Research and analysis

The impact of recommendation algorithms on the UK's music industry

Published 9 February 2023

1. Executive summary

The growth of music streaming through digital service providers (DSPs) has altered the UK’s music industry and consumer behaviours.

Consumers have a greater abundance of choice of music, and a growing expectation that their experience be more personalised. To address this, the process by which DSPs select and recommend music to consumers has become more automated.

Algorithmic recommendation systems form the basis of automated recommendations to consumers. They assume the role of a cultural intermediary between consumers and music. These systems employ a variety of technologies and techniques depending on the specific ways in which they are deployed. The idea of “a” or “the” recommendation algorithm is oversimplified.

Consumers, creators, and other stakeholders across the music ecosystem are interested and concerned about how these technologies function, and what their impact has been. This report provides a look at consumer, creator, and stakeholder sentiments towards these technologies, and existing academic research about whether these technologies are unfairly biased. This includes the role these technologies play in the wider music ecosystem, and how transparent DSPs are about how they use these technologies.

There are widely held beliefs that the use of these technologies might serve to unfairly advantage certain groups at the expense of others. This report finds that evidence proving or disproving whether these technologies embed, amplify, or introduce unfair biases is mixed, and at times inconclusive. It argues more research would need to be done to understand if and whether these types of impacts exist, and how they might be mitigated. However, this report also finds that collecting the types of data needed for this type of research may be intrusive.

This report goes on to suggest that the potential impacts of these technologies cannot be disentangled from the history and wider context of the UK’s music industry. These technologies should be considered in relation to other forms of cultural intermediaries, which continue to exist and influence the UK’s music industry.

Despite some efforts by DSPS, there is significant demand from creators, their representatives, and consumers for additional transparency about where these technologies are employed, and how they operate. This implies that existing forms of communication between DSPs and these groups have so far proved unsuccessful, even if well intentioned.

This report goes on to suggest additional steps that could be taken by DSPs, and other parties, to better understand the impacts of these technologies, and to resolve desires for greater transparency.

1.1 Key findings

Summary of key findings

1. The belief that recommendation algorithms result in unfair outcomes is widespread. Academic literature about these types of biases tends to be divided into those looking at popularity bias and those looking at biases according to demographic characteristics. There has been more research into popularity bias, and some evidence suggests that issues of popularity bias have reduced in recent years. There is limited evidence proving or disproving the amplification of existing bias, or the introduction of biases felt across different demographics.

2. The impact of recommendation algorithms cannot be considered in isolation from the wider history and condition of the UK’s music industry. These recommendation algorithms should be considered one among several different cultural intermediaries that influence how consumers engage with music. The majority of listens on DSPs remain unguided by recommendation algorithms. Whilst uses of recommendation algorithms bring their own risks and opportunities, these should be considered in the context of other cultural intermediaries.

3. Current approaches by DSPs towards transparency in relation to these technologies are not sufficiently alleviating the concerns of consumers, creators, and other stakeholders across the music industry. Feelings of mistreatment and suspicion, whilst not necessarily justified, are widespread.

2. Introduction

On 9 July 2021, the Digital, Culture, Media and Sport Committee published its Second Report of Session 2021-22, Economics of music streaming (HC 50). This report included a number of recommendations to the government. Recommendation 18 called for “research into the impact of streaming services’ algorithms on music consumption, including where creators are forgoing royalty payments in exchange for algorithmic promotion.”^{[footnote 1]} The government agreed that “there would be value in conducting further research on recommendation algorithms used by streaming services”^{[footnote 2]}, and CDEI was asked to lead on conducting further research in this area. This report draws together the findings of this research.

The introduction of streaming services has had a significant impact on the music industry. These impacts have been extensively explored by the Competition and Market Authority (CMA) in its market study into music and music streaming.^{[footnote 3]} According to the CMA, consumers have embraced music streaming - in 2021 there were 39 million active users of music streaming services, and over 138 billion streams. Streaming has been pivotal in the sector’s recovery from the impacts of piracy, which emerged in the early 2000s.

A British Phonographic Industry Report found that more than 80% of music is listened to via music streaming services.^{[footnote 4]} Our consumer polling provided a slightly conflicting view of the prominence of streaming. According to the consumers surveyed, radio remains the most popular way of listening to music, with nearly two thirds of all respondents (64%) using this method.^{[footnote 5]} Fewer respondents reported listening to music via a free streaming service (40%) or paid-for streaming service (40%).^{[footnote 6]} However, it could remain the case that consumers that listen to music via streaming services account for a greater portion of overall music consumption.

Digital Service Providers (DSPs) provide both free streaming services (with advertising) and paid subscription services. These DSPs use a variety of algorithms in delivering this service.

As explored in the literature review produced by Hesmondhalgh et al. for this project, many kinds of algorithms are used in ways that are relevant to the production and consumption of music.^{[footnote 7]} This report is primarily concerned with how algorithms are used in making recommendations of music to consumers.

Music is recommended on platforms in a variety of ways, some are automated and personalised, and others are not. Our work revealed there is confusion, from both consumers and creators, about the roles that algorithms and human editors play, particularly in the role of playlisting music. This work is concerned with algorithmically generated recommendations, and not playlists or other recommendations curated by human editors. However, this confusion over roles remains important.

The underlying technologies that are used to make automated recommendations are commonly known as recommendation systems. The recommendation systems that power music streaming recommendations are broadly similar to recommendation systems used elsewhere, for example in ordering content on social media platforms. Three approaches used by recommendation systems are:

1. Collaborative filtering - this approach draws on the preferences of other users on music tracks through platform features (such as “liking” a track) to make recommendations,

2. Content-based approaches - these draw on information about the content of the music (such as artist and song title), and feedback provided by experts or audio analysis to make recommendations,

3. Contextual approaches - these approaches draw on contextual factors (such as the device a consumer is using, what day of the week it is, the weather at the time of listening etc.) to make recommendations.

How these approaches are utilised in music recommendation systems specifically is covered in greater depth in our literature review.^{[footnote 8]} As a summary, music recommendation systems typically use a combination of collaborative filtering with either content-based or contextual approaches.

This report explores existing evidence of the impacts that these recommendation systems are having, as well as consumer, creator, and industry views on these technologies. Much of the existing conversation about these technologies, including submissions to the Select Committee inquiry, highlight concerns that they could in some way unfairly disadvantage certain groups of creators.

The rest of this report is structured as follows:

The next section, “Algorithmic bias”, presents allegations of biases that could result from recommendation systems, and evidence of whether these biases exist.

The following section, “The wider UK music industry”, considers the interplay between these technologies and the historic state of the music industry, and contrasts these technologies to other cultural intermediaries that have, and continue to, inform how consumers interact with music.

The following section of this report, “Transparency”, looks at existing levels of transparency about these technologies, and creator and consumer attitudes towards existing transparency.

This report then moves on to suggest further research that could enable a better understanding of if, and where, unfair biases exist, and how the use of these technologies could be better communicated about by DSPs, to creators and consumers, to foster trust in the “Ideas for further action” section.

The final section, “Methodology”, outlines the methods of research that have informed this work.

3. Algorithmic bias

The CDEI has previously explored how algorithmic systems can lead to biased decisions in our Review into bias in algorithmic decision-making.^{[footnote 9]} Whilst this work focuses on the use of algorithmic tools in four specific areas (Recruitment, Financial Services, Policing, and Local Government), many issues identified in this work are also felt in relation to the recommendation systems used in music streaming.

3.1 Perceived bias

Our engagement with creators, consumers, and industry showed significant concern that algorithmic recommendations made on DSPs could be “unfair”. Defining what “fair” or “unfair” recommendations would look like is difficult. For example, an algorithm could be considered “fair” or “unfair” compared to the previous level of fairness/unfairness in the music industry. Alternatively, an algorithm could be considered “fair” or “unfair” outside of this context (i.e. if we rebuilt the music industry to be as “fair” as possible, what would it look like?).

Different groups we engaged with tended to view “fairness” through different lenses. However, through our interviews we identified a number of recurrent perceived types of “unfair” bias. Some of the more widely held beliefs are as follows:

• the relationship between major labels and the DSPs are leveraged to provide greater promotion for the creators represented by these labels

• DSPs prioritise promoting music that they hold rights over, as a way of reducing the remuneration that they have to pay out

• music recommendation algorithms tend to favour already popular creators

• music recommendation algorithms could disadvantage certain demographic groups

• certain genres of music are disadvantaged compared to others by music recommendation algorithms

• introducing the capability for creators to sacrifice earnings for greater algorithmic promotion will create an uneven playing field, or if taken up by a significant number of creators only serve to benefit DSPs

These fears were voiced to us throughout our research. In highlighting them we are not passing judgement on their validity. We explore more below what evidence there is to support these allegations of unfair bias.

Many of the creators that responded to our survey expressed significant concern that bias in recommendation algorithms could lead to unfair prioritisation of certain music. For example, 89.2% of creators that responded to our survey agreed or strongly agreed that they are concerned that bias in recommendation algorithms could lead to certain artists, or labels, being prioritised over others.^{[footnote 10]} Similarly, 85.3% of respondents expressed concern that algorithmic bias could lead to music from certain genres being prioritised over others, while 67.7% of respondents are concerned that recommendation algorithms could lead to certain demographic groups (e.g., different ethnic groups, genders etc.) being prioritised over others.^{[footnote 11]}

Our consumer polling showed that although consumers are less worried about these types of bias than creators, concern remains. In the polling results, a significant minority of respondents (49%) agreed that they are concerned about unfair bias in recommendation algorithms and how these might impact on their own listening habits.^{[footnote 12]} Furthermore, this group was almost evenly divided over whether they agreed or disagreed that they were concerned about unfair bias in recommendation algorithms and how these might impact on artists (46% agree, vs. 45% disagree).^{[footnote 13]}

3.2 Evidence of unfair bias

The concept of bias is perceived inconsistently by different groups. The literature review produced for the CDEI identifies two main categories of bias in existing academic literature: popularity bias (i.e. favouring the most popular items in recommendations) and bias according to demographic characteristics (e.g. users from a particular demographic being recommended certain types of music, or creators from a particular demographic being advantaged or disadvantaged by recommendation systems.)^{[footnote 14]} Notably, these two types of bias can be interdependent - for example, bias towards one demographic group could result in a popularity bias towards this group. These biases could lead to harms, such as disadvantaging creators from certain groups, and limiting the types of music available to consumers.

The literature review explores in detail existing studies that examine the extent to which biases in these two categories exist.

The review finds that there is a wealth of research into popularity bias, and how it can be countered.^{[footnote 15]} Popularity bias is an issue both for more obscure artists, as it entrenches the advantages of successful artists, and the users who listen to more obscure artists, as it makes it harder for them to find artists they want to listen to. Academic researchers recognise the importance of novelty and discovery in recommendation algorithms and have outlined technical interventions to mitigate popularity bias.^{[footnote 16]} It is harder to gauge what DSPs have done to address this.^{[footnote 17]} Our conversations with DSP data scientists suggests they also consider the use of adaptive policies, which vary how wide a range of artists a user is exposed to based on their profile, to tackle popularity bias. Aside from lessening popularity bias, their research shows this can also improve user satisfaction.

Since 2018, there appears to be a slight shift in UK streaming data towards less popular artists and tracks.^{[footnote 18]} This could reflect how diversifying music recommendations engage new users or how DSPs are aware of public concerns around rewarding musicians on streaming platforms. It is worth noting that popularity bias cannot be extracted from its historical presence in the music industry. From the 1950s to 2000s, the music industry placed great emphasis on records being in the chart listing, giving these enormous amounts of media coverage.^{[footnote 19]} If the shift towards less popular artists is accurate, then DSP recommendation algorithms could help counter this historical popularity bias (this issue is covered in more detail below - see ‘Music Prior to Streaming’). However, further research into how music recommendation systems work on actual streaming platforms is needed.

By contrast, the review finds that research into the way that certain demographic characteristics affect the recommendations that certain users receive, and relatedly how artists from different demographic groups are recommended, is much less consistent.

Gender is the most explored category in research on demographic bias amongst music recommendation systems, according to our literature review.^{[footnote 20]} This is partly due to compulsory registration processes on streaming platforms requiring users to identify as “male” or “female”.^{[footnote 21]} Studies suggest that men and women are recommended different artists, indicating the user’s gender factors into recommendations made by algorithms. However, regardless of user gender, music streaming platforms appear to predominantly recommend white male artists to users at a significant rate.^{[footnote 22]} It is worth noting that this is partly due to factors such as the low share of music by female artists, both on platforms and in the industry more widely, and music critics’ recommendations.^{[footnote 23]} Since other studies suggest recommendation systems may increase the number of women’s songs listened to, more research on larger datasets and how recommendation systems affect actual streaming practices is needed.

For other demographics (such as race, ethnicity, social class, sexuality), there is a lack of substantial research evaluating demographic bias in music recommendation systems. Some research exists on bias based on nationality. Similar to research on gender, these experiments show that algorithms can reproduce the geographic advantages and disadvantages already present in the music industry and potentially show a preference for local content.^{[footnote 24]} In contrast to research on popularity bias, there seem to be no clear proposed interventions to mitigate demographic bias in the literature reviewed.

Ultimately, evidence proving or disproving the existence of “unfair” bias is incomplete and limited. An ability to better understand these biases, particularly demographic bias, would rely on greater access to data and more complete datasets.^{[footnote 25]} Collecting more complete demographic data about creators and consumers could be necessary for a better understanding of these potential biases. However, the collection of these types of data could itself prove harmful as much of this data is highly personal. Asking, or requiring, creators or consumers to provide demographic data about themselves to enable this type of research could be perceived as intrusive.

Our conversations with DSPs indicated that they are aware of the potential for these biases to exist and the harms they could cause, and are researching and implementing steps to mitigate against them. DSPs highlighted to us that their business models rely on high levels of user satisfaction, and their research shows that minimising these types of bias tend to align with this goal.

As the CDEI has previously explored in our Review into bias in algorithmic decision-making, where these types of bias exist, they can do so through entrenching previous human biases as well as introducing new ones. As such, the next section considers the condition of the music industry prior to the introduction of recommendation systems and considers how this could impact algorithmic recommendation. The section goes on to contrast algorithmic recommendations to more “traditional” forms of music recommendation, and considers the relative importance of algorithmic recommendation to that of “traditional recommendations”.

4. The wider UK music industry

4.1 Music prior to streaming

Many stakeholders we interviewed highlighted that the types of bias we were investigating in relation to algorithmic recommendations already existed in the UK’s music industry prior to the growth of streaming services. As such, the “fairness” of algorithmic recommendations, can be considered against the “fairness” of the music industry prior to these technologies.

There have always been cultural intermediaries that have shaped what music consumers hear, and how they hear it, be they radio presenters or music critics. For example, as Hesmondhalgh et al. note, the music industries of the 1950s to the 2000s had a focus on music that was listed in the charts.^{[footnote 26]} Chart listing music would receive much more attention, on radio and television, and in newspapers and magazines, than music that fell outside of the charts. The popularity bias of this system was immense, and creators with greater access to finance and resources were at a significant advantage in achieving success.

Likewise, the UK’s music industry has historically privileged certain demographic groups at the expense, and exclusion of others, both in terms of creators and other industry workers. There has been more of a concerted effort in recent years to measure and redress these inequalities. For example, UK Music launched its UK Music Diversity survey in 2016, to track progress and boost diversity and inclusion in the UK’s music industry.^{[footnote 27]}

Recommendation algorithms are in many senses a newer form of cultural intermediary, following those that came before. The stakeholders we interviewed held mixed opinions about how streaming services, and their recommendation algorithms, compare to previous forms of cultural intermediaries. Some argued that the movement towards more personalised recommendations has diminished the power of traditional cultural intermediaries, and provided greater opportunities for aspiring creators and more diverse creators to reach new audiences. Some argued that creators with greater access to resources still face significant advantages in positioning their music to receive more attention and recommendations. Others argued that algorithmic systems will have entrenched, or even exacerbated, existing biases.

As noted above, existing evidence on whether “unfair” biases exist in algorithmic music recommendation systems are fragmented and incomplete. However, we note that it is impossible to consider the impact of these systems outside of the context of the history of the UK’s music industry.

4.2 The significance of algorithmic recommendation

Algorithmic recommendations remain only one of a number of ways that consumers discover music. Although the focus of this work has been on studying the impact of algorithmic recommendations, this needs to be considered against the continued power of traditional cultural intermediaries and other types of music consumption outside of streaming.

Figures shared with us by some of the DSPs we interviewed suggest that approximately 30% percent of user listens are “guided” whereas around 70% of listens continue to be user led. This is to say, the majority of music that users listen to on streaming platforms is sought out by consumers, rather than suggested to them in that instance (of course, consumers may be relistening to music that they have previously been guided to by a streaming service).

Our consumer polling revealed that radio was the place from where respondents are most likely to get their new music recommendations (39%), followed by friends and family (31%) and social media (25%), while only 22% reported getting new music recommendations from DSPs.^{[footnote 28]} The ways in which consumers receive recommendations do vary across demographics, for example 36% of 18-25 year olds responded that they use DSPs for recommendations, compared to 9% of over 65 year olds. However, these results suggest that despite the increasing popularity of streaming services, the majority of music discovery does not occur on music streaming platforms.

This is not to downplay the significance of the role that algorithmic recommendations play in shaping music consumption, rather to reiterate that it is one, among many, influencing forces on consumer habits.

5. Transparency

The technologies that power algorithmic recommendations of music are similar to recommendation systems used elsewhere.

The specific technologies, data, and corresponding weightings that a DSP uses are closely guarded. Hesitancy by DSPs to share specific information about how their recommendation systems work is understandable. Consumers are broadly able to access the same sets of music, at a similar price across services, and therefore the quality of recommendations made is one of the core differentiating factors for the DSPs. Releasing too much information about specific recommendation systems used could undermine the competitive edge that a DSP has.

Further, DSPs expressed to us the fear that releasing specific information about how they make recommendations could increase efforts by creators and their representatives to optimise their output, with the goal of making it as recommendable as possible (also known as optimisation).

Outside of these concerns, even if DSPs aspired to be fully transparent about how their recommendation systems work, there are many practical barriers that could stand in the way. Hesmondhalgh et al. note the use of the term “black box” in academic literature, capturing how the functioning of algorithmic systems can be difficult to comprehend, even to their developers.^{[footnote 29]} Even if, as the review notes, this term is often used in a distorted, over simplified way, the challenges in explaining how complex algorithmic tools work to the public remain significant.

Despite this, some DSPs do publish information on how their technology and recommendation systems work, particularly in academic journals and blogs. For example, Deezer Research have published an array of papers about Music Analysis, Information Retrieval, Machine Learning, and Recommendation, on their website.^{[footnote 30]} Likewise, Spotify frequently publishes about the technology they use on their Engineering blog.^{[footnote 31]} However, much of this content is academic in nature, and shared in ways that creators, their representatives, and consumers are unlikely to engage with.

Outside of direct information about the function of recommendation algorithms, all DSPs provide information to creators and their representatives about how their music is being consumed (normally through a dashboard). From our conversations with the DSPs, access to this information appears to be equal, regardless of the size of the creator or who it is that represents them.

5.1 Creator feelings towards transparency

Regardless of the actual levels of information provided by the DSPs, creators and consumers generally feel that there is a lack of transparency about how these technologies work, and what they do.

An overwhelming majority of the creators that responded to our survey expressed the desire for increased transparency about how recommendation algorithms work, as many feel that DSPs do not provide sufficient information about the use of these technologies. Specifically, 83% of creators that responded to our survey disagreed or strongly disagreed that streaming services give sufficient information about how recommendations are made to consumers.^{[footnote 32]} Moreover, an even larger majority, 89%, expressed a desire for more detail as to how recommendations are made on music streaming platforms, demonstrating a call for DSPs to improve how they communicate their use of recommendation algorithms.^{[footnote 33]}

Our interviews highlighted a culture of mistrust, where creators and their representatives often feel that they receive different types and levels of information from the DSPs. In particular, there is a pervasive belief that those creators, and their representatives, signed to major labels would have access to more and better information about how consumers engage with their music, and how recommendations are made.

We found no evidence to suggest that this is the case; on the contrary this idea was rebuked in our conversations with some of the major labels and DSPs. In fact, we found in our conversations with major labels, they shared similar concerns to independent labels. However, that such an idea exists, and is commonly held, is still deserving of attention. It demonstrates widely held feelings of discontent, that themselves are concerning, and more could be done to counter.

Despite the opinion voiced by some DSPs that more information about recommendation systems could lead to greater optimisation, some interviewees flagged that a lack of knowledge about how recommendations are made on streaming services could still be affecting the way that, and the type of, music that is being made. For example, some provided anecdotal evidence about creators more frequently collaborating with one another, in the belief that multiple artists being tagged on a track would make them more likely to be recommended and to a wider audience. Some interviewees suggested that creators with greater resources would be better placed to experiment with different ways to optimise their music for recommendation in an environment of low transparency, and that greater transparency would serve to level the playing field.

This trend was also seen in our survey, which suggests that recommendation algorithms could be affecting the way creators make and release music. A majority of creators that responded to our survey (53.5%) reported that they, or their representatives, have changed the way they release music, in order to try and increase the likelihood of it being recommended. However, a much smaller percentage (16.7%) reported that they have changed the way that they make music, in order to increase the likelihood of recommendation across DSPs.

Much of the information given to us in interviews was anecdotal, and our creators survey pulls from a relatively small, unrepresentative sample. Our literature review explores existing literature on issues around optimisation of music for recommendation, emphasising the potential social impacts on how artists make music. Creators’ awareness of optimisation could pressure them to think about “designing” their music in ways that make it more searchable and discoverable by music recommendation systems.^{[footnote 34]}

Similarly, musicians might shape their music to better exploit the requirements of platforms, for example, by trying to grab attention in the first 30 seconds of a track as most DSPs remunerate on the basis of listens that last at least this long.^{[footnote 35]} If these changes have occured, they are not necessarily in and of themselves a bad thing. Music has continuously evolved as the means by which it is consumed changes. Responses to music streaming would be the latest in a series of evolutions in the face of an ever changing industry.

Currently much of the research in this space is speculative in nature, but could be explored in more detail to understand whether these effects are observed.

5.2 Consumer feelings about transparency

Public polling results reflect similar sentiments from consumers, who also report a limited understanding of recommendation algorithms and desire for increased transparency from DSPs. Among all respondents, fewer than one in five (18%) said they completely understand how these recommendations work, with one quarter (25%) reporting that they do not understand how they work at all.^{[footnote 36]} Of course, perceived understanding does not equate to true understanding, and it is likely that the levels of true consumer understanding are lower. Our interviewees highlighted that it is likely that consumers do not understand when a recommendation they have received is algorithmically or editorially generated. This is despite some naming conventions used by DSPs, such as indicating when a playlist was designed specifically for a consumer (and therefore algorithmically generated), or curated for a wider audience.

In light of this limited understanding, many respondents to our survey reported a desire for increased transparency, to better understand the technologies used. Two thirds of respondents who currently use a music streaming service agree that they would like more information about what types of data music streaming platforms collect (66%)^{[footnote 37]} as well as how that data is used (66%),^{[footnote 38]} while amongst those who have listened to recommendations, 61% agree that they would like *more information about why they get given specific recommendations.^{[footnote 39]}

These results suggest significant demand for increased transparency, and indicate that better transparency could be necessary to sustain the success of DSPs. Among those who had a streaming service in the past, a significant minority (42%) reported that a lack of information about what data music streaming services collect was a factor in their decision to stop using them.^{[footnote 40]} Similarly, 41% of respondents who had a streaming service in the past reported that a lack of information about how music streaming services used their data was a factor in their decision to stop using them. Notably, 48% of respondents who had a streaming service in the past relayed that more information about how music streaming services used their data would encourage them to use these services again.^{[footnote 41]}

6. Ideas for further action

6.1 Improvements to transparency

Changes to the ways that DSPs communicate about these technologies could help ease feelings of discontent, and better foster trust between parties in the music ecosystem.

In particular, DSPs could:

• provide a clearer indication to consumers about when a playlist is curated by algorithms, editors, or a combination of both

• offer consumers a “why am I seeing this recommendation” function, similar to features provided by other online platforms that allow consumers to understand why they receive certain advertisements

• better communicate to creators and their representatives about how to access data about how their music is consumed, and provide more information about what types of data are available and why this might vary across creators (e.g. a creator’s audience being too small)

• Produce more content, tailored to non-academic audiences about how their recommendation algorithms work.

A standardised approach to transparency could help to better foster trust between DSPs, creators, others in the industry, and consumers. DSPs could look to draw upon the Algorithmic Transparency Recording Standard developed by the CDEI and Central Digital and Data Office (CDDO) for use in the public sector. The CDEI is open to working with those in the industry to build towards this type of consistent approach to transparency.

6.2 Future potential avenues of research

This report highlights areas where an absence of evidence stands in the way of making firm conclusions about the impacts of music streaming recommendation algorithms.

We have identified several possible future avenues of research that would enhance evidence in this space:

• DSPs could look to better collaborate with academics to address gaps in the existing research landscape. For example, DSPs and academics could look to use proxies for demographic characteristics in existing data, to try to better understand the impacts of these systems on different demographic groups. The CDEI is researching ways to enable responsible access to demographic data for algorithmic bias detection.^{[footnote 42]}

• Those looking to collect better demographic data on creators and consumers could look to use a host of emerging technologies to mitigate potential harms that could emerge from collecting this data.

• For example, if demographic data on creators was held by a data intermediary, it could allow for further research on the impacts of these technologies without researchers, or DSPs, having direct access to this data. The CDEI has previously published research on data intermediaries in our Unlocking the value of data: Exploring the role of data intermediaries report.

7. Methodology

This research sits alongside other work in this space.

The Competition and Markets Authority (CMA) has conducted a market study on music and streaming.^{[footnote 43]} This includes some discussion and data regarding the use of playlisting, including algorithmic playlists, by music streaming services.

The Intellectual Property Office (IPO) chairs working groups composed of experts from across the music industry, developing industry-led solutions to issues around transparency and metadata.

The IPO has also commissioned independent research into the impacts of three potential changes to copyright law recommended by the Select Committee: equitable remuneration; contract adjustment mechanisms; and rights reversion. This research will inform the government’s approach on creator remuneration.

We have worked closely with colleagues from across DCMS, the IPO, and the CMA, to ensure that we draw lessons from their work, and minimise duplication.

Academic literature review

We commissioned an academic literature review from a team of researchers at the University of Leeds, led by Professor David Hesmondhalgh. This literature review explores how, and to what degree, existing research has addressed a number of issues surrounding algorithmically-driven music recommendation systems, in particular:

• The question of “bias” in music streaming algorithms: how might different groups of artists and consumers be affected by algorithms?

• The question of diversity: positive and negative impacts of algorithms on musical diversity.

• Questions of transparency, opacity, and oversight.

This literature review explores content from both computer science research and “critical” social science and humanities research (particularly a sub-field known as critical algorithm studies). This review was produced in April 2022, and therefore does not include any literature published after this date.

The full literature review has been published alongside this report. This report draws heavily on the findings of the literature review throughout.

Public polling

We partnered with Deltapoll and polled a representative sample of around 4,000 UK adults^{[footnote 44]} to better understand their music consumption, their thoughts and feelings towards both music streaming and algorithmic recommendations, and the levels to which they understand how and when these types of technology are used.

Results from this polling have been drawn upon throughout this report. The full results from this polling, the questions we asked consumers, and a breakdown of the demographics of those we polled have been published alongside this report.

Creator survey

We ran a survey for creators to report on their experiences with streaming services, and their understanding of how technologies, like recommendation algorithms, work. This survey was distributed through the CDEI’s social media channels, and the organisations that form the Council of Music Makers (CMM).^{[footnote 45]}

One limitation of this survey was its relatively small sample size (102 respondents). Another limitation of this survey was that we were not able to ensure that respondents were representative of the creator community. We asked respondents to optionally self-identify certain demographic characteristics, to better understand the diversity of respondents. The responses highlighted that respondents to our survey were not representative of the UK population, for example, 73.5% of respondents self-identified as male. Further, we are conscious that those opting to participate in this survey are likely to be more interested in, and hold opinions about, the technologies we are considering.

Results from this survey have been drawn upon throughout this report. The results of this survey, the questions asked, and the self-reported demographics of the participants have been published alongside this work.

Industry engagement

We conducted semi-structured interviews with stakeholders from across the music industry, including representatives from: Labels (both major and independent), publishers, distributors, the DSPs, and the organisations that make up the CMM. These interviews were conducted under the condition of anonymity, however we note in this report when a conclusion is based on statements given to us in these interviews that we do not have additional evidence to support. We believe that the thoughts and feelings of these interviewees remain important, regardless of their accuracy. Widely held, inaccurate beliefs might indicate a need for more communication between groups within the music industry.

We chose to use semi-structured interviews because they allowed for a flexible approach given the diversity of organisations we were engaging with. Further, this allowed interviewees the scope to guide conversation towards topics they were most interested in or concerned with.

Outside of these interviews, we have maintained a working relationship with stakeholders from across the industry, who have been able to provide further evidence and clarifications to us when necessary. We have tested early findings of this report with the IPO’s Music Industry Contact Group.

This project has relied on people from across the industry, government, and academia giving their time and expertise. We are grateful for all those that contributed to this work.

8. Glossary of terms

Cultural Intermediaries - People, organisations, and technologies that mediate how cultural goods are perceived and engaged with.

Demographic Bias - Bias according to demographic characteristics. “Demographic” can refer to many different characteristics, including race, ethnicity, gender, class, age, disability, sexuality, and nationality.

Digital Service Providers (DSPs) - Defined by the ICO as an organisation that provides “an online search engine, online marketplace or cloud computing service (either alone or in combination)”. In the context of this work, DSP refers to an entity that provides digital music services.

Popularity Bias - Popular items receiving a lot of exposure, while less popular items are under-represented.

Recommendation Systems - An information filtering system that provides suggestions for items that are most pertinent to a particular user.

1. House of Commons Digital, Culture, Media and Sport Committee, Economics of Music Streaming, 2021. ↩

2. House of Commons Digital, Culture, Media and Sport Committee, Economics of music streaming: Government and Competition and Markets Authority Responses to Committee’s Second Report, 2021. ↩

3. Competition and Markets Authority (CMA), Music and streaming market study, 2022. ↩

4. British Phonographic Industry (BPI), BPI publishes its yearbook “All About the Music 2021”, 2021. ↩

5. The CDEI, The impact of recommendation algorithms on the UK’s music industry: Deltapoll music streaming polling data, CD1. ↩

6. Deltapoll Survey Results, CD1. ↩

7. David Hesmondhalgh, Raquel Campos Valverde, D. Bondy Valdovinos Kaye, Zhongwei Li, The Impact of Algorithmically Driven Recommendation Systems on Music Consumption and Production: A Literature Review, 2022, s1.3 ↩

8. Hesmondhalgh et al., Literature Review, 2022, s3.2. ↩

9. The CDEI, Review into bias in algorithmic decision-making, 2020. ↩

10. The CDEI, The impact of recommendation algorithms on the UK’s music industry: creators’ survey results, Section 3, Q1. ↩

11. The Future of Music Streaming Survey, Section 3, Q2, Q3. ↩

12. Deltapoll Survey Results, CD23_2. ↩

13. Deltapoll Survey Results, CD23_3. ↩

14. Hesmondhalgh et al., Literature Review, s4.2. ↩

15. Hesmondhalgh et al., Literature Review, s4.2. ↩

16. Hesmondhalgh et al., Literature Review, s4.2. Jannach et al (2013), Celma (2010) ↩

17. Hesmondhalgh et al., Literature Review, s4.2. ↩

18. Hesmondhalgh et al., Literature Review, s4.2. Hesmondhalgh et al. (2021) and Ingham (2018). ↩

19. Hesmondhalgh et al., Literature Review, s4.2. ↩

20. Hesmondhalgh et al., Literature Review, s4.4. ↩

21. Note that only in 2016 did Spotify introduce “non-binary” as a category for registration, which Hesmondhalgh et al., Literature Review (s4.4) suggests explains the lack of studies on bias related to non-binary and genderqueer artists and listeners. ↩

22. Hesmondhalgh et al., Literature Review, s4.4.2. Eriksson and Johansson (2017), Werner (2020). ↩

23. Hesmondhalgh et al., Literature Review, s4.4.2. Aguiar et al. (2021). ↩

24. Hesmondhalgh et al., Literature Review , s4.4.3. Way et al (2020). ↩

25. Hesmondhalgh et al., Literature Review note that the majority of studies into bias currently rely on simulations rather than running experiments on actual users interacting with music recommendation systems. ↩

26. Hesmondhalgh et al., Literature Review , s4.2 ↩

27. UK Music, UK Music Diversity Report 2020, 2020. ↩

28. Deltapoll Survey Results, CD3. ↩

29. Hesmondhalgh et al., Literature Review, s6. ↩

30. Deezer, Understanding music and how people relate to it, accessed September 2022. ↩

31. Spotify, Spotify Research is dedicated to extending the state of the art in audio, accessed September 2022. ↩

32. The Future of Music Streaming Survey, Section 2, Q4. ↩

33. The Future of Music Streaming Survey, Section 2, Q5. ↩

34. Morris (2020), Morris et al (2021) - in lit review, s4.3. ↩

35. Hesmondhalgh et al., Literature Review, s4.3. ↩

36. Deltapoll Survey Results, CD13. ↩

37. Deltapoll Survey Results, CD20_1. ↩

38. Deltapoll Survey Results, CD20_2. ↩

39. Deltapoll Survey Results, CD23_1. ↩

40. Deltapoll Survey Results, CD21_1. ↩

41. Deltapoll Survey Results, CD21_2. ↩

42. The CDEI, Enabling responsible access to demographic data for algorithmic bias detection, 2022. ↩

43. Competition and Markets Authority, Music and streaming market study, 2022. ↩

44. 4120 consumers. ↩

45. The CMM consists of: The Ivors Academy, The Featured Artists Coalition, The Music Managers Forum, The Music Producers Guild, and The Musicians’ Union. We are grateful for their help in distributing this survey. ↩