A man named Michael Smith made more than $8 million producing music.
How?
Platforms like Spotify pay small royalties when a song is streamed. Each play is worth a fraction of a cent. Musicians complain about this all the time. How can platforms pay so little? How did music become so “cheap”?
But billions of streams still add up to millions of dollars. And that was exactly what Smith was counting on.
He didn’t have a single hit. Or a real fan base. Yet millions kept landing in his account, year after year.
He hadn’t written the songs himself. As you’ve probably guessed, AI had.
The kind of songs you could make today by downloading an app like Suno. Except he didn’t make a handful. He made hundreds of thousands.
AI songs have become surprisingly enjoyable. You’ll happily listen to them, especially the ones you made yourself. But who has the time to sit through everyone else’s AI experiments?
So who listened to Smith’s hundreds of thousands of songs and made him all that money?
Nobody.
The “listeners” were automated accounts pretending to be people. Bots. Thousands of them, playing the songs billions of times.
AI wrote the songs. Software played them. The platforms treated those plays as genuine listening and paid out, a fraction of a cent at a time.
Quite a setup.
Eventually, it was uncovered, and charges followed. It sounds like a straightforward fraud case, but it has been described as the world’s first AI music fraud case.
That isn’t even the strangest part.
Fine, the bots play the songs and you collect the money. But why generate hundreds of thousands of tracks instead of five or ten?
Hoping at least one might become a hit?
Exactly the opposite. He wanted none of them to stand out.
A single song getting a billion plays would make platforms suspicious. Smith knew that. In a 2018 email cited by prosecutors, he explained the strategy: he needed lots of content, with relatively few streams per track, to stay under the radar.
Artists usually try to rise above the crowd, make the best music, attract the most listeners, and eventually get paid. Smith was producing music to disappear into the crowd.
The scheme ran for years before he was caught. As mentioned, it became the first U.S. criminal case involving AI-assisted music streaming fraud. Prosecutors sought at least 46 months in prison. The defense asked for no prison time. On October 6, the sentence came down: 18 months.
Yes, it was a first. Yes, the scheme was devilishly clever. Yes, the sentence seems laughably light. But the sentence isn’t why I’m telling this story.
For years, the counters ticked up, reports arrived, and payments went out. Everything looked like a music career.
What made that possible?
AI Makes a Real Difference. It Doesn’t Take Sides.
The answer is AI. But not in the way you hear about every day.
Someone is always promising you can “get rich with AI.” Most of it is nonsense. This story isn’t, because we can see exactly what AI changed: scale, rather than creativity. Without it, hundreds of thousands of songs would have taken far more time and money to produce.
Fake streams didn’t begin with AI. According to the indictment, Smith started working with an AI music producer around 2018. The fraud already existed. What was missing was enough music. AI filled that gap.
AI is powerful. It doesn’t care whose purposes it serves. Whatever you put in its hands, it scales up.
We keep asking the same question about AI: Can it write a good song?
Here, that was the wrong question. The songs didn’t have to be good. They didn’t even need anyone to listen. They only needed to look as though someone had.
We know the familiar story: production gets cheaper, everyone starts producing, competition intensifies, and attention becomes harder to earn. Smith turned that story upside down. While the rest of us complained about the crowd, he used it as cover.
So the useful question goes beyond “What can I do with AI?”
What has AI made cheaper, and whose metrics does that make less trustworthy?
The same AI can help a musician find the sound they’ve been hearing in their head. It can also produce the material needed to make an audience appear to exist.
The audio file alone won’t tell you which happened.
Who Is the Counter Counting?
Read this as a story about measurement. Something similar could be happening in your own reports.
A streaming record tells you something very narrow: a track was played.
We fill in the rest. Someone chose it, listened, perhaps liked it, perhaps played it again. Of course, every stream isn’t a declaration of love. Background music counts too. But the logic behind the payout assumes that somewhere, a person had an experience.
Smith’s system contained the record of that experience, without anyone having it.
Here’s the trap: you have two signals. Lots of music. Lots of listening. Taken separately, they might suggest a thriving market. But the same machine produced both. Neither independently validates the other.
It’s like a sitcom laugh track. You hear laughter because someone pressed a button, not because anyone found the joke funny.
That’s where this story reaches beyond music: as producing things gets easier, so does producing the signals that suggest those things found an audience.
Who Paid the Bill?
It might look like a closed loop with no victims. A computer makes a song, a bot plays it, and nobody gets hurt.
But the money wasn’t created inside that loop. In the basic model Spotify describes, royalties come from a shared pool and are distributed according to each recipient’s share of total streams. Fake streams can shrink the share attributable to genuine listening. The Justice Department says Smith’s earnings were diverted from other musicians and rights holders.
The defense’s main argument was that no artist suffered any perceptible harm.
That captures the nature of the harm quite well. Nobody lost a fan. The same people kept loving the same songs. But that doesn’t mean there was no damage. Thousands of people’s shares quietly eroded, a fraction of a cent at a time, without any one of them noticing it in their own pocket.
That’s what makes it so insidious.
Your competitor isn’t persuading your customers to switch. They’re taking a cut of the money allocated according to your customers’ behavior.
Which is why debating whether AI music is as good as human music misses the point here. The songs could have been much worse. The system would still have paid.
The Innocent Version of the Same Trap
This time, there’s no fraud. The intentions are good. But the mechanism looks familiar.
You have an idea for an app. You build a prototype with AI, then ask simulated personas representing your target audience to evaluate it:
“The price is a little high, but I’d pay if there were a weekly plan.”
“I love that feature, but there are too many notifications.”
“I already do this in Excel. Why would I switch?”
Three different people. Three different opinions. A working prototype. It feels like customer feedback.
But there are no customers. Those voices aren’t independent people. They’re echoes of your own brief. Adding a hundred more personas makes the echo louder. It doesn’t strengthen the evidence.
Don’t get me wrong: this exercise can be useful. It can uncover a question you missed, a feature that conflicts with another, or a need worth investigating. It’s excellent for generating questions. The moment you use it to supply the answers, you’re listening to the laugh track.
Automation can play the same game. The report was generated. The message was sent. The record was created. All of those things really happened. But the benefit you expected may not have materialized.
And once you reward the system for those intermediate steps, it starts optimizing for them. There’s an old rule: when a measure becomes a target, it stops being a good measure. Smith’s story is one of its most expensive illustrations.
Automation itself isn’t the culprit. You might truly love an AI-generated song. An assistant might buy exactly the right product on your behalf. Who pressed the button matters less than what happened afterward.
Was a real need met outside the system that produced the activity?
The Real-Person Test
Before trusting a number, ask three questions:
Did someone outside the system measuring this signal generate it?
Did that person give something up for it: money, time, an existing habit?
If I shut the system down, would the signal still arrive?
If the answer to all three is no, what you have is the echo of your own voice, rather than evidence.
Smith’s file contained everything: music, streams, money.
Only one thing was missing: someone who heard a song and thought, “I want to play that again.”
Production got cheaper. Finding a real audience didn’t.
If you can’t measure the difference, you’ll call noise growth.