Behind the Takedown of 260,000 Fake Tracks: Artists' Voices Have Become the Easiest Asset to Steal

Updated:

On October 5, 2026, a report in the Financial Times pushed one number into the spotlight: by the end of September, Sony Music had asked digital platforms to take down more than 260,000 AI-generated tracks impersonating artists on its roster — among them Harry Styles, Adele, Britney Spears, Queen, and Michael Jackson. That figure is nearly double what it was six months earlier: at the end of March 2026, the same count stood at 135,000.

Read on its own, this looks like routine enforcement. Stretched across a timeline, it is a curve that keeps steepening. And behind the curve lies a question nobody can answer: 260,000 takedown requests were sent, but how many fake tracks actually disappeared from the platforms? The FT's report did not say how many tracks were removed, how quickly platforms acted, or which services received the requests. What the public saw, in other words, was the invoice the label sent out — not the receipt the platforms returned.

A Curve That Outruns Enforcement

String together Sony Music's public numbers and this is what emerges: in November 2023, Dennis Kooker, the label's President of Global Digital Business, disclosed at an AI forum hosted by then-Senate Majority Leader Chuck Schumer that the company had sent nearly 10,000 deepfake takedown notices; by March 2025, in a submission to the UK government's copyright consultation, the number had grown past 75,000; by March 2026, per the BBC, it reached 135,000; and by the end of September 2026, it exceeded 260,000.

By Music Business Worldwide's calculation, Sony issued more than 20,000 AI deepfake takedown requests per month on average between April 1 and September 30, 2026. Generation is accelerating, removal is accelerating, and the net stock of fakes does not appear to be shrinking. Sony's statement to the FT was blunt: "We will keep enforcing our artists' rights and expect platforms to act quickly and to prevent the same content from reappearing. But it's becoming an uphill struggle."

Kooker explained the parasitic logic of fake tracks to the BBC in March: "The problem with deepfakes are they are a demand-driven event. They are taking advantage of the fact an artist is out there promoting their music." Fake tracks are not surfing a technology trend — they are hitching a ride on real artists' promotional cycles. New albums, tours, and chart moments are all windows for soundalikes to climb aboard.

Fake Songs Are a Business

The reason these tracks cannot be deleted fast enough is that they are not pranks but a working monetization chain. According to the FT, fraudsters mass-produce tracks with AI, then deploy bot accounts to stream them, siphoning royalties that would otherwise flow to artists and rightsholders.

Data from platforms and researchers sketches the scale of that chain. French streaming service Deezer disclosed in July 2026 that in June alone it received a daily average of nearly 90,000 fully AI-generated tracks — more than half of all new music delivered on peak days. Deezer had previously calculated that in 2025, up to 85% of the streams earned by fully AI-generated songs were fraudulent — though such songs account for only 1% to 3% of total streams on the platform (Deezer data as reported by CNET), meaning fakes still amount to far less listening than real music.

The upstream of the chain is equally well documented. Jamie Njoku-Goodwin, former CEO of UK Music, told the FT: "AI streaming fraud is fast becoming an epidemic for the music industry, and the exponential growth should alarm everyone. Whether it's in deepfakes, streaming farms or AI slop being generated and distributed in bulk, organised fraudsters appear to be using platforms like Suno and DistroKid to rip off artists." A report by the Sonic Intelligence Academy, based on 1,743 verified submissions, found that 92.8% of the AI tracks in its registry were made with Suno, and that more than three-quarters of those with distributor data had been distributed via DistroKid.

The dollar figures remain contested. Kooker told Music Week in May that between 5% and 10% of all streams could be fraudulent, while conceding that "no one knows exactly." The FT, citing label executives, put the industry's annual cost of streaming fraud at up to USD $2.2 billion. These numbers are still estimates (author's note: methodologies vary widely and no third-party audit exists), but they point the same way: this is a low-risk, high-return business.

Law enforcement has just delivered its first conviction. Michael Smith, a North Carolina musician, used AI-generated songs and thousands of bot accounts to collect roughly $8 million in fraudulent royalties; he pleaded guilty in March 2026 and in early October was sentenced to 18 months in prison with forfeiture of proceeds by the US Attorney's Office for the Southern District of New York — the Justice Department's first criminal case of AI-assisted music streaming fraud (per DOJ statements and reporting by The New York Times and others). The deterrent value of a precedent is real, but set against 20,000 takedown requests a month, criminal enforcement is clearly the narrowest stretch of this front.

Why Deletion Can't Keep Up: A Legal Gap and a Numbers Fog

The deeper reason fakes are so durable is that under current US law, cloning an artist's voice does not clearly amount to copyright infringement. When "Heart on My Sleeve," the 2023 track cloning the voices of Drake and The Weeknd, went viral and was pulled, the removal rested more on platform policy than on any clearly defined statutory right. A voice is not a traditional subject of copyright, and likeness rights vary from state to state — which is precisely why the majors keep lobbying for legislation.

Platforms, meanwhile, keep patching. In September 2025, Spotify updated its impersonation policy: "Vocal impersonation is only allowed in music on Spotify when the impersonated artist has authorized the usage." In March 2026, Spotify began piloting Artist Profile Protection, letting participating artists vet releases before they appeared on their profiles; per MBW, the feature was later redesigned — the default mode became notification plus self-removal of incorrectly delivered releases, while hold-until-approved was demoted to a limited-access advanced option. The platform keeps recalibrating between protecting artists and not slowing down legitimate releases.

There is also an oft-overlooked measurement problem. As MBW noted, Spotify's rules distinguish between unauthorized AI voice clones and music fraudulently delivered to the wrong artist's profile — two different behaviors — while Sony's published tally of 260,000 does not distinguish between them. One is infringement; the other may be a metadata error or malicious misattribution. Mixing two different quantities into a single number may overstate the scale of "fake songs," and it leaves outsiders with no way to verify either.

The possibility of collateral damage deserves attention too. In late 2025, the song "I Run" drew listener comparisons between its AI-processed vocals and British singer Jorja Smith; The Orchard, Sony's distributor, along with the RIAA and IFPI, issued takedown notices, and Spotify subsequently detected and removed the track, paying out no royalties. But per MBW citing Billboard, Haven's Harrison Walker and co-producer Jacob Donaghue maintained that they had used Suno to process Walker's own vocals and never targeted Jorja Smith. Whether the song was "impersonation" or a coincidental resemblance remains unsettled; and its label FAMM alleged the track was repeatedly re-delivered through four different distributors to sidestep the usual takedown process. The case displays both faces of the problem: takedowns can be evaded with ease — and they can land on creators who claim innocence.

The Law Catches Up: From the ELVIS Act to the NO FAKES Act

Legal remedies are now advancing on two tracks. At the state level, Tennessee's ELVIS Act, which took effect in 2024, became the first US state law to protect a person's voice — it placed voice alongside name and likeness in the state's right-of-publicity statute and reached not only users of clones but also the makers of cloning tools. Senator Marsha Blackburn has called it the "first generation" of the federal effort.

At the federal level, the NO FAKES Act (S.4591) goes further. As MBW's breakdown explains, the bill would create, for the first time in US federal law, an intellectual-property-style right covering a person's voice and visual likeness: rightsholders could authorize or block AI-generated digital replicas; the right would survive death for up to 70 years and could be inherited. The bill exempts parody, commentary, news, and authorized sampling. Penalties are tiered — $5,000 per work for an individual, $25,000 per work for a company that creates or distributes a replica, and up to $750,000 per work for an online service that fails to comply. To keep its safe harbor, a platform must remove a flagged replica "as soon as is technologically and practically feasible." The July 2024 version died with that Congress; a 2025 reintroduction stalled in the Senate Judiciary Committee amid negotiations with big tech and warnings from free-speech groups; the latest version was reintroduced on May 20, 2026 by a bipartisan group and cleared the Senate Judiciary Committee unanimously on June 18 (full text at Congress.gov). Its backers span the three major labels, Spotify, Google, OpenAI, and YouTube.

At the committee hearing, country artist Martina McBride offered a line that has been widely quoted since: "My voice is my art form."

The Counterargument: Takedowns Don't Cure, and the Law May Overshoot

Treating takedown counts as a measure of success has its skeptics.

First, takedowns are whack-a-mole. The FT disclosed neither actual removal volumes nor response times, and the "I Run" case shows that the same content can be re-uploaded through a different distributor. As long as the gates upstream — the Sunos and DistroKids — stay open, deletion downstream will forever chase production upstream. A takedown count that doubles in six months is less a scoreboard of enforcement than a gauge of generation speed (this is the author's inference from the data above).

Second, detectors make mistakes. The Electronic Frontier Foundation's (EFF) critique of the NO FAKES Act targets its machinery: to preserve safe harbor, platforms would have to deploy "replica filters" and keep flagged content from reappearing — and the track record of existing copyright filters is not reassuring. They take down things that are similar but not the same (two different people playing the same public-domain piece, for instance), act on matches of mere seconds, and ignore the context that makes a use lawful. Extending that infrastructure from copyright to "voice similarity" can only widen the surface for false positives.

Third, the bill's teeth may bite the wrong parties. EFF notes that although parody and commentary are exempted, exemptions are realized through litigation — those who cannot afford lawyers effectively have none — and that the bill would let a claimed rightsholder obtain a subpoena from a court clerk, not a judge, requiring a platform to unmask an uploader based on a bare allegation. The Cato Institute renewed its censorship warnings after the committee vote. Inside the bill, the flashpoint is the ceiling: critics argue the $750,000 per-work penalty for platforms dwarfs the real-world harm. There is a more fundamental objection as well (in Public Knowledge's phrasing): a bill sold as a way to fight deepfakes is in fact a law designed to create a tradable market in likeness rights — answering, with property logic, a question that belongs to privacy law.

These objections do not defend fraudsters. They are a reminder that protecting voices through expanded property rights and mandated filtering may leave the bill with ordinary creators and anonymous users on platforms. Whether the authors of "I Run" were innocent remains unresolved — but disputes like theirs will multiply as filters spread, not shrink (author's inference).

Coda: Invoices and Receipts

A tally of 260,000 takedown requests shows the labels' enforcement machine running at full capacity — and shows that the machine cannot keep up with generation. The real variables sit elsewhere, in two places. One is legislation: whether the NO FAKES Act completes its final stretch in this Congress will determine whether a "voice" is a right that can be clearly asserted at all. The other is the design of responsibility at the platform and tool level: as long as generators and distributors stay ungated, every role downstream is just a courier. Between the invoice — the requests sent — and the receipt — the fake songs that actually vanished — lies a gap nobody can see. An artist's voice has never been easier to copy, and never harder to take back.

Sources