The Sad Tale of Music Algorithms (or) Why Your Indie Band Isn't Famous Yet
Melanie Bainbridge
Jan 18, 2024 - 6 min read
Updated: Jan 19, 2024
The Algorithm Has Taste - and its taste, it turns out, looks a lot like whoever is paying.
There is a question that surfaces, in one form or another, every time someone in the music industry is feeling either honest or sufficiently caffeinated to say the quiet part out loud: do streaming services play favourites?
The short answer is yes, obviously, in ways that are structural and consequential and not particularly secret. The longer answer is more interesting, because it turns out the favouritism is not simply a matter of corrupt executives with questionable taste. It is a feature of a system that was designed, with considerable sophistication, to look like neutrality while functioning as anything but.
I have some professional interest in this question, as co-founder of The Pack Music Co-operative — Australia's first musician-owned streaming platform, built specifically on the premise that the current arrangement is both broken and, importantly, not inevitable. So I want to be upfront about my position: I am not a disinterested observer. But I also think the argument stands on its own merits, and I will try to make it do so.
In 2021, Spotify introduced something called Discovery Mode.(1) The pitch was elegant: artists could apply to have their tracks given a boost in Spotify's recommendation algorithm, increasing their chances of appearing in personalised playlists and radio stations. The catch, disclosed in the fine print with the confidence of a company that knows its users do not read fine print, was that participating artists would accept a reduced royalty rate in exchange for the algorithmic lift. Pay to play, dressed up as an opportunity.
Spotify described this as a tool that gives artists of all sizes access to the same promotional infrastructure previously available only to major labels. Which is true, in the same way that it is technically true that everyone is free to sleep under a bridge. The option exists. The conditions under which most independent artists could meaningfully exercise it are another matter entirely.
Because the artists who can most easily absorb a reduced royalty rate are the ones already earning enough that the reduction is manageable. The artists for whom a royalty cut is genuinely painful are precisely the ones who most need the visibility boost. Discovery Mode did not democratise algorithmic promotion. It reproduced the existing power structure with a new interface and a gentler name.
To understand why this matters, it helps to understand the architecture of the industry the algorithm is operating inside.
Four major labels — Universal Music Group, Sony Music Entertainment, Warner Music Group, and Merlin (the independent label collective) — control the majority of the music available on streaming platforms.(2) They also hold licensing agreements, equity stakes, and commercial relationships with those platforms that are not available to independent artists. When Spotify negotiates with Universal, it is negotiating with a company that owns a piece of Spotify. This is not a secret. It is not, for the most part, even presented as a problem. It is simply the structure.
The consequence is that the algorithm does not operate on a level surface. It operates on a surface that was tilted before the algorithm arrived, and the algorithm, trained on data produced by that tilted surface, has learned to recommend accordingly. A track signed to a major label arrives in the system with advantages that have nothing to do with the quality of the music: promotional budgets, editorial relationships, playlist placement negotiated at a corporate level, and the accumulated streaming history that comes from having been the dominant force in popular music for decades. The algorithm sees all of this data. It learns from it. It recommends accordingly.
This is what researchers call a feedback loop, and it is the structural mechanism underneath all the more visible forms of streaming inequality. Popularity breeds recommendation. Recommendation breeds more plays. More plays breed more popularity. At the end of this loop, the already-popular get more popular, and the independent artist releasing their third album to an audience of a few thousand dedicated listeners stays, statistically, exactly where they are. Not because their music is worse. Because the system is not measuring quality. It is measuring momentum, and momentum is unevenly distributed by prior wealth and prior access.
The bias in music recommendation algorithms is not hypothetical. It is documented.(3) Studies have found consistent underrepresentation of female artists in algorithmically generated playlists, independent of listener preference data — which means the algorithm is not simply reflecting what people choose to hear, it is shaping what they are offered before the choice is made. Research has also found genre-based and geography-based biases: artists from outside the English-speaking world, artists working in genres not well-represented in mainstream commercial catalogues, artists whose sonic profiles do not fit neatly into the categories the algorithm was trained to recognise, are all systematically less likely to surface in recommendations.
An algorithm, like a child, is shaped by what it is fed. If the training data is predominantly mainstream commercial music from English-speaking markets signed to major labels, the algorithm learns that this is what music sounds like, what good music sounds like, what recommendable music sounds like. This is not malice. It is the ordinary operation of machine learning on biased data. But the ordinariness of the mechanism does not reduce the consequence, which is that the system continuously recommends the familiar, marginalises the different, and calls this process personalisation.
The filter bubble is the endpoint of this process. Your listening history is used to predict what you will like, and what you will like is assumed to resemble what you have already heard, and what you have already heard was already shaped by the algorithm's prior predictions, and around we go. The result is a listener who believes they have unlimited access to the entire history of recorded music and is, in practice, being served an increasingly narrow corridor of it, tailored to their existing preferences with the efficiency of a system that has confused familiarity with satisfaction.
The word fairness gets deployed a lot in these conversations, and it tends to dissolve under examination because people are using it to mean different things. There is fairness as equal access — every artist gets the same shot at recommendation. There is fairness as proportionality — artists are recommended in proportion to their actual audience size and engagement. And there is fairness as equity — the system actively accounts for structural disadvantages and adjusts accordingly.
The first definition is the one platforms claim to be pursuing. The second is roughly what they are actually doing. The third is the one that would require changing who owns the infrastructure, which is why nobody in a position to implement it is particularly enthusiastic about the concept.
I want to be honest about the limits of the fairness argument, because I think the honest version of it is more useful than the utopian one. Policing taste is not the goal and not possible. People listen to what they like. The question is not whether the algorithm should force everyone to listen to more obscure Tuvan throat singing — it should not, and the person who wants to spend their evening with Taylor Swift is making a perfectly valid choice that deserves to be respected. The question is whether the system that mediates between the listener and the available music is designed to expand or narrow what they are offered before the choice is made.
Research consistently finds that listeners with more diverse music consumption have, on average, broader taste development over time — which is to say that exposure to the unfamiliar tends to generate appetite for more of the unfamiliar, and appetite for the unfamiliar is good for independent artists, good for genre diversity, and good for a music ecosystem that is not entirely dependent on a small number of extremely powerful labels deciding what gets made.
There are practical interventions that exist and work. Re-ranking algorithms to weight diversity alongside predicted preference. Diversity metrics built into playlist generation. Human editorial curation that operates with genuine independence from label relationships — which costs money, which is why most major platforms have been systematically reducing their human curation teams in favour of fully automated systems, because fully automated systems are cheaper and the cost savings outweigh, in a quarterly earnings sense, the quality loss. Transparency requirements that force platforms to disclose how their recommendation systems work, who has paid for promotional placement, and what the royalty implications of algorithmic participation are.
Some of this is beginning to happen through regulation rather than voluntary reform.4 The EU's Digital Markets Act and various ongoing legislative processes around platform accountability are slowly creating pressure for disclosure that the platforms would not, left to their own devices, provide. This is the predictable shape of these arguments: the system does not reform itself. It reforms when the cost of not reforming exceeds the cost of reforming, and that calculation shifts primarily through regulatory pressure and, occasionally, through the market signal of listeners and artists choosing to go elsewhere.
Which brings me, with no particular embarrassment about the segue, to why The Pack exists. The cooperative model is not a nostalgic fantasy about a pre-digital music economy. It is a structural argument about who should own the infrastructure that mediates between artists and listeners, and what that ownership produces. When the people whose work generates the value also own the system that distributes it, the incentives align differently. The algorithm — if there is one — is answerable to its owners, and its owners are the artists. The question of whether the system plays favourites becomes, at minimum, a question that the affected parties have standing to ask and power to affect.
This is not a complete solution to the problems I have described. A cooperative streaming platform still operates in a music industry structured around the four major labels. It still serves listeners who have been trained by years of algorithmic narrowing to expect a particular kind of experience. It still has to be financially viable in a market where the dominant players have structural advantages that no amount of good intentions can fully offset. I am not selling a utopia.
I am selling a different set of incentives, and arguing that different incentives produce different outcomes, and that we know this from evidence rather than theory. The history of cooperative enterprise is not a history of failure. It is a history of enterprises that asked different questions, answered to different stakeholders, and produced different distributions of value. Mondragon. The Rochdale Pioneers. The dozens of worker-owned and artist-owned enterprises that operate at scale, under competitive conditions, and demonstrate consistently that the choice is not between the current arrangement and chaos.
The algorithm has taste. Its taste was taught to it by the data it was fed, and the data it was fed reflects a century of commercial music infrastructure organised around the interests of a small number of very large companies. This is not a conspiracy, but it is an outcome. And outcomes, unlike conspiracies, can be changed by changing the conditions that produce them.
The next time a streaming service serves you a recommendation, it is worth asking what the recommendation is for. Sometimes it is for you — a genuine attempt to match you with something you will love that you have not yet found. Often it is for the platform — keeping you listening, keeping you subscribed, keeping the engagement metrics healthy for the quarterly report. Sometimes it is for a label that has paid, in one form or another, for the placement. These are not mutually exclusive, but they are not identical, and the conflation of them is doing a great deal of work in the service of an industry that prefers you not to think too carefully about the difference.
Every artist whose music you have ever loved started somewhere. Most of them started in the margins of a system that was not designed to surface them. Some of them made it through anyway, through luck, through word of mouth, through the kind of human recommendation that happens when someone grabs your arm and says you have to hear this. Some of them did not.
The ones who did not make it are not footnotes. They are the majority of the music that was ever made. The algorithm does not know they exist. It was not built to.
That is the thing we are trying to fix.
* * *
References and inspirations
1. Spotify Discovery Mode was announced in November 2020 and launched in beta in 2021. The reduced royalty structure in exchange for algorithmic promotion was documented extensively at launch; see coverage in Music Business Worldwide and The Verge. The Federal Trade Commission has since examined the practice as part of broader scrutiny of streaming platform commercial arrangements.
2. On major label market share and structural relationships with streaming platforms, see the Music Industry Blog (Mark Mulligan) and various annual reports from IFPI (International Federation of the Phonographic Industry). The equity relationships between major labels and streaming platforms at various points in their development are documented in each platform's public filings.
3. On gender bias in music recommendation algorithms, see Celma and Herrera's foundational work on the long-tail problem in music recommendation (2008), updated by more recent studies including Ferraro et al., 'Artist Gender Representation in Music Streaming' (RecSys 2021). On geographic and genre bias, see the substantial literature on algorithmic fairness in recommender systems, including work from the ACM Conference on Recommender Systems.
4. On platform regulation and transparency requirements, see the EU Digital Markets Act (2022) and the ongoing legislative work in multiple jurisdictions around algorithmic accountability. The pace of implementation is, to put it generously, measured.