Impacts of AI on Music Consumption and Fairness

Abstract

The algorithmic systems utilized by music streaming services have the potential to positively influence individual choices by promoting new artists, but they are also often accused of perpetuating biases. In this research note, we aim to explore the impact of these platforms’ AI-based algorithms on fairness in music consumption. To address this question, we adopt a multidimensional approach that considers the legal, economic, and algorithmic dimensions of fairness. This approach is applied to our EU Horizon Europe Fair MusE project, which advocates for a fairer music ecosystem. However, it should be noted that we propose a tool to score output (playlists) based on fairness models instead of directly altering the algorithms. Data from end users, data brokers, and open-source databases will inform the model, while the processing of the data is aimed at providing users with insights into algorithmic biases and empower them to influence the output. Acknowledging this aspect, this research note serves as a prelude to highlight the need for increased transparency and explainability of algorithms. Furthermore, we seek to inform policy interventions that promote fairness, particularly regarding data sharing between creators and platform providers. Such interventions would foster trust among stakeholders and benefit both users and businesses.

The digitalization of the music ecosystem and the need for fairness

With the increasing digitalization of our world, media sectors are subject to important transformations. Among the first that was influenced by this escalation is the music sector (Arditi, 2014), which consists of several industries. This research note will focus more on the phonographic or recorded music industry, as its link with computing has quite the history. For instance, Alan Turing used a synthesizer already in the 50 s to create and record three melodies: “God Save the King,” “Baa, Baa Black Sheep,” and “In the Mood,” while technology itself has remained a consistent factor in reshaping the musical landscape. Forte (1967) illustrates this situation with the monochord, a device used for musical and scientific purposes, while technical progress not only transformed the creative process for musicians, but also dramatically shifted power dynamics within the industry. With the COVID-19 crisis, the significance of the link between music and computing has been reinforced (Mazziotti, 2020). The popularity of music streaming services (MSS) has skyrocketed, with 67% of the revenue growth generated by streaming (IFPI, 2023), making MSS the gateway to accessing music. This process of plateformization of cultural artifacts includes the key processes of selection, curation, datafication and commodification (Van Dijck et al., 2018).

Furthermore, there is a growing interdependence between major record labels and MSS, where services like Spotify have become much more than a distributor of recorded music. To differentiate their offer of an almost unlimited access to catalogs of recorded music, Spotify, for instance, positions itself as a producer of a branded and unique musical experience (Eriksson et al., 2019; Morris, 2020; Morris & Powers, 2015). This positioning is done through the provision of proprietary playlists, which are curated with the influence of machine learning algorithms and the monitoring activities of human editors in their recommender systems (Bonini & Gandini, 2019). These playlists carry a significant influence in promoting promising songs, as well as the discovery of new artists, to a major and wider list of global audiences (Aguiar et al., 2021). However, the ownership stakes that major record labels have on Spotify might provide a more advantageous bargaining power that favors the inclusion of more content from major record labels. These labels have control over the licensing rights of their catalog and Spotify relies on these rights to curate their playlists. To secure the rights, the service is constantly on the negotiating table with the major record labels, such as Universal Music Group, Sony Music Entertainment, and Warner Music Group (Prey et al., 2022).

Yet, there are noticeable competitive discrepancies between the share of content from major record labels and independent ones. According to Prey et al. (2022), over 50% of the Spotify-owned playlists that were available in 2018 include content that is owned by major record labels. Spotify's 2022 financial statement (2022) also shows that 75% of the global volume of streams is recorded by Universal Music Group, Sony Music Entertainment, Warner Music Group. Moreover, major labels own extremely large back catalogs which offers them significant financial advantages. Around 3,000 artists who have been around for a couple of decades with more than 80% of their streams coming from tracks at least 5 years old which gather more than 500,000 monthly listeners—which Spotify (2023) calls heritage artists—generate significant passive income on their catalogs. Such a situation leads to questions about their position in the industry, as well as their influence.

On the one hand, the influence of MSS on individual choices can result in the discoveries of new artists (IFPI, 2023). On the other hand, the model and the systems that they subsequently optimize have been accused of being unfair because they sustain or amplify biases against some categories of creators, performers and/or works. Such biases are often related to their training dataset or the developers’ bias (Abdollahpouri et al., 2020; Maslej et al., 2023). With this perspective in mind, addressing the concerns surrounding algorithms becomes a much more prevalent topic, as their objectives are also many fold and expand beyond consumer retention strategies.

When addressing the concerns, the outcome of fairness stands out as an important discussion in literature (Maslej et al., 2023; Mazziotti & Ranaivoson, 2024). ¹ Particularly, discussions on fair exposure and fair remuneration structures for music creators have yet to generate consensus. While Fair, Reasonable, and Non-Discriminatory (FRAND) principles and other EU legal frameworks have been proposed to cultivate a fairer ecosystem, the discourse surrounding remuneration cannot be isolated from other intertwined factors that contribute to the issue. To this end, we will draw upon three elements from our literature review: legal, economic, and algorithmic perspectives. These elements will facilitate an examination of what fairness entails and seek to answer the following question: How do AI algorithms in streaming services impact fairness in music consumption?

To answer the question, our research note focuses on the machine-learning algorithms that are responsible for mediating recommendations and curating playlists for music consumption. These algorithms have established themselves as pivotal in facilitating access to content that might otherwise be difficult for listeners to navigate. Additionally, at the end of this research note, we present our approach in aiming to reach a fairer music ecosystem.

A cross-domain analysis of fairness

The concept of fairness embodies a multifaceted nature that carries divergent interpretations across varying contextual frameworks. Deciding on what is deemed to be fair is often shaped by individuals’ experiences, perceptions, and the specific context, whether it is within a business, discipline, or sector. In our analysis, we adopt a multi-level approach that considers what fairness entails from the dimension of legal, economic, and algorithmic perspectives.

What does fairness entail from a legal perspective?

While rooted in varied values and principles among societies and cultures, fairness is linked to normative criteria (Dator, 2006). Fairness is also used in the context of pay equality (Adriaans & Targa, 2023), rewards for individual merits (Hogan, 2024) and life satisfaction (Nicolitsas, 2024). It can also be used in the context of promoting equality of opportunities, as well as outcomes, while referring to basic needs, fundamental rights, and welfare systems.

Despite these interpretations, fairness is generally recognized as part of the wider concept of justice (Hogan, 2024; Ruf & Detyniecki, 2021). When looking at the etymological development of the word “fair” in the late Old English of the early 1300s, it came to mean “virtuous/morally good” (Unebe, 2019), a character which according to Oesterle and Sugden (1970, p. 31), embraces “particular virtues such as courage, justice […].” More recently, however, fairness has gained traction against the backdrop of fair and proportionate remuneration and compensation of music creators (Mazziotti & Ranaivoson, 2024). Given its multifaceted application, the absence of a universally embraced definition of fairness is readily apparent.

Although not always explicitly mentioned, fairness is often a principle embedded in various EU legal frameworks such as Regulation (EU) 2019/1150 (Platform to Business Regulation—P2B Regulation), ² Regulation (EU) 2022/2065 (Digital Services Act—DSA), ³ and Regulation (EU) 2022/1925 (Digital Markets Act—DMA). ⁴ Concerning the matters surrounding competition, as well as the more recent antitrust cases brought against some online gatekeeping platforms, the FRAND principles were brought to the fore, especially in cases involving patents, licensing, access requirements to competitors and contractual agreements. More explicitly, fairness is mentioned in Chapter 3 of Directive (EU) 2019/790 ⁵ (Copyright Directive) which is entitled Fair remuneration in exploitation contracts of authors and performers. Regardless of fairness being an explicit regulatory objective or not, fairness is core to the legitimacy and effectiveness of any legal system, and honoring FRAND principles in business practices encourages a competitive behavior and fosters a healthy and competitive market environment to the benefit of the consumers and the economy as a whole.

In the music sector, according to Mazziotti and Ranaivoson (2024) fair and proportionate remuneration of music creators are one of the main issues that have dominated debates surrounding online platforms and EU music governance. However, discussions on remuneration are not a matter that can be discussed in isolation from other elements that are simultaneously part of the problem. One can think of problematic practices such as economic fairness and algorithmic fairness, whose ramifications ripple down to both listeners—whose consumption behavior is being influenced by the output of these algorithms—and music creators. The latter's remuneration depends on their creative work to reach an audience, which is linked to how the algorithm is employed, which playlists feature their creative work and how/if it is recommended among others (Pachali & Datta, 2022; Prey, 2020), but is also linked to the contractual deals and power relations between record labels—who distribute the revenue received from the streaming services—and music creators.

What does fairness entail from an economic perspective?

Since online platforms have become major enablers of music content flow with unparalleled gatekeeping powers (Vlassis, 2021) and the remuneration of creators deeply depends on the ways through which algorithms expose the content (Mazziotti, 2020), the notion of fairness can be associated with value networks. This includes an analysis on key stakeholders (Mazziotti & Ranaivoson, 2024) and comprehending the broader ramifications of music consumption that necessitates a multistakeholder perspective (Porcaro et al., 2021).

As for addressing the impact of AI algorithms in MSSs on fairness in music consumption, it should be noted that conventional studies in economics tend to favor econometric modeling and economic impact studies (see Towse, 2010). However, the economic analysis of issues concerning the music industry should also consider the translation of end user data, especially since it is integrated into the optimization of the streaming services’ machine learning algorithms for music consumption through personalized recommendations and playlists (Bonini & Magaudda, 2024).

On services like Spotify, playlists serve as a standout feature that repackages music in a form of curated collections that are generated through algorithmic systems and/or human–editorial collaboration. They contribute to the cultivation of moods, as users often select playlists based on their current emotional state or desired frame of mind (Prey, 2018). Additionally, in the current state of MSS, the conceptualization of playlists extends to the association with “datafication” (Kitchin, 2014). This concept involves the collection and analysis of vast amounts of user data, such as listening habits, skips, and likes, to personalize music recommendations (Prey, 2020; Tofalvy & Koltai, 2023). Spotify's Discover Weekly and Release Radar are prime examples of this as their curation process utilizes sophisticated machine learning algorithms that analyze audio features and user interactions to predict preferences and curate playlists accordingly (Freeman et al., 2023). However, the integration of human and machine curation is also accounted for, which dubs the system as having an “algo-torial” approach (Bonini & Gandini, 2019). This approach not only personalizes the listening experience for end users through context-aware adjustments that are based on factors like time of day and user activity, but also influences music consumption patterns by encouraging the promotion of certain tracks and artists over others.

Despite acting as powerful intermediaries in influencing music consumption and artist visibility, the promotion strategy of playlists entrenches existing hierarchies within the music industry. The model easily favors major record labels and their recording artists, as major record labels have the competitive advantage to leverage their connections with Spotify curators to secure prominent placements on popular playlists (Aguiar & Waldfogel, 2018; Prey, 2020). Furthermore, this relationship is strategic, since focusing on playlisting can ensure that the major record labels’ artists can reach larger audiences and receive higher royalty payouts (Morris, 2020). The role of playlists then highlights a complex interplay between technology, human curation, and commercial interests in shaping the way music is consumed, as well as valued on streaming services. They have become pivotal in mediating algorithmic success and reinforcing the power dynamics within recorded music.

Within this analytical framework, impacts on fairness can be observed through the restructuring of power dynamics that transcends beyond the traditional gatekeepers. For instance, before the widespread adoption of MSS, music retailers held a greater influence over the way in which music was presented and sold (Du Gay & Negus, 1994), while major record labels wielded a significant control over pricing strategies because of their monopoly power (Bockstedt et al., 2006). However, in the wake of algorithmic playlists representing the most prominent way that music is made “contingent” and “platform-dependent” (Morris et al., 2021), curatorial power emerges as the dominant logic that governs how music is produced, distributed, and consumed (Prey, 2020). Similarly, a unique form of information asymmetries emerges as the transparency surrounding how music reaches users becomes less opaque (O’Dair & Fry, 2020), while strategies that foster these asymmetries could contribute to higher royalty payouts for music creators who implement these strategies (Morris et al., 2021). Thus, from an economic perspective, impacts on fairness can be drawn upon the shifting locus of power, cultivation of information asymmetries, and the distribution of royalty payouts.

What does fairness entail from an algorithmic perspective?

In the context of computer science, the concept of algorithmic fairness has gained significant prominence due to the increasing prevalence of AI misuse incidents and the growing importance of algorithms in daily life (Maslej et al., 2023; O’Neil, 2017). The issue of algorithmic bias and toxicity is frequently linked to the training dataset or developer's bias.

As mentioned, there is a tension between algorithmic fairness and performance. For this reason, Maslej et al. (2023) emphasized the need to find a balance between the two. However, Wang et al. (2022) warned that some algorithms may disadvantage minority groups in order to perform better. This issue has sparked debates within computer science literature, as Ferraro et al. (2021) stated:

There is a multitude of definitions of (algorithmic) fairness. Two common definitions are ‘group’ and ‘individual’ fairness. Individual fairness reflects that similar individuals should be treated similarly. Group fairness ensures that people of a protected group should be treated in the same way as the rest of the population. (p. 2)

However, these approaches do not consider all possible situations an algorithm may encounter, as algorithms aim to maximize the selected variable and address as many situations as possible (Wang et al., 2022). It is also known that algorithms can reproduce bias from their creators (Mitchell et al., 2021). Since access to source code or training data may not be available, it is necessary to develop various approaches to identify algorithmic issues and biases.

Kleinberg et al. (2016) highlight three fundamental conditions that should be satisfied for an algorithm to be fair in a group fairness perspective: demographic parity, equalized odds, and predictive parity. However, several studies posit the “impossibility theorem of fairness” as those three conditions are not compatible from a statistical perspective (Miconi, 2017; Saravanakumar, 2021).

Moreover, to expand on the discussion of algorithmic fairness, Molina and Loiseau (2022) use the concept of intersectional fairness to ensure that no subgroup is unfairly treated by an algorithm. To achieve this, they calculate the unfairness (Danks & London, 2017) of the algorithm for each group of attributes and then approximate fairness at their intersections.

Impacts of AI algorithms in MSSs on fairness in music consumption

Looking at the EU regulatory landscape, including the legislations listed before, it becomes clear that there is no singular or main framework for observing the impact of fairness in the music sector, as is the case for the audiovisual sector overseen by the Audiovisual Media Services Directive (AVMSD). Requirements for increased transparency and explainability of algorithms and recommender systems are starting to be embedded in legislation, (see Article 27 DSA on recommender system transparency) and are applicable to all online platforms that use recommender systems, including those active in the music industry. Yet, with this in mind, comes the European Parliament's (2024) call for EU rules to ensure the music streaming sector is fair and sustainable, and to promote cultural diversity. Included in their proposition is fair pay for authors, visibility of European works, support for musical diversity and transparency of AI tools used to generate songs. Until further updates, the Copyright Directive, the newer DSA, DMA and P2B Regulation, alongside more established legal frameworks like the e-Commerce Directive and Competition law, remain the backbone for ensuring openness and fairness in digital markets and for defending European values in the online space and beyond.

From an economic perspective, however, implications can be drawn by looking into the services’ strategy of competing for the best recommender systems as a starting point. The increased influence of personalized recommendations and their role in creating platform dependency among rights holders signifies a shift in gatekeeping power. No longer does gatekeeping power solely rest with major record labels, as the outcomes of AI algorithms are assuming a more dominant role in governing consumption and are dubbed “the true hit makers” (Prey, 2020). Because of this increased gatekeeping power, end users who rely on personalized playlists can be easily ensnared in a filter bubble that limits their exposure to diverse content for repeated listening.

Furthermore, the limitation that stems from this gatekeeping power is compounded by how the algorithms are developed. The algorithms rely heavily on user feedback and the consequences of a cold-start problem (Schedl et al., 2018). Even on playlists that are monitored by in-house curators, certain items would always end up being more prominently featured through the readjustments of ranks or positions (Aguiar et al., 2021; Bonini & Magaudda, 2024; Prey, 2020). If this dynamic practice persists, it will imperil the visibility of items and perpetuate the occurrence of information asymmetries (O’Dair & Fry, 2020).

In the streaming economy, information asymmetries are intricately linked to concerns surrounding transparency. In particular, transparent reporting stands out in recorded music as many concerns are directed toward the unfair negotiating conditions that would only involve the streaming services and major record labels (De Voldere et al., 2017). Music creators are often excluded from these negotiations, which has led to an ongoing issue of a lack of transparent reporting of how their music is featured in certain playlists and other recommendation features (Siles et al., 2022). Responding to this asymmetry, music creators have sought to highlight their concerns by resorting to optimization strategies.

According to Morris (2020), the optimization of cultural items, which include songs on MSS, entails a strategic preparation and adaptation to suit the demands of specific platforms. Examples of this include shorter songs, cramming attention-grabbing devices into the first thirty seconds to prevent skips before one play count is registered, and hiring companies to inflate the play counts (Hesmondhalgh, 2021; Morris et al., 2021). Strategies that are aimed for playlist inclusions are also considered since playlists serve to contain uncertainty and contingency by making music amenable to mathematical calculation (Eriksson, 2020). An example of this is the practice of tailoring songs to fit the sonic features of the songs that specific services would include in their popular mood-based playlists (Morris, 2020). Regardless of the aim, the implementation of optimization strategies reflects a trend wherein various stakeholders engage in practices that are akin to manipulating digital platforms for economic gains (Morris, 2020), while contributing to the acceleration of more asymmetries.

Consequently, these optimization strategies have the potential to contribute to the inequitable division of royalty payouts. More specifically, a skewed distribution might occur due to the pro-rata model. This model entails a streaming payment system that is based on how many plays a track has in relation to all other tracks that are played simultaneously (Vonderau, 2019). When taking this model into account, smaller record labels and music creators with relatively lower capital for engaging in optimization strategies might experience a shrink in the revenue. Furthermore, through the continuous deployment of machine learning algorithms that rely heavily on user feedback, popularity bias might continue to occur and make visible those music creators with the most listens (Freeman et al., 2023). This would perpetuate the skewed distribution of royalty payouts, where top-tier artists account for the majority of revenue under the pro-rata model (Brown & Holt, 2018).

Certain MSS, however, have endeavored to address this issue by refining their payout models. Most notably, Spotify has implemented a revised approach that requires a minimum threshold of 1,000 streams within the preceding 12 months for a given track to begin generating royalties and be factored into the royalty pool calculation (Stassen, 2024). While this modification aims to combat “artificial streaming” and better distribute disregarded payments that often fail to reach music creators through distributors, its contributing consequence on fair distribution of royalty payouts remains elusive to precise measurement.

As for the impacts on algorithmic fairness, Aguiar et al. (2021) identifies several issues. These include the ranking of songs, ranking of artists, gender disparities, and the influence of factors such as commercial promise and label type on the algorithm's results. These algorithms, particularly their recommendation systems, have a crucial influence because they determine which music is presented to users. When MSS pays royalties to content providers, they can recommend more profitable commissions by steering users towards cheaper content and promoting emerging artists. This type of strategy is common among profit-maximizing digital platforms that tend to favor the most profitable transactions by biasing the order of recommendations.

Furthermore, in research pertaining to the music industry, “biasing” or biases are recognized as a concept that encompasses two categories: popularity and demographic-based biases. The former reinforce the popularity of already popular artists and songs (Kowald et al., 2020), while the latter are based on various identity categories, including ethnicity, gender, class, age, disability, sexuality, and nationality (Schedl et al., 2015). These biases raise questions of (in)justice and inequality because they expand beyond the traditional concept of bias. The least researched type of demographic-based biases is particularly concerning because it suggests that music recommendations may be influenced by the user's identity categories. This raises further issues that impact fairness, equality, and access to diverse music content for individuals from different backgrounds (Hesmondhalgh et al., 2023).

To address the issue, Veale and Binns (2017) propose and discuss three potential approaches. Firstly, trusted third parties could securely store data necessary for identifying instances of unfair discrimination and incorporating fairness constraints into model building while maintaining user privacy. Secondly, collaborative online platforms would enable diverse organizations to share contextual and experiential knowledge that promotes fairness in machine learning systems. Thirdly, unsupervised learning and interpretable algorithms could facilitate the construction of fairness hypotheses for further testing and exploration. Another approach takes form in debiasing methods, which include rebalancing, regularization, counterfactual intervention, and adversarial training (Melchiorre et al., 2021). Rebalancing involves adjusting data or recommendation results to meet specific fairness measures, while regularization applies constraints on the model's parameters to prevent unfair behavior. Counterfactual intervention involves modifying the input features to test their impact on outcomes and promote fairness, while adversarial training involves introducing additional “adversary” data into the training set to improve the model's robustness and reduce its susceptibility to unfair inputs.

Methodological proposal to analyze fairness through data

While methods like adversarial training may serve as a theoretical gateway to improve the impacts of AI algorithms on fairness, we are currently working on addressing the concerns surrounding algorithms in MSS through the Fair MusE project. The main goal of Fair MusE is to promote fairness for music creators and stakeholders, thus leading to a more transparent, competitive, and sustainable music sector in Europe. This requires proposing new criteria, methodologies, and tools to assess this concept, including the provision of a fairness score. Applied to the music ecosystem, this score will be usable directly by users to let them be able to have a quick overview of how fair the playlist they are listening to is. In this approach, we will not try to change the algorithm itself, yet we will score its output (the playlist) based on various variables that are coming from the legal, economic, and technical work done throughout the project. This tool, based on a parametric way, is able to score the output according to the variables defined and the prioritization of the user.

To achieve this purpose, we are modeling fairness and its dimensions. At the same time, we are gathering data from end users (history of listening and qualitative interviews), data brokers (Soundcharts) and open-source databases (like MusicBrainz that is acting as a pivot model to ensure the alignment between tracks and artists). This approach allows for a triangulation of data as we are collecting user's history of listening (including features related to the use of recommendation systems that is, contextual variables), algorithmic and algotorial playlists from MSS, radio stations (80 stations, 10 for each country represented in the consortium) and third-party applications (e.g., Genius, Shazam). By gathering this data, we are currently building a rich dataset with which we will be able to identify patterns in the history of listening of real users, to compare those patterns with their perceptions through interviews and patterns from radio stations.

This approach could lead us to a more comprehensive perspective of fairness for end users of MSS. From a user perspective, we would like to reinforce the user's agency, which can be done by informing the user of their biased music output, and thus ensure that they have a deeper understanding of how the algorithm is biased and what action they can take to influence the output. In addition, by giving access to an operational perspective for policymakers, our proposal could contribute to restoring balance between them and algorithmic platforms (in the same perspective as the Digital Service Act or the Digital Market Act).

Conclusion

In line with current academic and industry discourse, we assume that algorithms implemented by MSS have an impact on music consumption. However, the extent to which they affect fairness and whether this impact on music consumption is fair in and of itself remains subject to debate, compounded by the limited number of empirical observations. In response to this pressing need for clarity, the Fair MusE portal ⁶ is currently accepting donations of streaming consumption data with the aim of collecting such data donations of at least 1,000 real users, on the grounds of their right to access personal data collected and stored by streaming services and social media companies under Article 20 of the GDPR. By analyzing this data, it is anticipated that discernible evidence pertaining to the algorithmic fairness or lack thereof, as well as its impact on music consumption, will be elicited. Furthermore, the EU makes use of a complex and robust regulatory system, but is the system equipped to tackle algorithms and intervene to make the situation “fairer” if necessary? This question remains to be seen and answered in the remainder of this project.

Moreover, the purpose of this research note was not to exhaustively enumerate and dissect every framework that touches upon the concept of fairness, nor to propose a general definition of fairness under current regimes, but to highlight several aspects: (a) the need for increased transparency and explainability of algorithms and recommender systems employed by MSS (b) the need for increased transparency between the creators and exploiters of digital music when it comes to data collection and sharing, and proportionate remuneration among others (c) the fact that, as argued in the Manifesto by Mazziotti and Ranaivoson, the “[l]ack of transparency also prevents the development of policy measures to promote fairness and diversity in a post-COVID-19 context.” Addressing these aspects further through policy and regulatory intervention would not only serve the users of MSS, but will also serve the business users, just as highlighted among others, by the DSA, DMA, and the P2B Regulation. This will ultimately increase trust between the users, business users, MSS, record labels and regulators, which could pave the path for a fairer and more proportionate remuneration, a path which could also positively impact music consumption and diversity.