The S.D.N.Y. Gets Mad about AI Music
Till now, the AI enforcement cases have been mostly (entirely?) AI washing or misrepresention cases. A company says all the great things they’re doing with artificial intelligence, but they’re not actually doing them, or the claims outstrip the performance. It’s fun and all, but they’re pretty basic. People lie about all kinds of things to make money, and AI is one of those things. Thankfully, Michael Smith has arrived to show us a way people might use AI to actually do fraud. He did it with music on streaming services, but maybe we can use his model to imagine a brighter future where people are doing similar frauds in other areas.
How Streaming Payments Work
To understand what Smith did, we’re have to walk through how the music streaming industry works today. As you probably know and do, people stream music through platforms like Apple Music and Spotify. Each time a song is played, the songwriter, performer, and sometimes others are entitled to small royalty payments. The exact amount depends on a number of factors, but it’s often less than one cent per stream.
The streaming platforms pay the royalties. The amounts are calculated as a percentage of streaming revenues, and the royalty pool is then allocated proportionally among artists based on their respective percentages of total streams. Different mechanisms apply to songwriters versus performing artists, but that is the gist. If someone were to, say, use AI to create a bunch of songs and also bots to download and stream those songs, you can see how that might mess up the system we’ve arrived at here.
The S.D.N.Y.’s Allegations
So . . . on September 4th, federal prosecutors in the Southern District of New York indicted a guy named Michael Smith for “fraudulently obtain[ing] millions of dollars in royalty payments from the Streaming Platforms, Rights Organizations, and music distribution companies.” As you might expect, all of the players involved here frown upon this sort of thing and their terms of service prohibit it. But here’s what they say Smith did:[i]
Stage 1 – Create the Bots
First Smith created thousands of accounts on the streaming platforms that he could use to stream songs. To do that, he obtained thousands of email accounts from vendors who sell them in bulk. Then he used those emails to register as many as 10,000 bot streaming accounts in fake names. AI can do a lot, but it sadly can’t do everything yet, so signing up so many was a chore. He was smart about it, so he typically signed up for family plans to get the most bang for his buck. To make it look like each bot account used a different source of payment, Smith used a company that provides large numbers of corporate debit cards for corporate employees, saying they were for his employees.
Stage 2 – Release the Hounds (or Commence the Streaming)
After registering the bot accounts, Smith had them continuously stream songs he owned in this way: (a) he used cloud services so he could use lots of virtual computers at once. (b) he put some of the bot accounts on each virtual computer at the same time. He typically used the web players for each streaming platforms and kept a number of bots streaming music on separate browser tabs. (c) he bought pieces of computer code called “macros” that continuously played the music for him.
Stage 3 – Collect the Money
Then he watched the money roll in.
By late 2017, Smith had 52 cloud services accounts with 20 bots on each, for a total of 1,040 bot accounts. At 636 songs per bot per day, he could rack up about 661,000 streams every day. At a ½ cent each, that meant $3,300/day, $99,000/month, and $1.2 million/year. It would get bigger!
Creating the Songs
He didn’t want people to figure out what he was doing, and knew that if a single song was streamed a billion times, it would raise suspicions. A billion streams spread across tens of thousands of songs, though . . . that just might be crazy enough to work. But where could he get enough songs? He needed a LOT of them.
So in 2018, he began working with an AI music company and a music promoter to create hundreds of thousands of songs using artificial intelligence that he could stream. After some trial runs, the music company started providing Smith with thousands of songs every week. Smith would own the intellectual property rights in the songs in exchange for metadata and a share of revenue.
Titles
The songs typically had randomized file names, such as “n_7a2b2d74-16214385-895d-ble4af78d860.mp3.” Then he created randomly generated song and artist names to look less robotic.
Here's a sample of 25 of Smith’s song titles: "Zygophyceae," "Zygophyllaceae," "Zygophyllum," "Zygopteraceae," "Zygopteris," "Zygopteron," "Zygopterous," "Zygosporic," "Zygotenes," "Zygotes," "Zygotic," "Zygotic Lanie," "Zygotic Washstands," "Zyme Bedewing," "Zymes," "Zymite," "Zymo Phyte," "Zymogenes," "Zymogenic," "Zymologies," "Zymoplastic," "Zymopure," "Zymotechnical," "Zymotechny," and "Zyzomys."
And here’s a sample of the “artists”: "Calliope Bloom," "Calliope Erratum," "Callous," "Callous Humane," "Callous Post," "Callousness," "Calm Baseball," "Calm Connected," "Calm Force" "Calm Identity" "Calm Innovation" "Calm Knuckles" "Calm Market" "Calm The ' ' , ' ' Super," "Calm Weary," "Calms Scorching," "Calorie Event," "Calorie Screams," "Calvin Mann," "Calvinistic Dust," "Calypso Xored," "Camalus Disen," "Camaxtli Minerva," "Cambists Cagelings," and "Camel Edible."
Of course, there are emails. In June 2019, Smith emailed his partners, "We are at 88 million TOTAL STREAMS so far!!!"
People didn’t flip out about ChatGPT until 2022, so it’s kind of impressive that he was able to pull this off in 2017 to 2019. As you can imagine, the AI technology was getting better throughout his scheme. In August 2020, the music company CEO wrote to SMITH, "Song quality is 10x-20x better now, and we also have vocal generation capabilities. . . . Have a listen to the attached for an idea of what I'm talking about."
Hmmm . . .
I guess one question is how others are going to use AI in similar ways to do crimes in other fields. I don’t know! I wonder if there’s a fun AI wash trading scheme in our future. Maybe some AI health insurance fraud with fake patients and maladies? The instances there might be too high-dollar for AI to be an efficient vehicle. Just buckle up because we live in a magical time.
United States v. Smith, 24 Cr. 504 Sealed Indictment (Sept. 4, 2024)
[i] It’s just an indictment. He hasn’t been convicted, so he might not be guilty of everything that follows.