Fairness as a normative human concept should be defined for each specific domain and context [14, 19]. Fairness research often focuses on sub-groups of users or item providers who are underrepresented or underserved, aiming to identify and promote fairer outcomes [19]. Translating such goals into practice, and accurately measuring them, remains challenging, as demonstrated for dating apps [61]. For MRS, music artists identify popularity, niches, and imbalanced representations as underlying causes of unfairness [16, 24].

Popularity bias arises when already popular items are recommended more often because of their higher presence in past interactions [1, 11, 42]. MRS are known to suffer from such bias as well [44, 45]. The platforms’ responsibility in shaping a fair music ecosystem for artists by integrating more lesser-known items is recognized by both artists [16, 24] and music industry professionals [18]. For end users, popularity bias can lead to homogenization, as RS show a bias towards items preferred by the majority, making it more likely for minority users to receive poor recommendations [42, 68]. Still, algorithmically aligning RS with each user’s content consumption may not sufficiently reduce such bias, as many users inherently prefer popular items [40].

Niche or non-mainstream music or genres are generally less well-presented in MRS datasets and recommendations themselves [41, 42], leading to underserved artists and genres. This further contributes to a lack of item diversity in recommendations, which in turn could harm user satisfaction [10, 53]. As such, algorithmic bias and interaction bias amplify existing imbalances in society [8].

Imbalanced representations of groups in the data may also cause unfairness, such as disparities in gender and country or nationality [23, 47, 51, 57]. These imbalances also come up in interviews with artists, but with varying responses on what would constitute a fair system in these dimensions [16, 24].

Transparency is widely recognized as a foundational and significant principle in the development of ethical AI [36, 65]. In the domain of RS, it is considered vital to fostering user acceptance and trust [50, 59]. Transparency in a system involves explaining how it works^1^ [63], which can increase user confidence and contribute to quicker decision-making [59].

Explanations play an important role in people’s reception of recommendations [4, 28, 63]. Zhao et al. [71] show that natural language explanations for why a particular song is recommended to a user encourage higher engagement. Since explanations directly affect user decisions and interactions on the platform, one needs to recognize that not all users have the same goals [63].

Decision making. RS help with choice overload in decision-making processes by providing personalized suggestions. While much research focuses on algorithms, user traits and decision-making also shape RS effectiveness [12]. For example, personality may influence users’ preferences for diverse recommendations [13, 35]. Another work shows that users often consider the entirety of a playlist when evaluating its attractiveness and diversity [30].

Control for RS users can be increased by allowing them to provide inputs [64] or engage in active decision-making [7]. Such increased control can positively affect the perceived quality of recommendations [31]. Some argue that creating awareness and enabling users to support their values could also improve fairness for item providers [17, 69]. However, human decisions are affected by noise and limited insight into the system [37], and more control may also increase cognitive load [34]. Moreover, in a simulation study, re-ranking methods more effectively improved gender fairness in RS over time than user choice models [22].

Accountability.Almeida et al. [5] show how algorithms influence individuals and shape society as a whole. They argue that governance strategies are needed to make organizations accountable for the consequences of their systems. Leerssen [43] critically explore RS accountability in Europe, and what types of accountability these RS adopt. However, shared accountability through added user control and its effect on RS remains underexplored.

In this paper we investigate user perception of control and accountability in MRS, exploring how fairness-focused descriptions influence decision-making [60].

The study was held online through Microsoft Teams or in person at Utrecht University. Participants gave informed consent before starting. The study involved four steps, outlined in Figure 1^2^ and took 59 minutes on average (Step 1, 2.1, 2.2, 3, and 4, respectively took 2, 26, 5, 20, and 6 minutes on average).

Step 1: Pre-survey. Participants provided their demographics (age, gender, nationality), how much they listen to music per day (in general and attentively), whether they ever released music on a streaming platform, their familiarity with popular and niche music, their affinity to discover new music, and their preferred genres.

Step 2.1: Playlist selection Round 1. During the first round of the playlists selection task, participants were asked to select a playlist from a set of three options for 4 consecutive sets of playlists, while thinking aloud to share their decision-making process. The sets were based on fairness and diversity aspects discussed in previous MRS work (see Section 2.1): artist gender (Set 1), artist nationality (Set 2), song genre (Set 3, also as a proxy for ‘mainstreaminess’ [54]), and song popularity (Set 4). Within each set, 3 playlists were displayed: one that was not diverse, one moderately diverse, one very diverse. Having participants select one playlist among a couple, as opposed to rating separate songs, allowed us to recreate a more realistic music listening scenario in which we could control for overall diversity. The playlists’ order in each set was randomized to prevent middle option bias [58]. Each playlist contained 5 songs. More details on the song selection procedure can be found in Section 3.2. For each song, the title, artist, and album cover were shown. Participants could listen to an audio preview of maximum 30 seconds per song, similar to a study by Fatahi et al. [21].

Step 2.2: Playlist selection Round 2. In the second round of Step 2, participants revisited playlist Sets 1–4—now accompanied by fairness descriptions (see Table 1)—and could confirm or change their initial selections. This round aimed to explore participants’ responses to the fairness characteristics of the playlists, and whether they would want to change their choice after seeing those characteristics. Extending beyond previous work [6, 38], participants were asked to think aloud throughout this process.

Step 3: Semi-structured interview. We conducted a semi-structured interview to explore participants’ perception of the topics and fairness dimensions that came up in Step 2. The interview covered three topics: 1) task and description evaluation inquired about participants’ experience of Step 2, the influence of playlist descriptions on their choices, whether having playlist descriptions from the start would have influenced their choices differently, and whether they would appreciate such descriptions on music streaming platforms. 2) MRS artist fairness inquired about whether participants felt that artists are currently treated unfairly in MRS, and if so, in what ways—whether related to the dimensions identified in Step 2 or any others they considered relevant. 3) MRS end user fairness inquired about participants’ perception of MRS (un)fairness towards end users, whether they see their own music taste represented, and whether MRS help them explore new music.

Step 4: Think-aloud and post-survey. Here, we provided several facts illustrating some current imbalances in the music industry, illustrating that such imbalances indeed exist, for each fairness dimension from Step 2 (e.g.: “The US and Canada combined receive a 40,9% market share of the global music industry revenue. 7,5% of the world’s global population lives on this continent” [32, 67]). Participants were then asked “How important is it that music from [X] is recommended more on music streaming services?” on a 5-point Likert scale, where [X] was replaced with each of the four respective underrepresented categories. Next, participants viewed their answers plotted against answers from professionals in the music industry in [18]. Participants were asked to think aloud throughout. With these insights, we aimed to measure whether participants would see algorithmic interventions (i.e., disparate treatment [19]) in MRS as a viable way to counter current imbalances.

All participants received the same playlists in Step 2, which suits the study aims to investigate the influence of playlist diversity and descriptions in a controlled setting, rather than to model realistic recommendation behavior. To further minimize the influence of song characteristics, we purposefully sampled our songs as follows. All songs were released between 2013 and 2024 and available on Spotify. We chose to select songs categorized on Bandcamp^3^ under ‘pop’ and with English lyrics, due to its wide appeal to general audiences [27]. The only exceptions to this were playlists in Set 3 (Genre), which contained non-pop songs, and Set 2 (Nationality), which contained non-English lyrics.

To minimize the impact of familiarity on participant preference [33], we selected songs from non-Dutch artists that had been played < 100.000 times on Spotify, with no other songs of the artist having over 1 million plays. One exception is the Popularity dimension in Set 4, where the popular songs had > 100 million Spotify plays. We used Spotify play counts to gauge familiarity as Spotify is a prominent platform in The Netherlands^4^. We consulted artist profiles on Wikipedia, Bandcamp, Spotify and YouTube to retrieve artist information on Gender, Nationality and Genre. To ensure that playlists in one Set addressed only one fairness dimension, the other dimensions were carefully mixed^5^.

We analyzed the initial selection for cohesiveness within playlists. Since playlists perceived as too diverse are generally less well-received [53], we made sure five audio features were similar in each Set^6^. Outliers were identified using a k-nearest neighbors approach (k=4, distance-based threshold=2) applied to a UMAP-generated [46] embedding space of these audio features, and subsequently removed and replaced with different songs. The moderately and very diverse playlists in the Genre dimension were excluded from this analysis due to their inherent high variety. After pretesting the study, two songs were replaced due to both participants responding very negatively to them. The final playlists were rechecked to confirm sufficient cohesiveness.

We report demographics, music affinity, playlist choices, and survey responses using descriptive statistics. All recordings were transcribed and coded using NVivo 14, applying thematic analysis [9]. Two coders coded two transcripts together, then coded individually and cross-validated codes afterwards.

Table 1: Descriptions used in Step 2.2, focused on the fairness/diversity aspect of the respective Set. Each Set contained a playlist for the majority group, a moderate mix, and the minority group.

Codes from previous work were used as a starting point (deductive) [16]. Other codes were added or adapted inductively. Thematic analysis revealed five main themes, discussed in Section 4. While we note frequency for each theme, our focus remains primarily on the qualitative insights.

Participants were recruited through convenience sampling until thematic saturation was reached. 10 were Dutch, 6 European, and 2 non-European. 9 participants identified as women and 9 as men. The average age was 30.5 years (Med=29.5, SD=7.71). On average, participants listened to music for 3.1 hours daily, of which 2.9 via a streaming service, and 1.3 attentively. 3 participants indicated having released music on a streaming platform. Our survey also included 3 Likert scale questions and a genre affinity rating:

•Familiarity with popular music/artists: Agree (7), Disagree (4), Neither (6).

•Perception of their own taste as niche: Strongly Agree (1), Agree (4), Neither (5), Disagree (6), Strongly Disagree (1).

•Curiosity for unfamiliar music: Strongly Agree (5), Agree (9), Neither (2), Disagree (2).

•Avg. Genre affinity: Pop (3.6), Jazz (3.5), Rock (3.4), Classical (3.1), Metal (3.0), Folk (2.8), Electronic (2.6), Hip-Hop (2.2).

Table 2: Reasons for Playlist Selection

Here, we report 1) what participants were thinking aloud during Steps 2.1 and 2.2, and 2) what they mentioned in the interviews in Step 3, when asked about what influenced their playlist choices.

In Step 2.1, participants selected one playlist each from four sets of three playlists while thinking aloud. They discussed various aspects of the songs that they paid attention to whilst listening, including melodies (all participants), vibe (16), album covers (6), lyrics (5), and artist names (4). Then, when selecting their favorite playlist, participants mentioned similar factors (i.e., taste, familiarity, vibe, and diversity) to be important in their decision. Table 2 gives an overview of these themes. The left side of Figure 2 shows the choices users made between the three playlist of each set, with most people choosing mixed gender, moderately diverse genres, highly diverse nationalities, and highly popular artists.

Musical taste was the main factor in users’ choices (16 participants) including the playlists’ vibe: P10: "When I’m choosing a playlist, I would probably go for, like, overall vibe." Preferences included genre and style, and occasionally individual songs: P3: "From these two playlists, I like the first two songs a little bit more, because it’s like more electronic dance music, which I do tend to listen to.".

(Un)familiarity with music was frequently noted in Sets 1–3 (with unfamiliar songs by design), by 12 participants. In Set 4 (Song popularity), participants often expressed joy at recognizing songs. In that Set, participants often showed a preference for known songs: P13: "The main reason I chose this playlist is because I recognize the songs." They spent more time on unfamiliar songs than familiar ones, sometimes replaying them: P7: "I don’t know [this song] [...] So then I’m listening longer to it. But, yeah, a lot of these songs I know, so then I don’t have to listen to them very long to remember the song." They often shared opinions about artists familiar to them. Some participants also tried to identify the playlists’ themes unprompted: P5: "This seems like an international playlist, right? [...] All the songs are different languages."

The language of song lyrics was mentioned by 4 participants as an important aspect. For Set 2 (Artist nationality), some expressed a preference or dislike for music in a particular language: P10: "I do like French music a lot". Overall, the playlist containing the most diverse nationalities was selected by the majority of participants, as they found it refreshing to explore songs outside of their usual rotation: P8: "I think I’ll go for this one with all the different languages. I think that would be fun." However, they preferred to listen to diverse playlists when actively exploring new music.

Table 3: Perception of Playlist Descriptions

In Step 2.2, participants were asked to change or confirm their selected playlists, now with each dimension explicitly labeled. This process was much quicker, as most participants (15) felt confident in their initial choices. Some re-listened to certain songs when they were uncertain about their initial selections, but nearly all participants maintained those selections, taking into account several aspects outlined in Table 2. At times, the descriptions made them feel happier with their choices: P15: "Yeah, some things were good to see, that I tend to move towards playlists with diverse genres.". At other times, it led to disappointment: P11: "With the [popularity-based playlists] I was like: “oh, bummer”. [...] I would like to listen to new music, and it’s annoying [my choice is] so mainstream." While the descriptions influenced participants’ feelings about their choices, they did not affect their final selection (see Section 4.3 for details).

The perception of playlist descriptions is summarized in Table 3. 8 participants indicated feeling influenced by the descriptions, though often not for all fairness dimensions, or only for ones where they had already made the choice that corresponded to their values. Conversely, in 16 interviews, participants (also) expressed not being influenced by the description for at least one dimension. Furthermore, some stated that the descriptions did not provide information about the music itself: P2: "Just the description didn’t tell me much about if I was going to like a playlist or not. [...]’Diverse music’ can have literally any music in it." Still, 15 participants did want streaming platforms to include such descriptions more, e.g., to speed up their decision-making process: P7: "[The descriptions] would make it easier for me to choose, because it kind of explains what I’m going to listen to." 7 participants also mentioned ethical implications: P6: "They made me think more ethically about the choices I make. I do think it’s also just important to listen to music you like, you shouldn’t just listen to music purely because it’s from a certain group or not. But [it’s] good to be mindful of it." Lastly, 5 participants expressed that such descriptions would help other users to be more mindful about diversity in their own listening behavior: P11: "I think it would be more eye opening for [some people]."

Showing descriptions before choosing a playlist would supposedly influence some participants to choose a fairer and more diverse option from the start: P1: "I would have had an incentive because of curiosity to look at the diverse nationality playlist". Table 4 shows the influence the participants estimated for each dimension. Finally, some participants only wanted descriptions in certain contexts: P15: "If I want to have a review sometimes of my playlists, how I listen to music, how diverse it is, then those statistics would help. But [...] I wouldn’t want that in my face [laughing] all the time."

In summary, the playlist selection process was driven by diverse considerations outside the fairness dimensions. For some participants, the descriptions were helpful, but not for all participants or for all Sets. These conflicting values will be discussed next.

Table 4: Participant responses on whether a diversity-related description would have influenced their choices in Round 1.

Here, we explore the fairness values that participants expressed in Step 3 and 4, and discuss whether those values were reflected in the choices they made in the playlist selection task (Step 2).

Expressed values were collected when interviewing participants on what they consider to be fair for artists and for users. Table 5 shows the values mentioned. Participants focused on wanting to hear more music from artists of certain underrepresented or niche groups, which they related to artist fairness concerns (calling for better representation) and touches on user fairness issues (e.g., feeling that current recommendations do not meet their needs). One key value expressed in the interviews was discovery: participants sought more recommendations beyond their usual bubble, valuing greater diversity overall. This is further addressed in Section 4.5.

Table 5: Expressed Values of Participants

Supporting lesser-known artists was mentioned by all participants. Some related this to users not being able to find them on streaming platforms easily (see Section 4.5). Others emphasize the artist unfairness perspective, mentioning that either algorithms, or how the music industry as a whole functions, cause unfair advantages for already popular artists: P2: "If it’s too focused on what everyone likes, then it’s very hard for the lesser known people or lesser known genres to become more known. [...] And that’s kind of unfair to the lesser known people, I think." Some did note that recommending more popular artists is understandable from a business perspective: P15: "It’s more related to how [streaming platforms] make the money and business out of the big artists. So from their business perspective, [...] I understand, but it is not fair, obviously."

Artists’ gender was also discussed in all interviews, but participants here were less unanimous on whether they perceived the current platforms as unfair towards artists from underrepresented genders, and on whether action should be taken to address historical gender imbalances. Supporting minority genders was a priority for some: P2: "I am all for having more female voices in music. So especially in, for example, rap is a very male dominated genre." While for others, gender should not be a factor: P4: "What difference should it make? [...] What does the gender have to do with anything?"

Artists’ nationality was mentioned in 16 interviews in a fairness context. 6 participants would like to be able to better discover artists from other countries. Another 7 specifically mentioned that artists outside the US (and sometimes UK and Western Europe) should be promoted: P8: "I think we are doing the world a disservice by only selecting big pop stars from America. [...] [For example], a Dutch artist will only be pushed into the Netherlands and nowhere else. [...] We’re all missing out on a lot of good music because of this policy." Participants recognized that song lyrics affects an artist’s reach, noting that the prevalence of the English language provides an advantage to artists that write their lyrics in English: P11: "I don’t like the fact that [the US and Canada have] such a big share of the market, but they do have the luxury that English is such a commonly spoken language, so people can resonate with it a bit better."

Music genre was also mentioned in a fairness context by 16 participants. Here, the main sentiment was more nuanced, with participants expressing that they did not see it as a problem if music listeners have one preferred genre and only listen to that. They suggested giving the choice for more genre diversity to the user: P4: "If I ask for it, then yes. Otherwise no. [...] I’d assume that the people know what they roughly want to listen to. So if that happens to be pop, yeah." However, they noted that if one wants to discover a new genre, streaming platforms do not sufficiently cater to that need: P2: "I also like it, but then somehow I end up only listening to pop if I only click on [recommended] stuff, because it doesn’t recommend me any other stuff. And that’s kind of unfair to the lesser known people."

Figure 2:

Sankey Diagram of Playlist Selection in Step 2.2 and Survey Responses on Disparate Treatment in Step 4

As previously mentioned, nearly all participants confirmed their initial playlist choices after seeing the descriptions, with only 3 exceptions. 2 changed their choice in the Set 3 (Artist nationality) from ‘US artists’ to ‘highly diverse nationalities’, both indicating that they were triggered to explore new content. The participant stated that their third change was unrelated to the descriptions.

The conflict between values and actions is illustrated in Figure 2, which shows the relation between the playlists selected after seeing the fairness descriptions (on the left), and participants’ Step 4 responses on their level of concern regarding the fairness dimensions (on the right). Interestingly, participants’ playlist choices were often not coherent with their responses in Step 4.

The Popularity dimension shows most notable difference. Here, 11 participants chose the ‘Popular’ playlist, even though they identified the lack of recommendations of lesser-known artists as an issue during Step 3, and in Step 4, agreed to the statement that more lesser-known artists should be promoted on streaming platforms.

For Gender, Figure 2 shows that almost all who (strongly) agreed that gender balance should be improved, still preferred the mixed playlist (songs by men & women) over the playlist with songs primarily by women. Notably, the average response from the 9 male participants (avg = 3.2) was lower than that of the 9 female ones (avg = 4.3), whereas for other fairness dimensions the differences between participant genders were negligible. Regardless of their opinion, most participants chose the mixed playlist.

Lastly, regarding Nationality, 16 participants expressed a preference to promote non-US artists, which is in line with the majority choice for artists from diverse nationalities.

In summary, even though participants indicated to care about one or more fairness dimensions, this often did not influence their final playlist selection. Their decision process was largely driven by other considerations: P6: "I do want more diverse types of playlists. But I still went with the one I liked."

In addition to discovering participants’ specific values (see Section 4.3), in Step 3 (interviews) we asked participants whether they had ever considered the fairness of streaming platforms and MRS towards artists and towards users. If so, they were asked about their perception of what makes such a platform fair, and whether platforms are currently fair or not. As some participants mentioned ideas on how to improve fairness, we discuss those here, too.

Table 6: Fairness for Artists

Fairness for artists was considered by 11 participants before participating; 6 had never considered it, and for 1 it was unclear. 15 participants said they considered the MRS integrated in streaming platforms to currently be unfair. Table 6 shows the frequency of themes coded on this topic. The most frequently mentioned cause of unfairness was popularity bias (10 participants): P4: "Popularity is so bloody influenced by these algorithmic things apparently nowadays, it’s just like: “oh, hey, look, this guy has passed a certain threshold, so we instantly recommend him to many more people.” Like, I don’t know, is that fair?" Secondly, 6 participants noted that artists and labels with good connections and money get an unfair advantage: P1: "How much money you have when you start out, this is also something that’s connected to a lot of social factors, right? And whether you have access to strong support of music labels, or have to do it on your own." Some even suspected that some artists or labels are paying streaming platforms to be recommended more: P3: "I think music labels will probably pay Spotify money in order for them to put certain songs in certain playlists, which feels weird. Like I’m sort of getting tricked to listen to a certain song." The third cause for unfairness, mentioned by 4, was that they felt artists get paid (too) little: P5: "They’re not paid well enough. And that’s also probably because we pay too little for the streaming. And most of the money goes to popular artists or to the company." Only one considered MRS to be fair: P18: "I think unfairness is removed from such a platform because anybody can post anything as long as it’s within the guidelines of the service."

Table 7: Fairness for Users

Fairness towards users of music streaming platforms had never been considered by 4 participants. 5 were dissatisfied with it, while 9 found platforms mostly fair as they meet expectations and offer convenience: P10: "All of my preferences or the types of things that I could think about for playlist, I feel like I get those."

Table 7 shows the frequency of themes coded regarding user fairness. In terms of feeling that their taste was recognized in MRS, 8 participants agreed, 4 were undecided, and 6 disagreed, one mentioning: P7: "I don’t always feel that [the system] knows what I want to listen to at the moment. [...] But I understand [that if] the history is not representative, then it can’t make a good decision for the present."

Further, 6 participants viewed the lack of transparency or control as unfair: P1: "I don’t know why things are recommended to me. I don’t think I have the option to calibrate this a lot." 5 participants saw lack of diversity as unfair: P11: "I’m seeking more variety and diversity and different angles, different songs [...]. It’s difficult to stray away from the music norm or a music genre if you’re just being fed this the whole time." Finally, as with artists, some considered monetary interests: P16: "What I find frustrating is that sometimes I feel like because of a business model [...] certain platforms will sort of try to force-feed a certain type of [profitable] content to you."

Finally, 13 participants expressed being able to find music that resonated with their taste and associated community. Others mentioned being undecided about representation: P1: "[I personally feel represented], because I’m in a relatively privileged position. [...] but I wouldn’t say it’s all good for everybody." while the rest felt they were not well-represented: P4: "I felt caged. But maybe that’s because I did like too much of the same thing. And then it overfits to like: “Oh! More of this stuff, here we go.” [...] It was horrible."

In summary, participants were more aware of artist- than user-related fairness issues, focusing for the latter on limited taste capture and a desire for more discovery. We expand on this next.

Here, we show how participants discover music and feel supported by their streaming platform, as expressed in Step 3 (interviews).

Lack of support in discovering new music was mentioned by all participants: P15: "They don’t encourage you to listen to something different. [...] To explore something new." They reported that streaming platforms perform better in helping them discover music close to their preferences than for music that is further away. Participants did mention that platforms recognizing their usual taste offers convenience, despite the possible drawbacks. However, to find niche music, participants indicated the need for more specific queries: P15: "I think it’s more of our own effort. And then once they recognize that this is what you’re interested in, then they push it in, instead of [they] themselves taking some initiative to push something." They believed that this is caused by RS algorithms, or the workings of the music industry, mentioning: P2: "It knows that you like one thing, it will only recommend you that one thing. [...] it’s difficult to tell the system like: “no, I don’t only want this, I want some of this as well”." They also felt like this narrow focus on their preferences gets amplified over time: P15: "It’s pushing the same thing again and again and now it’s been to the point when I feel like they’re just repeating the same stuff [...]. So that is kind of frustrating"

More control by giving users the ability to set their own selection criteria was seen as ideal, as participants felt that currently the algorithm is setting the criteria for them: P1: "At the very least, the user should have the option of choosing certain criteria that they want to not impact the recommendations [...]. So, not only being transparent [...], but also allowing the user to completely influence them."

In summary, while participants felt streaming platforms help them discover music and recognize their taste, downsides of such platforms are that they lack diversity and control.

Finally, we examine where participants feel the responsibility lies to make music streaming platforms and MRS fairer, using data from Step 3 (interviews) and Step 4 (survey and think-aloud).

The role of streaming platforms in potentially improving fairness was mentioned by 11 participants. 5 saw them as key stakeholders responsible for implementing changes to enhance fairness: P5: "It should be also the job of, like, Spotify, to [...] not amplify popular artists, but to give a chance for smaller artists." 3 did not consider them responsible, and 3 were unsure.

Regarding the role of other stakeholders, some viewed the music industry as responsible, especially for countering inequalities between popular and lesser-known artists: P16: "I think it’s more the music industry as a whole that needs to do something about it. The streaming [platform] is at the very end of the chain." P12 noted that inequalities should be solved at their source, mentioning both algorithms and society as a whole: P12: "Is [gender imbalance] a result of the recommendation system, or of people’s individual biases? [...] If it is the former, then [women] should be recommended more, but if it’s not then I don’t think that is the job of the recommender system to fix, and it’s more something that society should fix." As previously described, participants also often identified financial gains as causes of unfairness, recognizing that the status quo may be more profitable for the platforms. In this case, broader societal tendencies, such as capitalism, were blamed for unfairness and viewed as inevitable: P6: "It feels a little bit silly to say, but I think the capitalism behind it. [...] In the end, they all want to make money."

Suggested improvements to achieve artist or user fairness were shared by 12 participants. Responses mainly concerned changing the recommendation algorithms or increasing transparency into, and control of, recommendations in the user interface. On transparency, P1 mentioned: P1: "I think [it would be fair] to be transparent. Make me understand as best possible why I’m being recommended certain things, [and others] are being hidden from me." More control was considered to increase fairness as well: P11: "[Streaming platforms] want you to stick to the product, and I think if it were to really be fair, they would let that choice be upon you, and they would also design it in such a way." A certain overall fairness preference could be set in a user’s profile: P7: "It would be good if users could indicate if they want to, like, support lesser known artists"

Regarding disparate treatment for underrepresented artists in MRS, participants were divided (see also Figure 2). In Step 4, some further reflected on whether disparate treatment was a solution, sometimes seeing it as potentially viable: P2: "It sounds at least somewhat fair that the [percentages of men and women in recommendations] are similar." In generated playlists, fairness goals could be directly integrated: P17: "If you put on a popular artists radio [...] and you put in lesser known artists, it would be really nice to get to know more artists that have the same kind of genre or something." Two participants were unsure, one noting: P2: "I think it’s strange to say: “Oh, this group is underrepresented, so we should [...] give them more attention.” Because that doesn’t sound like the right reason somehow. But it’s just to kind of make it fair - I don’t know."

In short, there was no consensus on who is responsible to improve fairness, with participants mentioning streaming platforms, the music industry, users themselves, and even society as a whole.