Streaming Fraud Is Not a Money Crime. It Is a Taste Crime.

Eleven minutes. That is how long it took me to make a track last month that I would not defend in a room full of musicians: 118 BPM, a plate-drenched Rhodes, a sub that dies below 40 Hz, hats off a preset I have used a hundred times. It is not good. It is playable. Drop it into a lo-fi playlist between two human records and nobody reaches for the skip button. Those eleven minutes sit underneath every current argument about streaming fraud, because the cost of manufacturing something streamable collapsed to near zero while the industry kept arguing about the demand side.

The verdict up front: this is not mainly a royalty-theft problem. It is a signal problem. The stolen money is recoverable, and against a global pool it is small. What is not recoverable is what the fake plays taught the recommender, the playlist editor, and the A&R who reads a chart the way the rest of us read a thermometer.

The myth everyone on the panel repeats

The version you have heard goes like this. Bad actors point bot farms at tracks and siphon a slice of the pro-rata pool. Platforms detect most of it, strip the streams, claw back the money. Honest artists lose fractions of a cent each, which is bad in aggregate and survivable in particular. It is a compliance problem with an engineering fix, and the fix is arriving.

I have heard that from people who understand the plumbing far better than I do, and it is not a stupid position. It is what the money line supports. It is also the account that lets every party keep its current job description: detection vendors sell detection, platforms publish removal counts, collecting societies audit distributions, and nobody has to answer for the part of the system that has no owner.

The trouble is that money is the least interesting thing the fraud touches.

What the record actually shows

Start with the case everyone cites. In September 2024, federal prosecutors in Manhattan charged a North Carolina musician, Michael Smith, over a scheme the indictment says ran for years and collected in the region of ten million dollars in royalties. The headline number got the coverage. The architecture is the part worth reading twice: prosecutors described hundreds of thousands of AI-generated tracks, streamed by thousands of bot accounts, with plays deliberately spread thin so that no individual song ever looked like anything at all. The design goal was not a hit. The design goal was invisibility.

Size the field around that. Beatdapp, one of the firms that sells detection to the industry, has put the fraudulent share of global streams in the neighbourhood of ten percent. Apply the usual discount for a vendor sizing its own market, then notice that nobody with better data has published a lower figure. France's Centre national de la musique, which has nothing to sell, studied 2021 streams for a report published in early 2023 and landed between one and three percent — of a single national market, that is north of a billion plays, and the authors expected the share to grow rather than shrink.

Then the supply shock arrived. Deezer began flagging fully AI-generated uploads in January 2025 and said at the time they were around a tenth of everything arriving daily. Through that year the company reported the share climbing to several times that level. The detail that deserved more attention than it got: Deezer also reported that the overwhelming majority of streams on those fully AI-generated tracks were fraudulent. The synthetic music was, for the most part, being listened to by synthetic listeners. Two automated systems, transacting with each other, against a real royalty pool.

Platforms have not been passive. Spotify's 2024 policy change — no recorded royalties below a floor of roughly a thousand annual streams per track, plus per-track penalties charged to distributors when artificial streams are detected — genuinely raised the cost of the wide-and-shallow approach, which is why it drew complaints from people whose catalogues live below that floor. Deezer publishing its AI-upload numbers at all is a service to everyone else. Credit where it is due, and then the honest caveat: both are measures of what the platform caught, disclosed on the platform's own schedule, with the methodology held close.

Meanwhile the culture side ran its own demonstration. In summer 2025 a band called The Velvet Sundown accumulated around a million monthly Spotify listeners, and picked up playlist placement and press, before its creators described the whole thing as a provocation with AI at the centre. No fraud was alleged there. That is exactly why it matters: an entirely legitimate listener count, built on an entity that did not exist, moved through the discovery machinery without friction.

And now there is a layer on top. EL PAÍS reported this summer on traders taking positions in prediction markets on chart outcomes, including one who spotted a streaming anomaly before the chart moved and was paid handsomely for it. Whatever you think of the size of that market as of writing, look at what it introduces: a financial reason for a complete stranger, with no stake in any artist, to want a specific number to move on a specific day.

The mechanism nobody owns

Here is the thing the money framing misses. A play count does two jobs at once. It is the key that unlocks a payout, and it is an input to ranking. The same integer that decides who gets paid also decides what gets surfaced, what gets playlisted, what shows up in the co-listen matrix beside a genuine hit, and what a human A&R sees on a Monday morning when they are deciding which of eleven artists to fly out.

Detection operates on the first job and cannot reach the second. Fraud identification is retrospective by nature: you need enough behavioural history to distinguish a bot from an insomniac, which means the stripping happens days or weeks after the streams landed. Ranking is not retrospective. It updated in real time, propagated to the editorial dashboard and the recommendation model, and produced downstream listening by actual humans that is now indistinguishable from organic demand, because it is organic demand — it was seeded by a lie. You can claw back the money. You cannot claw back the playlist add, the algorithmic association, or the memory of the person who saw the number and believed it.

The second mechanical consequence is a selection effect that quietly corrupts every statistic in the previous section. If enforcement runs on thresholds, the optimal fraud is wide and shallow: three hundred thousand tracks each sitting comfortably under the line, none of them ever charting, none of them ever worth investigating individually. That is precisely the shape prosecutors described in 2024. Which means the manipulation that gets caught, counted and reported is disproportionately the clumsy kind — the campaign that pushed one song too hard. Every published fraud figure is a floor wearing the costume of an estimate.

The third is jurisdictional. Ask who is responsible for the integrity of a chart and you get a shrug wearing four different lanyards. Platforms police their own catalogues and disclose what suits them. Distributors gate uploads and eat the penalties. Labels investigate their own artists with obvious enthusiasm limits. Collecting societies distribute against numbers they did not generate and cannot independently verify. Financial regulators supervise the betting venue but have no view on the underlying asset, and no music regulator anywhere has authority over a securities-adjacent contract. There is no arbiter. There is not even a body whose job it would be to convene one.

How I read a chart number now

I am a sound designer, not an auditor, but I have stopped taking these figures at face value and started applying five questions. They are not exotic, and they work on a press release.

• What is the denominator? Streams, monthly listeners, saves and "reach" are four different claims. Monthly listeners is the softest and the most quoted.
• Does it decay? Real listening has a tail. Purchased listening stops the day the invoice does. A number that goes vertical and then flat over a 28-day window is telling you something.
• Where is it? Play geography that has no relationship to where the artist has toured, charted, or been written about is the oldest signal in the book and still the most reliable.
• What is the ratio between passive and deliberate? Saves, playlist adds and repeat listens per stream are expensive to fake convincingly. Raw plays are cheap. The gap between them is the tell.
• Who profits from it being believed, as distinct from it being true? After prediction markets, this is no longer a rhetorical question, and the two groups are no longer the same people.

Who this changes the work for

If you run a platform, none of this is news, and your constraint is not detection capability but disclosure risk. If you sit in A&R, the practical implication is narrow and immediate: treat a streaming number as a claim requiring corroboration rather than as evidence, and weight the metrics that cost money to fake. If you work in policy, the gap worth legislating is not "ban the bots" — everyone already agrees on that — but a duty to publish methodology, so that removal statistics can be compared across services and years by somebody outside the building.

And the limits, since this publication does not pretend to more than it has. I cannot tell you what share of any given chart is manufactured. Neither can anyone else, credibly; the number is unknowable by construction, because the best fraud is designed to be statistically boring. Anyone quoting you a precise figure is selling either detection software or outrage. What is knowable is the direction of the incentive, and the incentive got stronger the month a competent track stopped costing a week and started costing eleven minutes.

One last thing, because it would be easy to read this as an argument against the tools. It is not. I use them, they are on my drive, they are an instrument and a decent one. The generator did not commit the fraud. The fraud lives in a payout scheme that pays per play and a discovery scheme that ranks by the same integer, which together offer a standing bounty to whoever can manufacture plays most cheaply. AI made manufacturing cheap. It did not write the bounty.

Charts were never a mirror held up to taste. They are an input to it, and inputs can be bought.

The money is recoverable. The listening is not.