous reflexivity is the most common type of reflexivity. However, communicative reflexivity can also be identified in some aspects of the participants’ use of music and recommendations. This reflexivity mode refers to the need to validate or confirm an internal conversation before acting and involves trust in and respect for those who are consulted (Archer, 2007). For example, several participants, such as William and Aurora, explain that they consider their friends when making a playlist or ask friends and family if a playlist they made is working, suggesting that they seek certain recommendations and validation from others. Thus, relying on people makes their algorithmic knowledge or imaginaries less relevant, and this reflexivity mode becomes more evident outside the platform. Nevertheless, communicative reflexives can also rely on algorithms to validate their internal conversations instead of other people, as seen, to some extent, in other studies that show how the algorithm is also personified (Siles et al., 2020). Similarly, participants like Ingrid and William address how algorithmic recommendations understand and help them, and even refer to Spotify as ‘a good friend’.

Finally, individuals who practice fractured reflexivity cannot develop reflexive actions due to their circumstances. This reflexivity mode, as well as, to some extent, meta-reflexivity, can lead to stress, indecision, and disorientation (Archer, 2007; Caetano, 2015; Golob and Makarovič, 2019). These feelings can also be identified when participants decide how to interact with the algorithmic system and its data collection. All Spotify users accept giving up their data when they sign up; however, some are more aware of or concerned about what this might entail. Filip, for example, is very aware of the existence of algorithms and explains that he knows there is nothing he can do about them and prefers to just give in:

I’m no expert in this field, but I know algorithms affect our lives, and I know I am feeding the algorithm, and for me, it’s inevitable. So, I just accept everything they ask me. I’m totally a give-upper, like, I give up my rights, if you like to say so… because I know it’s impossible to resist the algorithm, and I just think that… just take my data. And regarding Spotify, as long as you give me the music, I don’t really care. I mean, the algorithm is everywhere… you open a web page, the pop-up advertisements, and YouTube; it’s heavily algorithmed. So, to some extent, I like this, and I think it’s meeting my demands.

Although it is expected that users with fractured reflexivity will tend to agree with the algorithmic recommendations, like Filip, none of the participants are completely fractured, at least in this context, possibly due to the high level of education in the sample. When they have feelings of indecision and confusion, participants tend to transition to a more active use of the platform and a different reflexivity mode. Participants such as Nora and Filip show signs of fracturing when they explain that they often do not know what to play or are bored of their own playlists, so they put on a recommended playlist because they want something new or because it will be music they know they like but in a different order. Thus, they know they can find music similar to what they usually like, although not the same. Instead of passively accepting the recommendations, the recommendations are accepted in an active way, closer to autonomous and meta-reflexive modes, as Ella explains:

Sometimes, I know what to play, and I have a feeling of what I want to listen to, but sometimes I don’t. Sometimes, I’m sick of all my songs, and that’s usually when I open the radio function. And then, maybe, I get suggestions for new songs. Sometimes, I go to my old playlists, and I listen to old songs. That’s when I’m sick of it. And I often, sometimes, use Spotify playlists because they make personal ones for me. Sometimes, I use Daily Mixes and stuff. I feel that everything they make is tailored to the person. Like, I have a workout; it’s called… what’s it called? Workout Pop Mix by Spotify, and I feel like they made that for me. They wrote it, actually, that it’s picked just for me. So, that’s like a mix of songs, but I wouldn’t necessarily put that. But they know what I listen to, so they put it in. So, I use some of them, like the ones that Spotify has made, and sometimes, when I’m at a party, we just search for party music or something like that, because yeah, something easier just to use Spotify’s playlist.

Thus, instead of completely fracturing, Ella transitions to other reflexivity modes, depending on her circumstances. She practices autonomous and meta-reflexivity, considers her goals, concerns, and algorithmic knowledge, and leans towards communicative reflexivity in social circumstances if necessary, indicating how all modes of reflexivity can work together.

Concluding remarks

In this study, I applied a critical realist perspective to explore and identify how different reflexivity modes are practiced by users of music streaming platforms to further existing work on how people understand and use algorithmic recommendation systems. Participants can practice any reflexivity mode, but an autonomous and goal-oriented mode is the most common. The reflexivity mode they practice and how they ultimately engage with the platform and the music recommendations vary depending on each individual’s direct environment, algorithmic knowledge, goals, and concerns. While autonomous and meta-reflexive modes tend to be directed towards private uses of music, communicative and fractured modes rely more on social contexts. Nevertheless, participants make active and reflexive decisions about what music they want to listen to.

This is a step towards conceptualising reflexivity as the process related to how people think about data and algorithms and how they act based on this process. Contrary to Giddens’ structuration theory, the perspective presented here, based on Archer’s work, allows a separation of structure and agency and considers reflexivity as a process beyond self-monitoring (Bagguley, 2003; Burkitt, 2012). In other words, both structure and agency have their own causal powers, and reflexivity is the mechanism that mediates between them. In this way, reflexivity is more than just knowledge or perception; it is the process in which a person expresses their agency and produces social practices. In the findings discussed above, participants do not just think about recommendations but consider their meaning and usefulness, among other things, depending on their current needs, in order to engage with the recommendations. Furthermore, considering reflexivity as the capacity people have to mediate the conditioning they receive from social and cultural structures, media and technology become part of, and influence, reflexivity, just as reflexivity influences media practices.

Reflexivity is informed by the social and cultural structures people receive through new technologies (Mutch, 2010b). Technologies and their affordances influence users’ engagement with the platform through which users gain practical and technical knowledge (De Vito et al., 2018; Lomborg and Kapsch, 2020). Just as users can develop algorithmic imaginaries without technical knowledge, people can reflexively engage with algorithms without understanding them. However, their experiences and knowledge with and of them influence the reflexive process that leads to their interpretation and use. Here, past experiences and the experiential knowledge developed from them are part of reflexivity. As explored in the interviews, knowledge informs conscious action and decision-making when people reflexively engage with data and platforms. Thus, how a person practices a reflexivity mode and decides to act depends to some extent on how their algorithmic knowledge and imaginary has developed.

This article continues to bring forward a critical realist view of reflexivity in media and algorithm studies and a view of reflexivity that considers different reflexivity modes. However, I employ a simplified view of Archer’s approach and reflexivity modes as an introduction of this perspective to media and algorithm research. Certain elements have been omitted, such as the temporality of the process. Archer (1995, 2003) argues that structures are reproduced or changed over time and that reflexivity is a mechanism that changes and adapts. Researching such temporal elements is important to explore how algorithmic awareness and knowledge develop over time, especially considering the malleable nature of algorithm-based technologies that adapt their recommendations as they are used and updated.

Furthermore, emotion is part of imagining and using algorithmic recommendations (Bucher, 2017; Siles et al., 2020). Although I have addressed a few cases in which emotion plays a role in the reflexivity process, my findings have issues similar to those highlighted by Burkitt (2012), who argues that theories of reflexivity fail to put emotion at the centre of the process and come across as rational processes. In my findings, as in Archer’s work, emotions play a smaller role, as they influence motivations and concerns rather than directly affecting reflexivity. I agree that more needs to be done to reconsider the emotional and nonrational aspects of reflexivity in addition to its rational side to comprehensively grasp how people use and understand algorithmic recommendation systems. Thus, reflexivity becomes a process that encompasses both technical and affective elements influencing a person’s agency, which are often, to some extent, separated in other concepts.