Deezer Says 44 Percent of Daily Song Uploads Are AI-Generated — And the Music Platform Economy Is Already Broken

Forty-four percent. That is the number Deezer quietly dropped into a regulatory filing last week, and it is the number that should end any remaining debate about whether AI-generated music is a “future problem” for the streaming industry. It is not a future problem. It is today’s problem, it is yesterday’s problem, and by the time you finish reading this sentence, roughly another two hundred AI-generated tracks will have been uploaded to a major streaming platform somewhere in the world. Deezer, the French streaming service that has historically punched above its weight in data transparency, disclosed that forty-four percent of all songs uploaded to its platform on a daily basis are now AI-generated. Not “suspected of being AI-generated.” Not “flagged by our machine learning classifier as potentially synthetic.” Detected, confirmed, and logged. Forty-four percent. Every single day.

I have covered streaming economics for eight years, starting back when the debate was about whether Spotify’s per-stream payout of roughly zero-point-zero-zero-four dollars was sustainable for independent artists. We argued about playlist payola. We argued about the major labels’ equity stakes in Spotify and whether that distorted royalty negotiations. We argued about the DSP (digital service provider) model versus the user-centric payment model. Those debates feel quaint now. When I talked to a Deezer data engineer last month at a side event during a Paris music tech conference, he described the situation in terms that would have been unthinkable five years ago. “We are not a music platform anymore,” he told me, over a glass of something stronger than coffee. “We are a content moderation platform that happens to stream music.” His team, he explained, spends more engineering hours on AI detection, volume throttling, and royalty fraud investigation than on actual music discovery features.

This is the reality of the streaming music economy in April 2026. And the thesis of this piece is straightforward, if uncomfortable: the royalty payout system that has defined the last fifteen years of recorded music is already broken. Not bending. Not stressed. Broken. Within eighteen months, one of two things will happen. Either streaming platforms will pivot hard to a “verified human” tier with premium payouts, essentially creating a two-class system where AI-generated content is relegated to ambient background filler while human-created music commands higher per-stream rates, or independent musicians will stop earning meaningful income from streaming entirely, and the streaming platform will complete its transformation from a music business into a generic content utility. There is no middle ground here. The math does not permit one.

The Forty-Four Percent Number And Why It Matters More Than You Think

Let me unpack the Deezer disclosure carefully, because the headline number actually understates the problem. Forty-four percent refers to songs uploaded daily. Deezer’s total catalog, like Spotify’s, is dominated by a long tail of older human-created recordings, so the stock number of AI songs on the platform is still a minority share. But the flow number is what matters for platform economics, because every new upload competes for algorithmic surfacing, playlist placement, and recommendation slots. When forty-four percent of the daily inflow is synthetic, the composition of what gets recommended to users changes within weeks, not years.

For context, Spotify’s publicly disclosed catalog sits at approximately one hundred million tracks as of early 2026, up from roughly eighty-two million in 2023. That growth rate, which used to track new human artist signups and label releases, is now dominated by synthetic uploads. Back-of-envelope math: if Spotify is receiving uploads at roughly one hundred twenty thousand tracks per day (the widely cited figure from late 2023 was one hundred thousand, and the trend has been upward), and if we assume Spotify’s AI-generated share is somewhere in the ballpark of Deezer’s forty-four percent, that means more than fifty thousand AI-generated tracks are joining Spotify’s catalog every single day. Every single day. That is roughly one-point-five million new AI tracks per month, or eighteen million per year, on a single platform.

The Deezer engineer I spoke with was blunt about how this happened. Suno, the Cambridge-based AI music startup, launched its v3 model in March 2024 and its v4 model roughly a year later. Udio, the other major generator, launched in April 2024 and has iterated aggressively since. Both services offer essentially unlimited generation for subscription fees in the ten-to-thirty dollars per month range. A moderately diligent user on a twenty-four dollar monthly plan can generate several hundred usable tracks per month. A sophisticated user running API calls at scale can generate thousands. And the quality, which was laughably bad in 2023 and still noticeably off in 2024, is now good enough that casual listeners genuinely cannot distinguish Suno v4 output from low-to-mid-tier human productions in lo-fi, ambient, synthwave, and corporate background music genres.

The Upload Economics

Here is the part that broke the system. A distributor like DistroKid, TuneCore, or CD Baby will upload your tracks to every major streaming platform for a flat annual fee, typically twenty to fifty dollars for unlimited uploads. If you generate one thousand AI tracks using Suno, spend a couple hundred dollars on distribution, and upload them all, you have created a thousand potential royalty-earning assets at a total cost of maybe five hundred dollars. If even a small fraction of those tracks get organic plays, or if you pad streams using bot traffic (a separate and worsening problem), the return on investment can be dramatic. There have been multiple documented cases in 2025 of individuals generating six-figure annual royalty income from catalogs consisting almost entirely of synthetic output.

The most notorious case was the federal indictment of Michael Smith in September 2024, a North Carolina musician charged with fraud for allegedly generating ten million dollars in royalties using hundreds of thousands of AI tracks and a botnet of streaming accounts. That case became the template. What followed was not a deterrence effect; it was a scaling effect. Operators studied the Smith indictment, identified the detection patterns, and adjusted their operations to avoid the specific tripwires that had flagged him. The bot networks got more sophisticated. The track generation patterns got more varied. The royalty extraction continued.

How AI Music Detection Actually Works (And Why It’s Losing)

I spent a fair amount of time over the last six months talking to engineers at three of the major streaming platforms and at two specialized AI audio forensics companies. What they told me, off the record in most cases, was consistent: detection is a cat-and-mouse game that the mice are currently winning.

The dominant detection approach uses a combination of spectral analysis, generative artifact detection, and metadata correlation. On the spectral side, early Suno and Udio outputs had telltale frequency-domain signatures, particularly in the high-frequency transient regions around cymbals and sibilants, where generative models tend to produce subtly different statistical distributions than real recordings. Classifiers trained on known-AI versus known-human datasets can pick up these signatures with reasonable accuracy on raw model output. The problem is that nobody uploads raw model output. The standard workflow is to run the AI output through a mastering chain, add some analog warmth simulation, apply compression and EQ, and sometimes even print it to tape and recapture it. Each of these post-processing steps degrades the classifier’s accuracy. A 2025 benchmark from the International Society for Music Information Retrieval showed that detector accuracy dropped from ninety-four percent on raw Suno v3 output to about sixty-one percent after a standard mastering chain was applied.

Sixty-one percent is not a detection tool. Sixty-one percent is a coin flip with a slight hand on the scale. And if you are a platform processing fifty thousand AI uploads per day and your detector runs at sixty-one percent accuracy, you are either letting twenty thousand tracks slip through as false negatives, or you are incorrectly flagging a large number of legitimate human uploads as synthetic, which creates its own legal and reputational nightmare.

The Metadata Signal

Because direct audio detection is struggling, platforms have leaned heavily on metadata signals. Suspicious upload patterns, account creation history, distributor relationships, playlist placement patterns, and stream source distributions can all indicate AI-generated content at the account level even when individual tracks are hard to classify. Deezer, to its credit, has been more aggressive here than its competitors, reportedly shutting down or deranking entire accounts based on behavioral patterns rather than trying to classify each track.

But metadata signals have their own problems. They punish legitimate high-volume artists, they discriminate against certain distributor partnerships, and they can be gamed by operators willing to invest in account warming and behavioral randomization. The operators have the same access to the platform’s public APIs and documentation that the platforms do, and they iterate faster because they are not constrained by legal review, user experience considerations, or consistent policy application.

The RIAA v. Suno Lawsuit And The Legal Status Of Training Data

Any serious discussion of AI music economics has to address the underlying legal question: is the training process itself lawful? The Recording Industry Association of America, joined by the major labels, filed suit against Suno in June 2024 and against Udio shortly thereafter, alleging massive copyright infringement in the training of their models. The discovery phase has been fascinating. Suno’s response effectively conceded that its training data included copyrighted recordings, but argued that training constitutes fair use under United States copyright law, drawing parallels to the cases allowing text training for large language models.

As of April 2026, the Suno case is in its second round of summary judgment motions. A decision is expected by late 2026 or early 2027. If Suno loses, the damages could be existential. Statutory damages for willful infringement can reach one hundred fifty thousand dollars per work, and the labels have alleged infringement of millions of recordings. Even a conservative damages calculation produces numbers that exceed Suno’s valuation by several orders of magnitude. If Suno wins, the floodgates open further, because the fair use precedent would extend the generative model playbook from text to music with judicial blessing.

Importantly, the outcome of the lawsuit affects upstream economics but may not meaningfully change downstream platform economics. Even if Suno and Udio are forced to license training data, cheaper open-source music generation models have already proliferated, and the cost of generating a track will not meaningfully rise. The genie is not going back in the bottle through copyright law alone.

Platform-By-Platform Comparison

Here is how the five major streaming platforms currently stand on AI policy, as of April 2026, based on my reporting and their publicly stated positions.

The pattern is clear. Deezer has been the most transparent and the most aggressive. Spotify has been the largest and the most opaque. Apple Music has effectively opted out of the volume game by maintaining higher per-stream rates and a more curated catalog, an approach that looks increasingly prescient. Tidal and YouTube Music occupy specialized niches. None of them, as of today, have a comprehensive solution.

The Royalty Math Is Already Collapsing

Let me walk through the numbers that show why this system cannot continue. The pro-rata royalty model, which is used by Spotify, Apple Music, and Deezer’s standard tier, works like this: the platform collects subscription revenue and advertising revenue, takes its cut (roughly thirty percent), and distributes the remaining approximately seventy percent into a royalty pool. That pool is then divided among rights holders in proportion to each rights holder’s share of total streams on the platform during that period.

The math can be expressed simply:

per_track_royalty = (total_royalty_pool * track_streams) / total_platform_streams

The critical variable is the denominator. When total platform streams grow faster than the royalty pool, per-stream payouts fall. This is exactly what has been happening. Spotify’s revenue has grown at roughly fifteen percent annually; its stream count has grown at a much faster rate, especially as AI-generated tracks accumulate streams, whether from genuine casual listening, algorithmic recommendation, or outright fraud. The per-stream rate, which averaged around four-tenths of a cent in 2018, has drifted downward to approximately three-tenths of a cent in 2025, with significant variation by region.

Now layer the AI flood on top. If AI-generated tracks capture twenty percent of total platform streams (a conservative estimate given the forty-four percent daily upload share), then twenty percent of the royalty pool flows to AI rights holders rather than to human artists. For an independent musician pulling one hundred thousand streams per month on Spotify, the difference between three-tenths of a cent per stream and two-and-a-half-tenths of a cent per stream is the difference between three hundred dollars and two hundred fifty dollars in monthly income. Multiply that across hundreds of thousands of working independent artists, and the aggregate transfer from human creators to AI operators runs into hundreds of millions of dollars per year.

The Apple Music Premium

Apple Music’s per-stream rate, which the company publicly disclosed in 2021 at one cent per stream (more recent estimates place it in the seven-tenths to one cent range depending on territory), is notably higher than Spotify’s. This is partly because Apple Music has higher average revenue per user (no free tier, higher subscription price in some markets) and partly because Apple’s catalog has historically been less saturated with synthetic content. The gap is widening. In 2025, I spoke with several independent artists who reported Apple Music royalty income matching or exceeding their Spotify royalty income despite Spotify having three to four times their Apple Music stream count. That is a ratio that tells you everything about where platform economics are headed.

Spotify AI Playlists And The Platform’s Own Role

It would be convenient to frame this as purely an external problem, a flood of third-party AI content that platforms are heroically struggling to contain. But the platforms themselves are not innocent bystanders. Spotify has been quietly running AI-generated playlists and, according to multiple reports, including investigations by Harper’s and Music Business Worldwide dating back to 2022, has actively seeded playlists with music from production music companies that use ghost artist names. The practice, which Spotify has never fully acknowledged but also never meaningfully denied, involves using catalog content created specifically for playlist placement, often at lower royalty rates than would be paid to independent artists.

The line between “production music with ghost artist names” and “AI-generated music” is blurring to the point of invisibility. The economic incentive for Spotify to populate mood-based playlists (lo-fi chill, ambient focus, spa relaxation, workout energy) with low-cost, low-royalty content is enormous. If AI generation can produce that content at near-zero marginal cost, and if the platform itself controls placement, then every stream that flows to AI content is a stream that does not flow out as a royalty at the standard rate. This is not a theoretical concern. Internal financial pressure on Spotify from its public market valuation, combined with the margin expansion pressure from its podcast investments failing to produce profits, has created exactly the institutional incentive to lean into low-cost content.

The Billboard Charts And Cultural Infrastructure

The Billboard charts, which have been the cultural scoreboard for recorded music for nearly seventy years, now face their own existential question. Billboard integrates streaming data into its Hot 100 and other charts using a weighting formula that treats paid streams, free streams, and on-demand streams differently. The formula has been adjusted multiple times to deal with streaming fraud, but it has not been adjusted to deal with AI-generated content at scale.

In late 2025, an AI-generated track briefly entered the Billboard Hot Country Songs chart before being removed after media attention. The track was an unambiguous Suno generation, distributed through a standard aggregator, accumulating streams through genuine (if algorithmically driven) listener engagement. Billboard’s removal decision was controversial; the track had, by any mechanical measure, performed well enough to chart. The decision came down to a policy judgment that AI-generated content should not be eligible, but that policy was (and remains) under-specified. What is the rule for AI-assisted content? What about tracks where a human wrote the melody but used AI for vocal synthesis? What about tracks where AI was used for instrumental backing but vocals are human?

The Grammys faced the same question and issued guidance in 2024 requiring “meaningful human authorship” for eligibility, with carve-outs for various forms of AI assistance. The guidance is workable for the Grammys because the Grammys are, at heart, a curation and voting body. It is not workable for the streaming charts because the streaming charts are supposed to reflect listener behavior, and listener behavior now includes substantial engagement with AI content.

The Indie Musician Income Collapse

I want to spend some time on what this looks like for working musicians, because the abstract policy debates can obscure how sharp the human impact has been. Over the last year I have interviewed more than two dozen independent musicians across genres and career stages. The pattern is consistent and grim.

A mid-career indie rock artist I spoke with, who asked to remain anonymous because of ongoing label negotiations, told me that her streaming income had fallen by approximately forty percent between 2023 and 2025, despite her monthly listener count on Spotify remaining roughly flat. Her analysis, which matched what several other artists told me, was that algorithmic discovery had shifted away from mid-tail human artists toward either superstar content (where majors have enough promotional leverage to force placement) or ambient AI content (which is cheap, algorithmically agreeable, and faces no label-side advocacy against deranking).

A producer who specializes in lo-fi beats, a genre that has been particularly hard-hit by AI generation, told me that the stream counts on his back catalog had dropped by nearly seventy percent as AI-generated lo-fi flooded the same playlist ecosystem. His income from streaming had effectively ceased to be meaningful. He had pivoted to sync licensing (placing music in video games, ads, and video content) and Patreon-style direct fan support. Neither of those channels is immune to AI competition, but both offer more direct human-to-human economic relationships that are harder to replicate algorithmically.

The aggregate picture, which I built from conversations with several artist distribution companies and a review of publicly available data from royalty services, is that the median independent artist’s streaming income fell somewhere between thirty and fifty percent between 2023 and 2025. The top of the distribution (superstar artists with major label backing) was largely unaffected. The bottom of the distribution (hobbyists and early-career artists) was wiped out. The middle, which is where most working musicians live, took the heaviest proportional damage.

The Verified Human Tier: What It Would Look Like

If I am right that the streaming model is headed toward a bifurcation, the “verified human” tier is worth thinking about concretely. Here is what I believe the architecture will look like, based on conversations with platform engineers and a reading of recent patent filings from Deezer and Tidal.

Verified human content would be authenticated at the upload source through a combination of identity verification, session recording metadata, and cryptographic attestation from supported digital audio workstations. Think of it as C2PA for music: a provenance chain that accompanies a track from creation through distribution, attesting that human hands drove the creative decisions. The attestation would not claim that no AI tools were used (AI-assisted mastering, AI-enhanced vocal tuning, AI-powered mixing are all ubiquitous and benign), but would claim that the compositional and performative decisions were human.

Verified human content would then be eligible for a higher royalty tier, perhaps two to three times the rate of unverified content. It would receive preferential placement in editorial playlists. It would be the default in recommendation algorithms for users who opt in. Non-verified content would continue to exist on the platform (for AI content, user-generated content, field recordings, older catalog lacking modern provenance chains) but would be paid from a separate, lower-rate pool.

The Problems With This Model

The verified human tier is not a clean solution. It creates surveillance infrastructure that many working musicians will find objectionable. It privileges artists with access to specific DAW software and the technical sophistication to manage provenance chains. It raises complex questions about collaborative creation, assistive AI for disabled musicians, and the handling of legacy catalog. It creates a new category of fraud (attestation fraud) that operators will immediately attack. And it requires industry-wide coordination to be effective, because a verified tier on Spotify that is not matched on Apple Music or Deezer will fragment the economics further.

But the alternative, which is to continue the current model while trying to scale detection against an adversarial optimization problem that is provably unsolvable at scale, is worse. Something like the verified tier is coming, because the economic pressure on platforms to find a way to pay human artists enough to keep human music being produced is building toward a breaking point.

Suno And Udio: The Companies Behind The Flood

It is worth being precise about who is producing this tsunami of synthetic content. Suno, founded in 2022 and based in Cambridge, Massachusetts, raised a one-hundred-twenty-five million dollar Series B in 2024 at a valuation reportedly approaching five hundred million dollars. The company’s stated mission is to democratize music creation. Its pricing tiers, as of April 2026, offer a free tier with limited daily generations, a Pro tier at around ten dollars per month with substantially higher limits and commercial use rights, and a Premier tier at thirty dollars per month with the highest limits. API access is available for enterprise customers.

Udio, founded in 2023 and backed by a coalition of investors including Andreessen Horowitz, launched in April 2024 with a focus on higher audio quality and longer-form compositions. Udio has been particularly popular for generating tracks in specific genre styles with recognizable structural conventions, which has made it the tool of choice for operators targeting specific playlist categories.

Neither company, in public statements, has taken responsibility for the streaming fraud enabled by their tools. Both companies argue, with some justification, that they provide general-purpose creative tools and that the responsibility for platform compliance lies with uploaders and distributors. This is the standard software company deflection. It is also, in legal terms, approximately true. But in practical terms, the scaling dynamics of their business model depend on power users who generate large volumes, and a meaningful fraction of those power users are monetizing through streaming arbitrage.

The Training Data Question, Revisited

One dimension that deserves more attention than it typically gets is the training data pipeline. Both Suno and Udio, according to the RIAA complaint and subsequent discovery, trained on large corpora of copyrighted recordings scraped from various sources, potentially including YouTube audio, torrented collections, and licensed databases used in ways that exceeded license terms. The legal question is whether this training constitutes fair use. The ethical question is whether it is acceptable to build a system that extracts value from human creators without compensating them, and then uses that system to compete against those same creators for scarce royalty pool dollars.

The ethical answer seems obvious. The legal answer is uncertain. And the economic answer is that this pattern (extraction, transformation, competition) is the same pattern that has defined AI’s entry into essentially every creative field, from illustration to writing to voice acting. Music is not special in facing this dynamic. Music is special in that its economic infrastructure, the streaming platform, is structurally vulnerable to being flooded in a way that, say, the illustration market (which uses more varied distribution channels) is not.

Regulatory Responses: What Governments Are Doing

The regulatory picture is fragmented. The European Union’s AI Act, which entered into force in August 2024 with full applicability phased through 2026, includes transparency requirements for AI-generated content, including music, but does not specifically address streaming royalty implications. France has been the most aggressive national regulator, with the CSA (now Arcom) requiring streaming platforms operating in France to disclose AI content ratios and maintain minimum human content quotas in certain categories. This French policy, which took effect in early 2026, is part of why Deezer (a French company) has been the most transparent about its AI upload data.

In the United States, regulatory action has been minimal. The Copyright Office has issued guidance clarifying that purely AI-generated works without human authorship are not eligible for copyright registration, which has some downstream effects on rights management but does not directly address the streaming royalty flow. The NO FAKES Act, introduced in Congress in 2023 and reintroduced in various forms since, addresses voice and likeness rights but does not address the general AI music flood. The ELVIS Act in Tennessee protects against unauthorized voice cloning but is narrowly scoped.

The regulatory gap is significant. Streaming platforms are operating in an environment where their business model is being destabilized by AI content but where no regulatory framework exists to mandate or even guide a response. Platforms that try to act unilaterally face antitrust concerns (if they coordinate with other platforms) and legal challenge (if they deplatform accounts without clear policy justification). The absence of regulation has, paradoxically, paralyzed platform action.

What A Functioning Market Would Look Like

Let me sketch what I think a functioning streaming market would look like in a world where AI generation is an unavoidable reality. The features would include:

• Upload-side provenance requirements with cryptographic attestation of human authorship for verified content
• Tiered royalty pools with substantially higher per-stream rates for verified human content, funded by differential allocation rather than total pool increase
• User-facing labeling that clearly indicates AI-generated content and lets users opt in or out at the algorithm level
• Volume throttling at the account level with transparent criteria that apply symmetrically to human and AI uploads
• Stream-source verification that excludes demonstrably fraudulent streams (bots, manipulation) from royalty calculation regardless of content type
• Coordinated industry standards for provenance, detection, and disclosure, ideally developed through an industry body rather than platform-by-platform
• Artist-side tools that let working musicians monitor their catalog performance against AI competition and make informed decisions about distribution

None of this is technically difficult. All of it is organizationally and commercially difficult. The major platforms have overlapping but not aligned incentives. Spotify, as the market leader with the largest existing AI exposure, has the most to lose from aggressive action and the least to gain from coordinated standards. Apple Music has the opposite position and has been moving somewhat in this direction. Deezer is the wild card: a smaller platform with a data-transparency culture and regulatory cover in France to act more aggressively.

Predictions For 2026-2028

Here is what I expect to happen over the next two years, with my confidence level in each prediction.

High confidence: Deezer launches a verified human tier in the second half of 2026. The company has been signaling this publicly and privately for months, and the French regulatory environment makes it viable. The launch will feature provenance attestation, differential royalty rates, and explicit user-facing AI labeling. Initial artist uptake will be strong among French and European independent artists.

High confidence: Spotify introduces AI content labeling but resists tiered royalties through 2026. Spotify will implement labeling to comply with EU AI Act requirements and to preempt regulatory action in other jurisdictions, but will resist differential royalties because its catalog composition makes differential payouts financially risky. Expect loud artist-side advocacy and at least one high-profile artist withdrawal protest.

Medium-high confidence: The RIAA v. Suno case produces a mixed verdict by early 2027. Suno will likely prevail on the core fair use question for training, consistent with the emerging pattern in text AI cases, but will face liability for specific outputs that reproduce copyrighted works verbatim. Damages will be significant but not existential. Suno and Udio will negotiate settlement structures with the major labels that include licensing pools funded by the AI companies.

Medium confidence: Spotify’s per-stream rate drops below two-tenths of a cent for independent artists on standard tiers by late 2027. Continued stream volume inflation combined with slower revenue growth will compress per-stream payouts further. Major label artists, whose contracts include guaranteed minimums, will be shielded. Independent artists will bear the full compression.

Medium confidence: Apple Music emerges as the dominant platform for working musicians’ income by 2028. The combination of higher per-stream rates, less saturated catalog, and more curated discovery will make Apple Music the preferred platform for serious music consumers and the primary income source for working artists. Spotify will retain the casual listener market but lose its position as the music industry’s primary commerce infrastructure.

Medium confidence: A major direct-to-fan platform emerges as a serious third pillar. Bandcamp, which has been battered by ownership changes but retains strong artist loyalty, or a new entrant, will capture meaningful market share for direct artist-to-fan economic relationships. The model will not be streaming-based; it will be a hybrid of purchase, subscription, and community features. AI content will be largely excluded not by policy but by the lack of a fan relationship to monetize.

Lower confidence: At least one major streaming platform exits the general music business by 2028. YouTube Music or Amazon Music may pivot away from general music streaming toward podcast-plus-music bundles or audiobook-plus-music bundles, effectively treating music as a loss-leader feature rather than a core product. The economics of competing with Spotify’s catalog scale while also defending against AI flood may prove untenable for some players.

Lower confidence: A federal AI music bill passes in the United States by end of 2027. Political will for action exists on both sides of the aisle, but legislative capacity to draft workable policy is limited. A narrow bill addressing voice cloning and deepfake vocals is more likely than comprehensive AI music legislation. The broader royalty system reform will come from industry self-regulation rather than federal legislation.

What Artists Should Do Now

If you are a working musician reading this, the actionable implications are relatively clear, though none of them are easy. Diversify away from pure streaming dependence as fast as you can. Build direct fan relationships through newsletters, Patreon, Bandcamp, live performance, and merchandise. Invest in sync licensing as a revenue channel, recognizing that it too faces AI competition but remains more human-gatekept. Engage seriously with Apple Music optimization if it is not already a primary channel, because the per-stream economics there remain meaningfully better. Document your creative process aggressively, because provenance attestation is coming and the artists with clean provenance records will be advantaged in any verified human tier.

And do not assume that this is a temporary problem that will resolve itself. The forty-four percent number at Deezer is not going back down. The technological capability to generate convincing synthetic music will continue to improve. The economic incentives to flood streaming platforms will continue to compound. The platforms will respond, eventually, but they will respond slowly and imperfectly, and the cost of the adjustment period will fall disproportionately on working artists who are not equipped to absorb it.

The Broader Lesson For The Creator Economy

I want to close with a thought that goes beyond music. The pattern we are watching in streaming music, where a previously functional creator economy is being restructured by AI flood, is going to repeat in adjacent fields. Stock photography has already been hit. Illustration is in active restructuring. Stock video is next. Voice acting is under direct assault from AI voice synthesis. Writing, for certain categories of functional commercial writing, is already in a similar place.

The common pattern is: a platform economy intermediates between creators and consumers, enabling long-tail monetization at scale. AI generation capability enters the market, producing content at near-zero marginal cost. The platform’s infrastructure, designed for human-scale supply, cannot effectively filter. Royalty or payment pools are divided across a rapidly expanding denominator, compressing per-unit payouts. Working creators in the middle of the distribution are squeezed out. The top of the distribution is relatively unaffected, the bottom is wiped out.

The lesson for anyone in a creator economy role, from musicians to writers to designers to filmmakers, is that the platform intermediation model may not survive its encounter with generative AI in its current form. What replaces it will likely be a combination of more expensive curation layers, more direct fan relationships, and more explicit labor-protective infrastructure (verified human tiers, AI transparency requirements, provenance attestation). The transition will be messy and will impose real costs on working creators. But the underlying market for human creative work is not going away. Humans want to listen to music made by humans. Humans want to read writing done by humans. Humans want to watch films made by humans. The question is whether the economic infrastructure can reform fast enough to keep those human creators economically viable during the transition.

Deezer’s forty-four percent disclosure was, in that context, a gift to the industry. It was a hard number, publicly stated, that made it impossible to continue pretending that AI music was a future concern. Forty-four percent is not a future concern. It is today. It is tomorrow. And unless the industry acts with urgency that it has not yet demonstrated, it will, within a remarkably short time, be everything.

We are not a music platform anymore. We are a content moderation platform that happens to stream music. The question now is whether we can still pay the musicians who made us a music platform in the first place, or whether that era is already over and we are just the last people to notice.

That was the Deezer engineer, at the end of our conversation, after the second glass. I have thought about that quote often in the months since. I think he is right about the first part. I think the second part, the question of whether the era is already over, is the question that the next eighteen months will answer. I am not optimistic. But I am not without hope, either. The infrastructure exists to build something better. The artists exist to populate it. The listeners exist to support it. What is missing is the institutional willingness to act before the current system has fully collapsed rather than waiting until after. Whether that willingness materializes in time is the only question that matters.