Algorithmic transparency, fairness, and explainability in music discovery are at the very center of our work at Music Tomorrow. Over the years, we’ve explored how algorithmic biases can take shape — from structural popularity loops to the way personalization may inadvertently limit diversity — and examined gender gaps in music production and consumption. Our goal was never to claim that the streaming landscape is inherently unfair — but rather raise important questions about how fairness could be defined and measured in a digital music ecosystem.

Fast forward to 2025, and those questions remain as urgent as ever. What does “fairness” mean across such a complex landscape of artists, platforms, and audiences? How transparent have streaming services truly become under new regulations like the Digital Services Act (DSA)? And what progress has been made — through research, policy, and new tools — toward understanding and correcting the biases that shape music discovery today?

This article takes stock of where we stand at the end of 2025: examining what fairness means in an ecosystem shaped by many — often conflicting — interests, the gaps between research and industry practices, as well as emerging initiatives and frameworks that push the music industry toward greater accountability and transparency.

Why defining fairness in music recommendations is a challenge — and what do we mean by “fair”?

Working at the crossroads of research, industry, and artist practice has given us a close-up view of how the conversation around fairness in music recommendations has evolved. What once felt like an abstract, technical topic is now part of a much bigger discussion — one that spans research circles, label boardrooms, and artist communities alike. In the past few years, we’ve seen fairness move from theory to practice, becoming a shared concern about how music is discovered and consumed in an increasingly AI-driven world.

Fairness in music recommendations is a multi-stakeholder issue — one that touches everyone from artists and rights holders to listeners, platforms, and public institutions. Yet, there’s a major caveat: each of these groups views fairness through a different lens, often leading to conflicting priorities and definitions.

For listeners, fairness might mean getting recommendations that are personalized without pigeonholing or discrimination. They want algorithms to avoid biases that could, say, stereotype their tastes by age or gender, and instead fairly reflect their individual interests. A listener might consider it “unfair” if a recommender assumes they only like a narrow genre because of their past listening or demographic profile. Fairness, to users, is about being treated impartially and, ultimately, being exposed to music they’ll actually enjoy – not just the same mainstream hits or what people “like them” typically listen to.

For music creators and providers, fairness is largely about exposure. Artists and labels question whether the system gives their songs a reasonable chance to be recommended, or if the algorithms disproportionately favor other content (often major hits or big-name artists). From this standpoint, a fair recommender system is one that doesn’t only reward established acts; it also surfaces emerging artists, local scenes, and niche genres to listeners.

Artists tend to frame fairness in human terms: are we being given a fair shot, or are we trapped in algorithmic limbo? As algorithms increasingly mediate both discovery and monetization, creators and copyright holders have raised understandable transparency concerns — seeking insight into how platforms treat their music and explanations for why a track might be picked up (or overlooked) by streaming recommenders.

Cultural institutions and regulators, on the other hand, view fairness in collective terms — as a question of cultural diversity and equity. They worry about the macro effects of recommendation systems on culture: are algorithms limiting the public’s exposure to emerging or minority artists, or are they fostering a rich ecosystem that represents varied, local voices? Regulators are concerned that opaque recommendation engines might amplify structural inequities — favoring Anglophone content or major-run catalogues — and undermine cultural diversity goals. This perspective often translates into calls for transparency and accountability, pushing platforms to prove their algorithms aren’t unfairly disadvantageous to certain groups.

Finally, streaming platforms themselves — Spotify, Apple Music, Deezer, and others — are caught in the middle, balancing business objectives with fairness expectations. On one hand, their primary goal is user satisfaction and retention; on the other, they face increasing pressure to ensure a level playing field for artists and avoid reputational or regulatory backlash. For platforms, “fairness” means maintaining algorithms that maximize engagement while treating all users and content providers equally. In practice, this balancing act is tricky. As we previously covered, short-term engagement optimization tends to reward familiarity — the hits, the comfort zone — which can bury emerging artists or trap users in feedback loops. Heavy-handed diversity re-ranking, on the other hand, risks creating uneven experiences and compromising user satisfaction.

Crucially, these differing ideals of fairness can and do clash. Each stakeholder has valid concerns, but optimizing for one can compromise another. The heated discussions around Spotify’s Discovery Mode are a perfect example of these tensions.

From one perspective, Discovery Mode offers artists — particularly smaller ones — a new tool to boost discoverability by accepting a reduced royalty rate. That’s one interpretation of fairness: giving artists more agency over their algorithmic exposure.

But critics see it differently — they argue Discovery Mode creates a two-tier system where those who can afford a revenue cut gain privileged algorithmic treatment. Listeners, meanwhile, aren’t explicitly told what tracks are boosted or given an option to opt out of sponsored recommendations — raising further transparency concerns.

In the end, what one group sees as empowerment, another sees as bias. Any meaningful progress will require navigating these trade-offs — finding a balance that allows recommendation engines to serve culture, artists, listeners, and business goals without unfairly tipping the scales toward any one side.

What gets built? Disconnect Between Research and Industry Practices

While the industry wrestles with defining fairness, the academic world has been busy tackling algorithmic bias in a more technical way – and there’s often a major disconnect between the two approaches.

In computer science, “fairness in recommender systems” often becomes a mathematical optimization problem: ensuring equal exposure, accounting for biases in training datasets, adjusting relevance weighting, or balancing re-ranking constraints. Over the past few years, researchers have proposed dozens of models to make recommendations more equitable — from re-ranking algorithms that deliberately include underrepresented artists to loss functions that penalize popularity bias. Technically, many of these approaches work — but they tend to define fairness narrowly, optimizing for a single metric rather than a broader cultural outcome.

Meanwhile, the music industry speaks an entirely different language. Artists and labels rarely talk about statistical parity; they talk about whether they can reach audiences organically, or whether their releases are “stuck in the algorithmic limbo.” Fairness, to them, means opportunity and transparency, not necessarily balanced exposure ratios. This mismatch creates a persistent gap between research and practice: engineers focus on metrics, while practitioners debate values.

On a more positive side, there’s growing recognition that this gap must be bridged. Scholars in the fairness domain have started calling for more interdisciplinary work: they acknowledge that algorithmic solutions need conceptual clarity and buy-in from domain experts. This means involving artists, curators, institutions, and listeners in defining what the goals of a fair recommender should be before setting technical objectives.

The key is aligning on the right fairness goals together, then letting the technical experts figure out how to implement them. We’re seeing early steps in this direction – more dialogues between music stakeholders and AI experts at conferences and in initiatives like the FAccTRec workshop series that explicitly bring together researchers and practitioners.

It’s an ongoing journey, but the hope is that by 2025, “fairness in music AI” is no longer just a lab experiment or abstract ideal – it’s a collaborative effort to define and build recommender systems that reflect the music ecosystem’s values. But getting there would require a shared evidence base — a way for both researchers and practitioners to see, test, and measure how recommendation systems behave in the real world. But that’s precisely what’s missing today:

Why are fairness audits so difficult to conduct in practice?

One major obstacle hanging over much of these efforts is something rather basic: who has access to the data and the algorithms? The short answer is, the platforms do – and almost no one else. This asymmetry has created a data access bottleneck that makes it incredibly hard for outsiders to assess or audit fairness in music recommendation systems.

Most academic studies have had to rely on a handful of publicly available datasets that only partially reflect real-world streaming scenarios. A recent review noted that the majority of research papers on music recommender fairness use data from Last.fm – specifically, various datasets of listening history released by the platform. Last.fm, while popular in its day, now has a niche user base and a different usage model (“scrobbling” listening from multiple sources) compared to mainstream streamers. The common Last.fm datasets (LFM-1b, LFM-360K, etc.) date back several years and are not directly representative of today’s Spotify or Apple Music environment.

In other cases, researchers use proprietary datasets shared through one-off collaborations or simplified data like the Spotify Million Playlist Dataset (which contains playlists but not the context of algorithmic feeds). By and large, external researchers are analyzing whatever data they can get – and that often means older or incomplete proxies for the real thing.

Meanwhile, Spotify, Apple Music, and others sit atop real-time interaction data that could enable meaningful audits — showing how recommendations differ across demographics, how often artists appear in algorithmic playlists, or how exposure evolves over time. However, platforms have been extremely cautious about sharing detailed recommendation data or algorithm logs with the public — citing legitimate privacy and commercial confidentiality concerns.

Whatever the reasoning is, however, the data access challenges means that external researchers can’t easily audit industrial recommenders. Some now use indirect methods — like creating “sock puppet” accounts to observe recommendation differences, or collecting listener-sourced consumption data — but these experiments remain limited in scale. The data bottleneck remains one of the biggest obstacles to credible progress — and until access improves, academic fairness assessments will remain somewhat speculative.

As researchers struggle with limited visibility, regulators have begun to take notice. The lack of transparency in algorithmic systems isn’t just an academic concern — it’s now recognized as a public-interest issue. This recognition has shaped a wave of policy interventions, especially in Europe, where new digital legislation is redefining what platforms owe to users, creators, and society at large.

How are EU laws reshaping algorithmic accountability?

With transparency and fairness concerns mounting, regulators have begun to step in. Most notably, the EU’s twin legislative pillars — the Digital Services Act (DSA) and Digital Markets Act (DMA) — were introduced to redefine the obligations of online platforms.

The DSA, which took effect in 2024, directly addresses recommendation systems. Large platforms must now disclose “the main parameters” that determine what content users see, explain those mechanisms in accessible language, and offer at least one option for a non-personalized feed. In theory, this means streaming services like Spotify, YouTube Music, or TikTok must clarify how personalization works and give users the ability to opt out.

The DMA, meanwhile, targets systemic gatekeepers like Apple, Google, and Meta. It bans self-preferencing and mandates that business users (including rightsholders on YouTube or TikTok) can access the data generated by their own activity. While Spotify and other independent streaming services aren’t currently designated as gatekeepers, these rules still set important precedents — particularly around data access and competitive fairness.

The real-world impact of this legislation so far has been modest but visible. Platforms have quietly published explanations of their recommendation parameters and introduced opt-out toggles. These disclosures remain very high-level — and often buried in legal pages — but they mark the beginning of a transparency baseline that regulators can build on.

Longer term, the DSA’s requirement for independent algorithmic audits may become the most significant element: if properly enforced, it could finally enable third parties to assess systemic risks such as cultural bias or discriminatory patterns — issues that have, until now, remained internal to platforms. For now, however, the primary result of the EU’s digital accountability legislation is cultural: for platforms, transparency is shifting from a public relations issue to a compliance requirement — and that has the impact on how the industry views and talks about algorithmic accountability.

Who’s driving fairness and transparency today? Emerging Fairness & Transparency Initiatives To Follow

While regulation provides the stick, innovation provides the compass. In the past years, a growing ecosystem of initiatives has emerged to study, measure, and promote fairness in music recommendation systems. To highlight a few notable contributors:

Fair MusE: A major EU-funded consortium uniting researchers in law, economics, and computer science with music industry and policy experts. Fair MusE investigates how streaming and social media algorithms affect artists and audiences, developing a “Fairness Score” framework, data transparency portal, and policy recommendations for a fairer digital music ecosystem.

Open Music Observatory: Originating from the OpenMusE project (a sister project of Fair MusE under the Horizon Europe programme), the observatory is building an open-source data platform integrating streaming statistics, metadata, and cultural indicators. Its aim is to visualize disparities across genres, languages, and geographies, allowing institutions to monitor representation patterns over time.

UK Code of Good Practice on Transparency in Music Streaming: A voluntary, industry-wide agreement launched in 2024 under the oversight of the UK Intellectual Property Office. Developed by 12 major organisations representing every link in the music value chain — from artists and managers to labels, publishers, and collecting societies — the Code establishes shared standards for transparency around contracts, licensing, royalties, and usage data. While streaming platforms are not direct signatories, they are represented indirectly through the Entertainment Retailers Association (ERA). 

The initiative marks an important collective step toward a more open and accountable streaming economy, aiming to build trust between creators and rights-holders, encouraging consistent data and reporting practices, and sets a public benchmark for fair conduct. The IPO will convene review meetings every six months and conduct a formal evaluation in 2026, ensuring the Code continues to evolve as the industry’s transparency expectations mature.‍

FAccTRec workshop series is hosted as part of the ACM Conference on Recommender Systems, focusing on new standards and frameworks for measuring bias and building explainable recommendation systems — while bringing together scientists and industry practitioners to build a more practical vision of fairness in RecSys. These developments, though academic-first, and not music-specific, still influence how fairness is conceptualized across creative industries.

Music Tomorrow’s Work: Explainability Framework for Algorithmic Discovery

At Music Tomorrow, algorithmic observability has become a cornerstone of our mission to make recommendation systems more transparent and explainable. Over the past few years, we’ve developed a framework that models how streaming algorithms “see” artists — what you could think of as their algorithmic profile.

An algorithmic profile maps an artist’s position within the recommendation space: which other artists, playlists, and listener clusters are most closely associated with them according to the platform’s underlying dynamics. It doesn’t replicate the algorithm itself but makes visible the relationships that shape exposure — turning abstract algorithmic behavior into something that can be observed and measured, both quantitatively and qualitatively.

We’ve already put our models to work to offer insights into the streaming discoverability landscape to cultural institutions and policymakers. In joint policy-oriented research projects, our models allowed us to explore how different groups of artists are positioned within algorithmic ecosystems, quantifying their relative exposure and discoverability to provide a backbone for broader discussions on content diversity and cultural representation.

Our goal is to create the conditions for evidence-based dialogue — to understand how recommendation systems shape visibility, and how that understanding can inform cultural policies and help build a fairer digital music ecosystem. We’ll be sharing more findings and case studies from this work in the coming months: subscribe to our newsletter to not miss out on future updates!

Conclusion

In recent years, fairness and transparency in music recommender systems have evolved from theoretical debate into a more grounded, multi-layered challenge. Researchers have developed robust methods to detect and mitigate bias; regulators have established a legal foundation for algorithmic accountability; and platforms are beginning to assess their own systems with a new sense of responsibility.

Still, meaningful and sustained progress will depend on continued collaboration. Fairness in music recommendations is not a box you can tick — but an evolving negotiation between creativity, commerce, and technology. At Music Tomorrow, we believe that the next frontier lies in observability — the ability for creators, institutions, and researchers to see concrete evidence for how algorithms shape exposure, and use that understanding to push toward more diverse and equitable outcomes for the music industry and culture at large.

That will remain our focus: opening up the black boxes of the music industry, and transforming algorithmic transparency from an abstract principle into a practical foundation for a healthier, fairer music ecosystem.