1. The quiet transfer of gatekeeping
For most of the recorded-music era, gatekeeping was visible. A radio programmer, an A&R executive, a magazine editor, a record-store clerk — were the people who decided which songs reached which ears. They were imperfect. They were also legible. A musician who got cut from a station's playlist could call the station and ask why; a listener who disliked a song could identify the human who had chosen it. Curation, in other words, had a face.
The new curators are outputs. A neural network trained on a billion listening sessions and a million track embeddings does not pick songs because it has heard them and found them moving. It picks songs because they sit in a region of latent space where other songs have performed well for listeners who resemble you. The decision is statistical, not aesthetic, and it is also invisible. By the time a track reaches a listener's "Discover Weekly," it has been filtered through dozens of intermediate steps, none of which are labeled, and most of which the platform itself would struggle to describe in plain language.
What changed is not whether curation happens, but who is accountable for it. When a human editor overlooks a worthy artist, the artist can at least name the editor. When a model does, the artist faces a wall of probability. This is the first ethical problem of AI music curation, and it is the one that all the others depend on: the gatekeeper is now a system that no one is empowered to argue with.
2. Whose music is the model?
The training data tells the story. A model trained primarily on Western popular music, on commercially successful tracks, on songs that have already survived the algorithms of previous outputs, will reproduce those patterns. It will be confident in the center of the distribution and timid at its edges. This is not a bug; it is how dense representations work. But it is also an ethical problem, because the edges of the distribution are where most of the world's music lives.
Consider a few examples. A model trained on Spotify's top charts will struggle with West African highlife, with Kurdish dengbêj, with the regional variants of Carnatic music, with the underground scenes that exist primarily on Bandcamp and SoundCloud. The model will not say these genres do not exist. It will simply render them invisible, by failing to recommend them and failing to retrieve them when a listener asks. The musician in Kano who records a brilliant album and uploads it to a DSP will find that the recommendation system has no slot for her. The listener in Berlin who searches for something like it will find suggestions for African-influenced pop instead.
A second, related problem is appropriation. AI systems are now capable of producing tracks that imitate specific regional styles — reggaeton, gamelan — platforms are beginning to license or host these imitations alongside the originals. The economic effect is straightforward. A generated track costs nothing to produce and pays nothing to the community that developed the style. The listener often cannot tell the difference. The original artist, meanwhile, sees her streaming numbers fall because a model has learned to do a competent impersonation of her tradition and is now competing with her for the same recommendation slots.
The deeper issue is that "bias" in this context is not a deviation from a neutral baseline. There is no neutral baseline in music. Every dataset reflects choices about what was recorded, what was archived, what was digitized, what was labeled, and what was promoted. A recommender system trained on such a dataset inherits those choices and amplifies them. The output is not biased against the periphery; it is biased toward the center, and the center has been built by fifty years of gatekeeping decisions that were themselves biased toward the center.
3. The consent of listeners and musicians
The recommendation system runs on data. It runs on every play, every skip, every search, every moment of silence mid-track. Most users have agreed to this data collection, in the broad and unread way that most users agree to most terms of service. Few have a meaningful picture of what their agreement produces. The platform knows, with high accuracy, when a listener is sad, when she is working out, when she is trying to fall asleep, and when she is in love. It can act on this knowledge in real time. It does not ask her permission each time it does so, because the permission was bundled into the original click.
This is not a new problem. Surveillance capitalism has been the subject of extensive critique. What is new is that the surveillance is now being used to shape cultural experience, not merely to monetize it. A streaming service that knows a listener is grieving, and chooses to serve her an upbeat playlist rather than a melancholy one, is making an editorial decision. A service that chooses to keep her in a mood-matching loop, because mood-matching loops drive longer sessions, is making a different editorial decision with the same data. Neither decision is visible to the listener unless she goes looking for it, and few do.
Musicians face a parallel asymmetry. A new artist uploads a track and immediately becomes a data source for the recommender. Her song is fingerprinted, embedded, clustered, and scored. She is rarely told what score it received, why, or how to improve it. If the platform decides that her audience is small and her growth rate is slow, it may quietly stop recommending her. The artist will see her numbers plateau, and she will have no way to know whether the cause is her music, her metadata, her cover art, the platform's mood, or a quiet decision made by a model that no one at the company fully understands either.
There are partial fixes. Some platforms now offer an "exclude from AI training" toggle on uploads — for generative systems, less meaningful for recommendation outputs, since recommendations always learn from listening data. Some offer artists dashboards that show roughly which listener segments engage with their work. Few offer explanations in the sense that matters most: a description, in plain language, of why a particular song was recommended to a particular listener at a particular time. That kind of explainability remains an open research problem, and an unresolved ethical one.
4. What a more honest curation economy could look like
There is no clean path forward. The technology is not going to be unwound, and the volume of music being made is too large for human curation to scale to it. The question is what conditions would make the current system defensible.
Four conditions are worth naming.
Auditable outputs. Platforms should publish, at least in summary form, what their recommendation systems are optimizing for. "Engagement," "session length," and "diversity" are different objectives with different downstream effects, and listeners and artists deserve to know which one is in play. The audit does not need to expose proprietary model weights. It needs to expose the goals.
Visible gates. When a recommendation system down-ranks a track, the artist should be able to learn that it happened, and ideally why. This is the digital equivalent of the phone call a musician could once make to a radio station. It will not be perfect — decisions are not always articulable in plain language — the current state, in which the algorithm's reasoning is wholly opaque to the people most affected by it, is not acceptable.
Fair compensation for the periphery. The long tail of music is disproportionately underpaid relative to its cultural weight, and the recommendation system is a significant driver of that imbalance. A serious ethics of curation has to include a serious economics. If an output is recommending generated music alongside the regional styles it learned from, the communities that produced those styles should be compensated in a way that scales with the model's success. This is not a solved problem, and it is not one the platforms will solve on their own. It is a regulatory question.
A place for human judgment. The argument is not that AI should be removed from curation. It is that human curators should remain in the loop at the points that matter most: the discovery of genuinely new music, the support of regional scenes, the recognition of artists whose importance is not yet reflected in their streaming numbers. Some platforms have begun to invest in this. The economics are harder than they look, because human curation does not scale, and because its results are harder to A/B test. It is, however, the only mechanism we have for the kinds of editorial decisions that do not reduce to engagement.
The industry does not yet have a settled position on any of these. There are working groups, white papers, and a small number of lawsuits. There is not yet a norm. This is the moment when the norm gets made. The companies that build curation systems, the artists whose work those systems sort, the regulators who will eventually be asked to constrain them, and the listeners who live inside the resulting playlists are all, at present, participating in an experiment whose protocol is still being written. It is worth being careful. The playlists we are served shape the musicians we hear from. They shape, in the long run, the musicians who get to keep making music.