$8M in Fake Streams, 18 Months in Prison — You Call That a Deterrent?
Summary
AI music streaming fraud has emerged as a structural crisis for the music industry, with AI-generated tracks now comprising over 50% of new daily uploads on Deezer — roughly 90,000 tracks per day — while accounting for only 1–3% of actual streams, an extreme divergence that suggests much of that upload volume exists to feed bot-driven royalty extraction rather than real listeners. In 2025, up to 85% of streams from fully AI-generated tracks were flagged as fraudulent and blocked from monetization, revealing how deeply fraud has saturated the AI music ecosystem on at least one major platform. The case of Michael Smith — who defrauded streaming platforms of $8,091,843.64 over seven years and received just 18 months in prison — was designated by the U.S. Department of Justice as the "first criminally charged AI-assisted music streaming fraud," yet the sentence falls well short of the 46-month minimum requested by prosecutors, raising fundamental questions about deterrence. Music industry manipulation has an unbroken history stretching from the payola scandals of the 1950s to today's AI bot schemes, but AI did not invent a new category of fraud — it collapsed the marginal cost of existing fraud to near zero, eliminating the economic barriers that once limited entry. While platform detection systems, AI-labeling initiatives, and industry coalitions like the Music Fights Fraud Alliance provide real but structurally constrained defenses, the asymmetry between near-zero fraud entry costs and the sustained expense of detection suggests that no single intervention will resolve this crisis.
Key Points
The Extreme Gap Between AI Upload Volume and Actual Listening
Deezer's newsroom data published in July 2026 shows that more than 50% of new tracks uploaded to the platform every single day are AI-generated, averaging roughly 90,000 tracks daily. TechCrunch mapped the growth: AI uploads accounted for about 10% of daily additions in January 2025, climbed to 18% by April, 28% by September, 34% by November, and 39% by January 2026, before crossing 50% in June 2026 — a fivefold increase in roughly 18 months. Despite representing more than half of all daily uploads, these AI-generated tracks account for only 1–3% of actual streams on the platform, an extreme divergence that reveals the dominant motive behind mass AI uploading as royalty extraction via bot networks rather than any genuine attempt to reach listeners. In 2025 alone, Deezer detected and tagged over 13.4 million AI-generated tracks — a figure that describes the scale of industrial-volume content injection on a single platform over a single year. This gap is not a minor anomaly but a structural fingerprint: when upload share exceeds 50% while listen share stays below 3%, the content is not being uploaded for people to discover, it is being uploaded for bots to play. Data on comparable scale from Spotify, Apple Music, YouTube Music, and Amazon Music has not been publicly released, which means the full scope of AI-generated track infiltration across the global streaming ecosystem remains substantially unknown.
The First Criminal Case and the Limits of Deterrence
Michael Smith ran his scheme from 2017 to 2024 — roughly seven years — using AI to generate hundreds of thousands of songs, distributing them across Spotify, Apple Music, Amazon Music, YouTube Music, and other platforms, and operating up to 10,000 bot accounts simultaneously to generate billions of fraudulent streams. In April 2023 alone, his bots logged more than 80.9 million plays on YouTube Music's family plan tier — just one month on one subscription tier, suggesting the full operation was substantially larger. Judge John G. Koeltl of the Southern District of New York sentenced him on a single count of wire fraud conspiracy: 18 months' imprisonment, 2 years' probation, and full forfeiture of the $8,091,843.64 he extracted. The U.S. Department of Justice labeled this the "first criminally charged AI-assisted music streaming fraud," making the resulting sentence symbolically significant as the benchmark the legal system has established for this category of crime. Prosecutors had requested at least 46 months — the low end of federal sentencing guidelines — but the judge chose less than half of that, reportedly suggesting the guidelines were somewhat excessive. When the annual yield of a fraud scheme exceeds $1 million and the worst-case criminal outcome is 18 months, the deterrence value of that punishment is fundamentally limited relative to the expected return.
Seventy Years of Music Manipulation and What AI Changed
Music industry manipulation has an unbroken lineage stretching back to at least 1959, when U.S. Congressional payola hearings exposed radio DJs accepting cash to play specific songs — a Chicago DJ named Phil Lind testified to receiving $22,000 for a single broadcast placement, and the more prominent Alan Freed was fired and later convicted of commercial bribery in 1962. The medium shifted from radio to digital streaming, but the underlying logic remained: money follows influence over what gets heard, and wherever there's a financial distribution mechanism, someone will attempt to manipulate the inputs. The 2018 Bulgaria case illustrated the pre-AI ceiling: one person manually managing roughly 1,200 Spotify accounts to loop 467 songs for an estimated $1 million required sustained human labor, which acted as a natural entry barrier. What Michael Smith did revealed what happens when that barrier disappears: AI-generated music at scale plus cloud-provisioned bot infrastructure reduced the operational cost to near zero, transforming what had been a labor-limited cottage fraud into something that one person could run at a scale that overwhelmed manual detection. AI did not create a new type of fraud — it eliminated the cost friction that had always constrained the old type, and that is a structurally different problem from anything the industry has previously encountered.
Platform Detection and the Structural Disadvantage of Defense
Platform responses have been faster and more substantive than many expected. Deezer has excluded AI-generated tracks from algorithmic recommendations and editorial playlists, blocked royalty monetization for detected fraud streams, and is planning a user-facing AI music filter toggle before the end of 2026. Spotify operates AI-assisted track labeling under the DDEX standard and removes unauthorized voice clones. Apple Music introduced a "Made With AI" label in early 2026, though the limitation of relying on distributor self-disclosure is real and significant. Tidal withholds royalties entirely from 100% AI-generated tracks, and Bandcamp prohibits AI-led music outright. The fact that Deezer detected and blocked up to 85% of streams from fully AI-generated tracks as fraudulent in 2025 is genuine evidence that defensive infrastructure is working at a meaningful level. However, an Ipsos blind test involving 9,000 participants across eight countries found that 97% could not distinguish AI music from human music — which means the detection problem the platforms are solving requires algorithmic precision that consistently exceeds human perceptual capability. The arms-race dynamic is structurally unfavorable for defenders: detection must succeed every time, while evasion only needs to succeed once per new technique, and AI evasion tools evolve alongside detection tools.
Creator Income at Risk and the Structural Economics of AI Music
CISAC's December 2024 research provides the clearest economic framework for the stakes: music creators' income is estimated to be at risk of losing up to 24% by 2028 from AI-related impacts, with potential annual losses reaching €40 billion and cumulative five-year losses estimated at €100 billion. At the same time, AI service companies' music-related revenues are projected to grow from roughly €100 million in 2023 to approximately €4 billion by 2028, while the generative AI music market overall could reach approximately €16 billion by 2028 — up from roughly €1 billion in 2023. These projections describe a structural economic shift: the music market grows, but the value distribution moves away from creators and toward technology companies and infrastructure platforms. Streaming fraud is not the sole driver of this shift, but it accelerates it directly — every dollar extracted through fake streams is a dollar removed from the legitimate royalty pool that would otherwise have flowed to artists. The fact that 80% of listeners in surveys say they want clear AI music labeling suggests that consumer awareness and concern is real, but Deezer's finding that only 45% actively want a filter reveals a gap between stated preference and behavioral pressure that leaves the system vulnerable. James Blake's Instagram statement in the summer of 2026 — "If you're an artist, remember that in 2026 there's not a single part of the system that isn't faked" — captures the lived reality that economic projections alone cannot fully convey.
Positive & Negative Analysis
Positive Aspects
- Platform Detection Systems Are Actually Working at Scale
Deezer's detection of up to 85% of fully AI-generated track streams as fraudulent in 2025 — and blocking them from monetization — is concrete evidence that the defensive infrastructure is functioning at a meaningful level, not just in theory. Without this detection capability, the amount of royalty pool money diverted to fraud would be substantially higher than it currently is, and the economic harm to legitimate artists would be correspondingly worse. Beyond Deezer, Spotify, Apple Music, Amazon Music, Tidal, and Bandcamp have all built out distinct AI detection and labeling mechanisms, creating a layered defensive environment that makes the entire streaming ecosystem harder to exploit than it would otherwise be. Tidal's decision to withhold royalties entirely from 100% AI-generated tracks represents the most direct possible economic intervention — it removes the financial incentive at the source rather than trying to filter output after the fact. The platforms' response has been faster and more coordinated than critics often acknowledge, and the detection investment represents genuine harm reduction even where it falls short of a complete solution.
- Cross-Platform Coalition Infrastructure Is Taking Shape
The Music Fights Fraud Alliance, launched in June 2023 and grown to more than 28 member organizations by late 2025, represents a structural advance that individual platform defenses cannot replicate on their own. Spotify, Amazon Music, DistroKid, Believe, EMPIRE, and others competing for listeners in the same market have nonetheless chosen to collaborate on fraud detection — that alignment of incentives across competitors is not trivial and signals how seriously the major players regard the threat. The coalition's core value proposition is cross-platform intelligence sharing: when a fraud pattern is identified on one service, it can be deployed defensively across the others, making it harder for a bad actor blocked in one place to simply migrate to another. Qobuz's separate AI charter and proprietary detection tools further extend the coverage of coordinated industry-level response. This architecture is still maturing, but its existence creates the foundation for a genuinely industry-wide defense that no single platform could build alone, and the pace of its member growth suggests the commitment to it is durable.
- The Legal Precedent Exists and Lowers Future Barriers to Prosecution
The 18-month sentence in Michael Smith's case is rightly criticized for its deterrence limitations, but the establishment of legal precedent deserves recognition as a distinct and meaningful outcome. The DOJ proving that wire fraud conspiracy applies to AI-assisted music streaming fraud, building a successful digital forensics case around bot traffic analysis and revenue tracing, and creating an evidentiary and prosecutorial template for future cases — these outcomes reduce the marginal cost and complexity of bringing subsequent charges. The DOJ's public framing of the case as the "first criminally charged AI-assisted music streaming fraud" also ensures that the category of offense is now formally defined, which matters for how future prosecutors, judges, and investigators approach similar schemes. Media coverage of the case has also meaningfully increased public awareness of AI streaming fraud's existence and scale — which creates modest but real reputational deterrence even where sentencing deterrence is weak. A bad actor now has to factor in that this category of fraud is federally prosecuted, forensically traceable, and publicly known, which is marginally different from the near-total impunity that existed before 2026.
- User-Level Control Is Moving in the Right Direction
Deezer's planned AI music filter toggle represents a philosophically important shift: rather than making fraud-prevention decisions entirely on behalf of users, the platform is extending user agency over what kind of content populates their experience. The fact that 45% of Deezer's users say they want to filter AI music means there is a real and substantial market segment that values human-created music enough to actively seek it out — and giving those users the tool to act on that preference is both commercially rational and structurally meaningful. If the toggle succeeds and gains adoption, it creates pressure on other platforms to offer equivalent features, potentially normalizing user-controlled AI content filtering across the streaming ecosystem. Surveys showing that 80% of listeners want clearer AI music labeling suggest the demand for transparency substantially exceeds current supply, which means platforms that move fastest on labeling and filtering infrastructure may build durable loyalty advantages. User choice, aggregated across a large enough listener base, can become an organic market mechanism that reduces the economic value of flooding platforms with AI content that informed consumers actively reject.
Concerns
- The Criminal Deterrence Is Structurally Insufficient Relative to Expected Returns
The Smith case laid out the core deterrence problem in explicit numbers: seven years, $8,091,843.64 extracted, 18 months sentenced. At roughly $1.16 million per year in fraudulent income, the worst-case criminal outcome after arrest is a year and a half — a risk-to-return ratio that does not constitute meaningful deterrence for anyone rationally calculating expected values in this space. The judge's decision to sentence below the guideline minimum, combined with the defense's request for no prison time at all, signals that the judiciary is still calibrating its understanding of AI-enabled fraud's systemic harm — and until that calibration catches up, the sentencing benchmark will remain too low to shift behavior. The low detection rate — Smith ran his scheme for approximately seven years before consequences arrived — compounds the deterrence failure: multiplying a low capture probability by a modest sentence produces an expected punishment that is dwarfed by the expected return. For the criminal deterrence dimension of this problem to function, either sentencing levels need to rise substantially or detection rates need to improve dramatically, and neither trajectory is currently on a trajectory fast enough to change the near-term incentive structure.
- The Arms Race Structurally Favors the Attacker
The finding that 97% of listeners in an Ipsos study across eight countries could not distinguish AI-generated music from human-created music in a blind test reveals the ceiling of detection through perceptual means — and raises genuine questions about whether algorithmic detection can consistently exceed human perceptual capability at the scale required. Detection technology evolves, but evasion technology evolves alongside it; that is the defining feature of every security arms race, from email spam filtering to network intrusion detection. The asymmetry is structurally unfavorable for defenders: they must correctly classify every fraudulent attempt, while attackers only need each new evasion technique to work until a new defensive countermeasure is developed. AI generation tools are becoming more capable at producing output that evades watermarking, spectral analysis, and pattern detection — and those evasion improvements arrive as a byproduct of general AI capability improvements rather than requiring dedicated adversarial investment. Additionally, the platforms carrying the detection burden spend real money on infrastructure and human review, while fraudsters operate at near-zero cost — a cost asymmetry that the defender cannot reverse through effort alone.
- Near-Zero Entry Cost Has Permanently Changed the Threat Landscape
The colloquial way to understand the structural change AI brought to music fraud is this: what once required $22,000 in payola payments, then required managing 1,200 manual accounts, now requires a laptop and a cloud subscription. AI tools for music generation are free or near-free, bot infrastructure is cheap and scalable, and the operational knowledge is widely available. This cost collapse is not reversible — AI tools only become more capable and more accessible over time, which means the marginal cost of entry will not rise back to the level where it functions as a natural deterrent. Historically, economic friction limited music fraud to a relatively small number of organized actors willing to invest significant resources. With that friction gone, the threat landscape shifts from a small number of large-scale operations — detectable, prosecutable, attributable — to a potentially large number of small-to-medium operations that individually fall below detection thresholds and collectively represent a diffuse, hard-to-address background level of fraud. Thousands of people each running smaller versions of Smith's scheme produce a collective problem that is substantially harder to investigate and prosecute than a single large operator.
- Fragmented Platform Responses Leave Systemic Gaps
The current state of platform-level defense is a patchwork: Deezer publishes detailed data and deploys multiple interventions; Bandcamp prohibits AI music entirely; Tidal withholds royalties; Apple Music labels but relies on self-disclosure; others apply varying degrees of detection and filtering. This fragmentation creates exploitable gaps. A fraudster blocked or restricted on one platform can migrate operations to another where policies are less stringent or detection is less mature — and without a universal cross-platform standard, the most aggressive defenses create a pressure gradient that pushes bad actors toward softer targets rather than eliminating them from the ecosystem. MFFA membership of 28-plus organizations is meaningful, but the global streaming and distribution ecosystem encompasses dozens more services and hundreds of distributors not yet integrated into the coalition's information-sharing architecture. For platform-level defense to function as a genuine systemic barrier rather than a game of Whack-a-Mole, the coverage needs to be comprehensive enough that migration between services ceases to be a viable fraud strategy — and that level of universality remains a long way from current reality.
Outlook
In the near term — the next six months or so — the most visible change will be a strengthening of platform-level defenses. Deezer's planned AI music filter toggle is the first real test: its stated goal is to let users opt out of 100% AI-generated music from services like Suno, Udio, and Riffusion, while deliberately keeping AI-assisted human creation outside the filter's scope. If this feature lands successfully and sees meaningful adoption, it creates competitive pressure for other platforms to follow with equivalent user-control mechanisms. Spotify and Apple Music's existing AI labeling systems — built on the DDEX standard and the "Made With AI" tag respectively — are likely to become more precise and harder to circumvent as distributor-level reporting requirements tighten. I expect to see the DDEX framework expand its requirements for AI origin disclosure at the point of delivery, moving the verification burden closer to the source rather than leaving it entirely to downstream detection. These short-term measures will have some real effect in reducing the marginal profitability of fraud, but they're not structural solutions — they're better-calibrated buckets in a room that's taking on water.
The legal and enforcement picture following Smith's sentencing is genuinely mixed, and I think it's important not to over- or understate its significance. On the positive side, the case established a legal precedent: wire fraud conspiracy applies to AI-assisted music streaming fraud, the digital forensics methodology for proving such a case has been road-tested, and future prosecutors have a template to work from. U.S. Attorney Jamie McDonald described Smith as having flooded platforms with automated bots posing as consumers and fake songs standing in for creativity — framing that creates a usable prosecutorial narrative for similar cases going forward. The practical threshold for bringing charges of this type is now meaningfully lower. But I won't oversell this. Smith evaded detection for seven of the roughly seven years his scheme ran. Cross-border investigations remain complex and expensive. And the most sophisticated operators today are likely far more distributed in both their music generation and bot infrastructure than Smith was. The legal deterrent will apply to a subset of large-scale operators who leave enough of a trail to follow — not to the broader ecosystem of smaller-scale fraud that is likely already operating below the detection threshold.
Over the medium term — the next one to two years — the critical variable is whether the Music Fights Fraud Alliance can scale its coordination into something that actually functions like a real-time cross-platform alert system. Growing from inception to more than 28 members in roughly two years is genuinely impressive, and it reflects how seriously the major players regard this threat. But there's a meaningful gap between having a coalition and having functional, real-time data sharing across competitors who have natural incentives to guard their detection algorithms and fraud pattern intelligence. The architectural challenge is significant: for MFFA to deliver its theoretical value, member platforms need to share suspicious actor profiles fast enough that a fraudster blocked on one service faces an immediate barrier on the others. Getting there requires trust, standardized data formats, legal frameworks for sharing, and ongoing technical investment across all participating parties. I think this infrastructure will get built — the incentives are strong enough — but the timeline for it to reach operational effectiveness is likely longer than optimists would project.
Simultaneously, the medium term brings what I think is the hardest definitional challenge in this entire space: how do you draw a coherent, enforceable line between legitimate AI-assisted human creation and fraudulent fully-automated content? A growing number of artists use AI tools in their composition, production, and mixing workflows. Regulating all AI-touched music the same way would be both technically unjustifiable and artistically counterproductive. Deezer's approach — targeting specifically 100% AI-generated output from dedicated AI music services — is a pragmatic attempt to navigate this, but it depends on those services accurately identifying their own output, and it doesn't cover the increasingly blurry middle ground where a human artist uses AI for 60% of a track's composition. As AI tools become more capable, the boundary between "AI-assisted" and "AI-generated" will continue to dissolve. Regulators haven't established a clear standard, and I think that ambiguity will function as a gray zone that sophisticated bad actors will deliberately inhabit. Getting the definitional framework right — technically, legally, and artistically — is probably the most underrated challenge in the entire fraud-prevention effort.
Looking further out — a two-to-five year horizon — the CISAC research provides the clearest framework for thinking about structural stakes. The generative AI music market is projected to grow from roughly €1 billion in 2023 to approximately €16 billion by 2028. That's a 16x increase in five years. The question isn't whether this market grows — it's who captures the value. CISAC's projections suggest AI service companies' music-related revenues could surge from roughly €100 million in 2023 to approximately €4 billion by 2028, while music creators' income faces an estimated 24% at-risk exposure over the same period — with cumulative potential losses reaching €100 billion over five years. If those projections are even partially accurate, they describe a substantial transfer of economic value from the people who create music to the platforms and technology companies that distribute and generate it. Streaming fraud accelerates this transfer by siphoning royalty pool money away from legitimate artists and routing it to bad actors — it's not the only mechanism of this shift, but it's one of the more direct ones, and it's already measurable.
There's a longer-term risk that I think doesn't get enough attention: the slow erosion of listener trust in streaming curation. If AI fraud tracks continue to infiltrate playlists at scale, and 97% of listeners genuinely can't tell the difference, the problem doesn't manifest as sudden audience revolt — it manifests as a gradual, invisible degradation of confidence in whether what you're hearing is real. Listeners who can't consciously identify fraud might still develop a vague, unplaceable sense that their feeds feel less authentic, less interesting, less connected to actual human expression. That's the kind of brand damage that shows up in engagement metrics years after the underlying cause. Platforms that can credibly guarantee human-curated, authenticated content — whether through blockchain-based creative provenance systems, certified artist programs, or simply better-enforced labeling — may develop a durable competitive advantage over those that don't. Bandcamp's categorical AI music ban represents the extreme version of this strategy; whether it accumulates a meaningfully large "human music purist" subscriber base is worth watching. If the market naturally separates into AI-inclusive and human-only tiers, that segmentation itself could become a mechanism for reducing fraud's economic value within the premium tier.
Here's how I see the range of outcomes across three scenarios. In the optimistic scenario, MFFA-centered cross-platform intelligence sharing reaches functional real-time operation, one or more major regulatory jurisdictions — most plausibly the EU under its AI Act framework — mandates content provenance disclosure for AI-generated music at the distributor level, and platform detection technology advances faster than evasion techniques. Under this scenario, the fraud rate stabilizes or begins declining from current levels, and the economic incentive structure shifts meaningfully against entry. For this to materialize, platform data-sharing agreements need to become genuinely operational rather than aspirational, and regulatory timelines need to accelerate beyond their current pace. In the base scenario, the arms race between detection and evasion continues with neither side establishing durable superiority, fraud persists at roughly current scale but doesn't dramatically accelerate, and the industry absorbs an ongoing defense cost that places downward pressure on artist royalties. This is the scenario I find most plausible given current trajectories. In the pessimistic scenario, AI generation tool sophistication outpaces detection infrastructure on a sustained basis, trust in streaming curation degrades visibly enough to affect platform engagement metrics, and the industry faces structural pressure to rethink its basic operating model. I don't think this scenario is likely in the near term, but the conditions for it are not implausible.
I want to be honest about where I could be wrong. If the major platforms decide to operate separate royalty pools for AI-generated and human-created music — fundamentally changing the economic incentive for fraud — the return on AI music bot schemes drops substantially. If technologies like audio watermarking or digital provenance verification mature faster than current timelines suggest, the structural asymmetry between fraud cost and detection cost could narrow significantly. And if the EU's AI Act or other regulatory frameworks extend specific, enforceable requirements to music distributors faster than expected, the legal overhead of operating fraud schemes could increase in ways that current projections don't account for. I think the realistic probability of these accelerating factors arriving fast enough to outpace the fraud is not high, based on current data — but I want to be clear that this is a judgment under genuine uncertainty, not a confident forecast.
For each group that has skin in this game: platforms should be investing in MFFA-style cross-platform data sharing as a priority over building better individual silos — the collective defense is structurally superior, and a fraudster blocked on one service needs to face an immediate barrier everywhere else. Artists should be monitoring their own streaming data regularly and reporting unusual play pattern spikes to platforms as soon as they notice them, both to protect their own royalty integrity and to contribute detection intelligence. Regulators need to recognize that criminal sentencing at current levels is not generating meaningful deterrence, and that the more tractable near-term interventions are detection infrastructure investment and distributor oversight requirements. And for listeners: the most direct act of resistance is intentional listening — finding the artists you care about and seeking them out deliberately, rather than relying entirely on algorithmic playlists. I know that sounds small against the scale of what's described here. But in a system where 97% of people can't tell real from fake, conscious engagement with specific human artists matters more than it ever has.
Sources / References
- AI Music Exceeds 50 Percent Daily Uploads on Deezer — Deezer Newsroom
- Music Streamer Deezer Says More Than 50% of Daily Uploads Are AI-Generated — TechCrunch
- Michael Smith Jailed 18 Months for $8M AI Streaming Scam — AI Weekly
- Music Streaming Fraudster Prison Sentence — Complex
- Is 2026 the Fakest Year in Music? — NPR / Hawaii Public Radio
- Global Economic Study Shows Human Creators' Future at Risk from Generative AI — CISAC
- A Quarter of Musician Revenue to Be Lost to AI by 2028 — Euronews
- AI Rules at Major Streaming Platforms — Digital Music News