AI music: who gets paid for fake streams?

Press play on an album and the counter does more than measure popularity. It helps decide where money goes. On Spotify, music revenue is allocated to rights holders according to their share of the streams that count towards payment. Software that simulates an audience can therefore redirect income without winning a single new listener. Spotify explains that allocation system itself.

On 24 September 2026, Qobuz introduced a label for albums containing tracks identified as AI-generated. Those tracks account for 0.38% of its streams; 60% of their streams are classified as fraudulent and excluded from royalty payments, according to the company. Figures published by Qobuz.

The distance between those percentages matters. Machines can fill a catalogue without attracting much of an audience. They can also manufacture the records of consumption that never took place. Understanding what that costs musicians and subscribers requires following the payments, rather than deciding by ear whether a song deserves to exist.

How your subscription is shared

A subscription buys access to a service. It does not necessarily direct its entire contribution to music royalties towards the artists that particular subscriber plays. In Spotify’s proportional system, rights holders receive a share of the relevant market’s royalty pool based on their share of listening. The company calls this streamshare. There is no single fixed price attached to every press of the play button. Spotify’s explanation.

There are also different rights behind the same track. The recording and the composition it performs generate separate payment streams. Recording royalties typically pass through labels or distributors; publishing royalties go through publishers and organisations representing songwriters. Contracts determine what ultimately reaches each creator. A payment from a streaming service to a rights holder is not the performer’s take-home income. Spotify’s guide maps the two routes.

That changes how the fraud works. The perpetrator does not have to persuade you to buy a song. Getting an illegitimate claim accepted into the allocation calculation may be enough. Other artists can receive less even though their genuine listeners have not changed their habits.

How fake plays dilute royalties

Consider an entirely fictional month. A service has €1,000 already set aside for recording royalties. This is not its revenue: the example begins after the royalty pool has been determined. All streams carry equal weight and every track is assumed to qualify for payment.

Listeners generate 100,000 genuine streams. One artist accounts for 10,000 of them, or 10%. That artist’s recording earns €100 before any subsequent contractual split.

Now add 20,000 fraudulent streams, credited to a perpetrator’s catalogue and not detected in time. The amount available is unchanged, but the service now divides it across 120,000 recorded streams. The artist still has 10,000 genuine plays and receives only €83.33. The fraudulent catalogue captures €166.67, one-sixth of the pool.

Adding fake streams equal to 20% of the original total does not divert 20% of the resulting payout. The denominator has grown too. The artist’s relative loss is about 16.7%. These numbers do not estimate fraud on any platform. They isolate how invented listening can dilute legitimate payments.

Real services are more complicated. Revenue can change, some fraudulent accounts may pay subscription fees, and licensing agreements contain additional rules. Those expenses affect a perpetrator’s net proceeds. They do not ensure an account can recover only the money it contributed: a shared pool breaks that direct link between contribution and allocation.

Timing matters just as much. Removing the fake streams before allocating the pool prevents the diversion in this example. Finding anomalies after payment raises different questions: stopping subsequent transfers, correcting accounts or recovering money already sent. Dashboard counts and royalty-eligible streams can differ. Spotify distinguishes visible counts from royalty eligibility.

AI supplied a catalogue for an older fraud

The Michael Smith case in the United States takes the issue beyond hypothetical arithmetic. The indictment made public on 4 September 2024 described an operation combining a large catalogue of generated music with automated accounts. Making the recordings and manufacturing the audience were separate parts of the same scheme. The prosecutor’s account set out the allegations at that stage.

On 19 March 2026, Smith pleaded guilty to conspiracy to commit wire fraud. Prosecutors reported billions of automated streams and more than $8 million in fraudulently obtained royalties. Smith agreed to $8.09 million in forfeiture. Actual repayments to affected artists remain undocumented here. The guilty-plea announcement.

In this case, AI supplied a huge catalogue for an operation whose proceeds depended on an artificial audience. That is materially different from a musician using a generative tool and attracting real listeners. Disliking the creative process does not establish fraud in the listening figures.

The problem also predates today’s generators. In its study of listening in France in 2021, the Centre national de la musique estimated detected fake streams at at least 1–3%, using data from several participants with different detection methods. That historical result measures neither global fraud in 2026 nor manipulation that escaped detection. The CNM study.

AI adds production capacity. It does not make every generated track fraudulent, and it is not required to manufacture listening activity for a human-made recording.

The gap between uploads and listening

Deezer’s figures describe a flood of deliveries, not a comparable takeover of listening. On 21 July 2026, the service reported close to 90,000 generated tracks delivered per day on average in June, with their share of daily uploads exceeding half at its peak. The same announcement put generated tracks at 1–3% of streams. Figures published by Deezer.

The first measure concerns incoming music over a period; the second concerns use of the service. It is not possible to infer that half of all listening goes to AI music from the upload figure. A library can acquire a great many books that stay on the shelves.

Deezer also excludes those tracks from recommendations. Their small listening share is therefore observed after that intervention, not in the absence of selection. The published policy.

Qobuz’s two percentages likewise refer to different populations: all streams in one case, streams of the detected tracks in the other. Its announcement does not provide a precise statistical window that would let us reconcile them. We therefore derive neither a euro loss nor an overall fraud rate. Detection retains blind spots in both track provenance and account activity.

Another misleading number concerns Spotify. On 25 September 2025, it announced that it had removed more than 75 million spam tracks over the preceding twelve months. It did not say that every one was generated by AI. The figure does not describe a single week in September 2026 either. The original announcement.

Track provenance and listener activity

Services do not always attach the same meaning to an AI label. In February, Qobuz announced a tool targeting fully generated content. Its September charter acknowledges imperfect detection, especially where human work and generation are combined. The February announcement and current charter.

On 11 August 2026, Spotify announced its AI Persona badge. Its current documentation describes a rollout from mid-September and availability on mobile. Badge documentation. It concerns the profile’s public identity, not directly the production of each recording. Spotify also provides for appeals against profiles it flags. An artificial photographic persona and a synthetic voice are different things. The badge should not be read as a universal diagnosis of a song. Spotify’s description.

Production credits can be more specific. DistroKid distinguishes generated lyrics, compositions and audio, without requiring the same disclosure for tools that merely assist mixing or pitch correction. Those are categories in its disclosure system, not a single legal definition of creativity. Its AI credits documentation.

Above all, two plays of the same file can mean opposite things. In one, a person genuinely chooses to hear the music. In the other, software pretends to be a listener to claim payment. A detector examining only the audio cannot observe that difference. Listening activity still needs checking even when the recording’s origin is fully disclosed.

This distinction also matters for legitimate creators. DistroKid accepts music made with AI tools subject to rights requirements and prohibitions on impersonation and mass-generated spam. Distribution does not guarantee acceptance by every streaming service. The distributor’s rules.

What does “99% accurate” really tell an artist?

The question sounds technical until a flag makes an album harder to find or prevents someone from being paid.

Imagine 10,000 recordings, of which 100 are fully generated. A fictional detector correctly identifies 99 of those 100 tracks. It also correctly clears 99% of the other 9,900 recordings, leaving 99 false alarms.

Across the entire catalogue, it makes the right decision 99% of the time. Yet 99 of the 198 tracks it flags are misclassified. Half its alerts are wrong. Its strong overall result does not, on its own, describe the risk facing the person whose recording is flagged.

These parameters are invented for the explanation. They describe neither Qobuz, Deezer nor Spotify. The share of generated recordings is an assumption about this fictional catalogue, unrelated to the listening percentages discussed earlier.Deezer publishes its own detection and false-positive rates in its FAQ; this article has not independently reproduced those reported results.

The example identifies the evidence worth asking for. Which recordings were used to test the system? What kinds of human intervention did they include? How often does it falsely flag a track? What happens when a musician supplies working files or a production history? Success on one test set does not automatically establish performance on every future catalogue.

Rejecting detectors simply because they are fallible would be equally unhelpful. They can help identify cases for review. But raising a concern, removing a recommendation and withholding money have different consequences. More consequential decisions require stronger justification and a workable way to challenge them.

Qobuz directs artists who contest its badge to their distributor. There is an appeal route; its effectiveness would be easier to assess with information about response times and outcomes. Qobuz’s charter.

Checks begin before anyone presses play

The distributor supplies recordings and their associated information to streaming services. It may know who uploaded a track before its first play. The initiative launched by IFPI on 14 September 2026 calls on distributors to verify identities and rights, screen content and share reliable fraud indicators where legally permissible. These are industry commitments, not public certification of every signatory’s catalogue. The Streaming Integrity Initiative.

The checks address different problems. Identifying the claimant helps investigations and recovery. It does not ensure that person will never buy fake streams. Recognising a copied recording may protect its owner. It does not authenticate the accounts playing it. An effective system has to follow the file, the uploader and the activity behind the payment.

Since April 2024, Spotify has charged labels and distributors in certain flagrant cases of artificial streaming. The incentive is to make repeatedly supplying content associated with manipulation more costly. The same documentation acknowledges that artists can be targeted with artificial plays without knowing it. Bots playing a track do not automatically establish its performer’s involvement. Spotify’s published rules and qualifications.

The industry is also selling its response. Deezer has commercially licensed its detection technology since January 2026. The business may support useful protection while giving the company a commercial interest in demonstrating its importance. Those possibilities are compatible. The commercial launch announcement.

This is a reason to examine results, not to dismiss every alert. A count of removed files measures work performed. Assessing the economic benefit also requires knowing which payments were prevented, what money was actually recovered and which decisions were reversed on appeal. Adding those categories without following the same funds risks inventing savings.

Bring the payment closer to the listener

Changing the allocation system offers another response. In a user-centric model, a subscriber’s distributable contribution goes to the artists that person listened to, according to their use. Activity on another account no longer directly dilutes that allocation. The CNM examined this approach under the name User Centric Payment System. Its study overview.

Take €10 of distributable revenue in a fictional example, not an actual subscription price. If the sum belongs to one account’s allocation, multiplying that account’s plays does not create another €10. Listening volume divides a bounded contribution rather than claiming a share of a larger common pool.

That structure reduces the appeal of some manipulation without preventing compromised accounts or subscription abuse. The CNM specifically considered fraud shifting towards less active users and sub-accounts. SoundCloud, meanwhile, offers listener-linked payments under its fan-powered royalties model for participating content. This is not a universal rule across services or rights. CNM; SoundCloud’s explanation.

Choosing a reassuring name for a payment model is not enough. The relevant questions are which fraud becomes less profitable and what could replace it. A different allocation formula does not eliminate the need to check accounts and recipients.

The conditions for earning royalties

Anti-fraud efforts should not obscure other distribution choices. On Spotify, a track must reach 1,000 streams over the preceding twelve months, alongside an undisclosed minimum number of unique listeners, to enter the recording-royalty calculation. In force since April 2024, the eligibility rule does not apply to publishing royalties. The eligibility conditions.

A small catalogue can therefore be authentic, attract genuine listeners and remain below the threshold. That is not an allegation of fraud. It is a commercial rule deciding which recordings share in the pool. Understanding a payment requires separating that selection from the rejection of artificial activity.

Legal protection for the music is a further question. In its January 2025 report, the U.S. Copyright Office concluded that AI assistance does not exclude protection for human creative contributions; supplying prompts alone is not enough to establish it. That U.S. analysis determines neither the status of every individual work nor French law. The official report summary.

Acceptance by a distributor, availability in a catalogue, a streaming counter and enforceable rights are not interchangeable. Treating them as equivalent creates several ways to misread a payment or a takedown.

Accountability to subscribers and artists

The issue does not require defending every use of AI or removing every recording that contains it. One person may enjoy generated music; another may care deeply about who performed and wrote a track. Neither preference permits anyone to invent the listening activity that triggers payment.

The available evidence establishes a form of fraud capable of diverting real income, alongside measures intended to prevent it. It does not provide a reliable global total for the money currently stolen through generated music. Detection rates, content categories and reporting periods differ too much to build that total by adding them together.

The most dramatic counter is not necessarily the most useful measure of progress. A service should be able to explain what it excluded from payment, what it recovered and how it corrected a mistaken classification, without publishing instructions for defeating its controls.

Subscribers should not need to become experts in audio watermarks to understand where their money goes. They should be able to distinguish editorial selection, a technical suspicion and established fraud. Artists losing visibility or income should know which decision affected them and how to contest it. Checking the audience only does its job if it also protects the people the system might miscount.

Further reading

Classification errors also feature in our investigation of AI and requests for help in debt collection. For voice impersonation, see the Fideuram case.

Sources and documents

• Qobuz : Qobuz Rolls Out a Tag to Flag AI-Generated Music (2026-09-24).
• Qobuz : AI Charter, version 2 (2026-09).
• Qobuz : Qobuz Moves to Protect Artists and Listeners from AI Content (2026-02-26).
• Spotify : Understanding Spotify royalties.
• Spotify : Royalties Guide.
• U.S. Department of Justice, SDNY : North Carolina Musician Charged With Music Streaming Fraud Aided By Artificial Intelligence (2024-09-04).
• U.S. Department of Justice, SDNY : North Carolina Man Pleads Guilty To Music Streaming Fraud Aided By Artificial Intelligence (2026-03-19).
• Centre national de la musique : Étude manipulation des écoutes en ligne.
• Deezer : AI music has surpassed 50% of new music uploads for the first time (2026-07-21).
• Deezer : Deezer confirms demonetization of up to 85% of AI-music streams due to fraud and moves to sell AI-detection Technology (2026-01-29).
• Spotify : Spotify Strengthens AI Protections for Artists, Songwriters, and Producers (2025-09-25).
• Spotify : Introducing a New Label for AI-Generated Artist Identities on Spotify (2026-08-11).
• DistroKid : Can I Upload Music Made With AI Tools to DistroKid?.
• DistroKid : What Are AI Credits?.
• IFPI : Music Companies Unite in New Commitment to Combat Streaming Fraud in Order to Protect Artists, Songwriters and Fans (2026-09-14).
• Spotify : Artificial Streaming.
• Spotify : Track monetization eligibility.
• Centre national de la musique : User Centric Payment System (UCPS).
• SoundCloud : How to get paid on SoundCloud.
• U.S. Copyright Office : Copyright Office Releases Part 2 of Artificial Intelligence Report (2025-01-29).
• Spotify : Your Questions, Answered, Loud & Clear.
• Spotify : AI Persona Badging.

Method and limitations

Sources accessed on 28 September 2026. Companies describe their own rules and results; IFPI represents the industry. Periods and methods differ. No catalogue audit or detector benchmark was conducted. The simulations isolate two mechanisms using fictional assumptions; calculations appear in their captions. This account of the Smith case documents the plea of 19 March 2026. The hearing date announced in that release confirms neither that it took place nor a final sentence or restitution. The copyright discussion concerns the cited US report.