The Streaming Algorithm as Gatekeeper

In 2023, a music industry investigation revealed that several distributors had been uploading vast quantities of AI-generated music to Spotify under fake artist names, gaming the platform’s royalty-splitting mechanism. The scheme worked because Spotify’s payment structure distributes a pool of subscription revenue to rights holders proportional to streams — and streams can be generated by placing music in algorithmic playlists, which respond to early engagement signals in ways that can be manipulated. Spotify removed hundreds of thousands of tracks and issued statements about protecting human artists. The statement did not address the underlying structural issue.

The underlying structural issue is this: streaming platforms built their recommendation algorithms to maximize engagement without designing for the distinction between human-created and AI-generated content. That distinction was not relevant when the algorithms were built, and retrofitting it is harder than any press statement makes it sound.

How the Recommendation Engine Works

To understand the problem, you need to understand how Spotify’s Discover Weekly, Release Radar, and radio-style recommendation systems actually function.

These systems are collaborative filtering models: they build a representation of a user’s taste based on their listening history, identify users with similar taste vectors, and recommend tracks that similar users have engaged with positively. The definition of “positive engagement” is not just a listen — it includes whether you skipped, how long you listened, whether you added the track to a playlist, and whether you came back to it. These signals train a model of what you want to hear.

The model is about you and tracks. It has no intrinsic knowledge of who or what made the tracks. A track generated by Suno that matches a user’s taste vector will be recommended just as readily as a track recorded by a human musician, because the algorithm cannot tell the difference unless it is explicitly told (by metadata, by content analysis, or by some other external signal).

Spotify has argued that its Content ID-like detection systems can flag AI-generated tracks. This is partially true for tracks that closely match existing copyrighted recordings and partially false for novel AI-generated music. A Suno-generated track in the style of a specific artist may be generically similar but not close enough to trigger content matching. An entirely novel AI-generated track of ambient music for focus has no match anywhere.

The Playlist Economy

The specific harm to human musicians is most acute in the playlist economy — the system by which music reaches new listeners through algorithmic and editorial playlists.

Algorithmic playlists (Discover Weekly, the Radio features) are entirely algorithmic. Editorial playlists (Today’s Top Hits, specific mood or genre playlists) are curated by human editors but shaped by algorithmic performance data. Both rely on engagement signals that are agnostic about human versus AI origin.

AI-generated music for specific contexts — focus and concentration music, sleep music, ambient music for various activities — optimizes extremely well for engagement signals in those contexts. The music has been generated to fit a specific emotional/functional brief and has no commercial identity to maintain; it will never tour, so there is no tension between streaming optimization and artistic integrity. Several analysis firms have found that a substantial fraction of Spotify’s most-streamed ambient and focus music is AI-generated, a number that has grown rapidly since 2024.

This matters to human musicians in those genres because the playlist positions are finite. A sleep music playlist with 2 million followers has room for perhaps 100 tracks. If 70 of those slots are occupied by optimized AI-generated content, 70 fewer slots exist for human musicians making sleep music. The royalty pool is split between them; if AI-generated tracks collectively accumulate more streams, the per-stream royalty for human tracks declines further.

What the Labels Negotiated

The major labels — Universal, Sony, Warner — have had more success in negotiating protections than independent artists, largely because they have the contractual leverage to demand them.

In 2025-2026 contract renegotiations with streaming platforms, the major labels extracted commitments including: AI-generated content must be labeled as such in metadata; AI-generated content cannot be included in editorial playlists without explicit platform curator decision; royalty splits between human-created and AI-generated content will be separately accounted for. Whether these commitments are enforced is a different question from whether they exist.

Independent artists, who represent approximately 40% of music on streaming platforms and have no equivalent negotiating leverage, have not obtained these protections. The Distrokid and TuneCore distributors they use have adopted minimal AI disclosure requirements, but enforcement is essentially voluntary.

The Royalty Pool Math

There is a specific arithmetic problem baked into the streaming royalty model that AI generation makes much worse.

Spotify and most other major streaming services use what is called a “pro-rata” royalty model: the platform collects subscription revenue, deducts operating costs and platform fees, and distributes the remainder to rights holders in proportion to their share of total streams. If you account for 1% of all streams on Spotify in a given month, you receive 1% of the royalty pool.

The model was designed in a world where total stream counts were bounded by the supply of music humans could produce. AI generation removes that bound. When AI-generated tracks enter the stream count at scale — and given their low production cost, there is no economic constraint on how many can be produced — the denominator of the royalty equation grows without limit. Every AI-generated stream dilutes the per-stream value of every human-generated stream.

This is not a theoretical risk. The musicologist Will Page, formerly Spotify’s chief economist, estimated in 2025 that total stream counts on major platforms had grown approximately 35% between 2023 and 2025, while listener populations had grown approximately 12%. The gap is attributable, at least in part, to AI-generated content accumulating streams through algorithmic recommendation. The result: per-stream royalty rates have declined even as platform subscription revenue has grown. The music industry’s total royalty pool is larger. Its distribution has been diluted.

The alternative to pro-rata is “user-centric” royalty allocation: each user’s subscription fee is distributed to the artists they personally listened to, not to the platform-wide stream count. This model would be less vulnerable to AI stream dilution because it ties royalties to the relationships between specific listeners and specific artists. Deezer adopted user-centric payouts in France in 2023, and early data showed measurable redistribution toward independent artists and away from streaming-optimized bulk content. Spotify has been piloting a modified version. Adoption has been slow because the major labels, whose catalogs benefit from pro-rata’s popularity-weighting, have not been enthusiastic supporters of the change.

The Spatial Audio Lever

One dimension of the streaming competition that has not been fully litigated: spatial audio and immersive audio formats give human musicians a temporary quality differentiation over AI-generated content.

Apple Music’s Dolby Atmos rollout, Tidal’s spatial audio library, and equivalent offerings on Amazon Music have created a category of music experience where the spatial positioning of instruments, the sense of room and acoustic environment, and the three-dimensional presence of performers are aesthetically central. AI-generated music has not kept pace with this format. The reasons are partly technical — generating convincing spatial audio requires understanding of acoustic physics that current models don’t implement well — and partly economic: the additional production cost of spatial audio engineering means that the economic case for AI-generated ambient music is less compelling in this format.

Whether this gap is durable or a temporary technical lag is uncertain. The working assumption among music technology researchers is that AI spatial audio will improve substantially over the next two to three years. For now, it represents one of the few format-specific differentiators that human musicians have in the streaming context.

The Economic Model That Survived

The live music economy has been the structural survivor of AI disruption to music — not because AI cannot compose live music (it can be used to generate music for live AI performances, which exist) but because the value of live performance is social and situational in ways that recorded music is not.

Ticket prices for live music have increased substantially in real terms since 2022, and not only because of the Ticketmaster pricing controversies. There is genuine demand growth. The theory most consistently offered by market observers: as recorded music has become abundant and cheap (and its origin increasingly uncertain), the scarcity and authenticity of live human performance has become more valuable. You cannot generate a live concert. The sweat is real, the mistakes are real, the energy between performer and audience is real. In a world of infinite generated content, presence costs something.

This creates an uncomfortable implication: the economic future of professional musicianship may run more strongly through touring and live performance than through recording, reversing the 20th-century economic structure of the music business. The recording was how musicians built the audiences that supported touring. In the AI era, the relationship may invert: touring — local performances, intimate events, direct audience relationships — may become the primary economic activity, with recorded music as the calling card rather than the product.

The streaming algorithm was not designed to be neutral between human and AI creativity. It was designed to maximize engagement, which is also not a neutral goal. When the most engagement-optimized content is AI-generated content that has no artistic integrity to protect and no touring schedule to maintain, the algorithm will serve that content. That is not a design failure; it is the design working exactly as intended, in a context that has changed around it.