If you pay for a streaming subscription, you assume your listens help real artists. The Michael Smith case says that’s not always true. Prosecutors described a playbook built for scale: hundreds of thousands of AI-generated tracks, automated bot accounts, VPN masking, and a farm of cloud accounts pushing fake plays for years. The result wasn’t just “bad data.” It was allegedly millions pulled from the same royalty pool legitimate artists rely on—ending in an 18‑month sentence, $8.09M forfeiture, and two years of supervised release.
What prosecutors say happened (and why it worked for so long)
The alleged play wasn’t subtle. It was just scaled.
Prosecutors say Michael Smith built a streaming fraud machine with two moving parts: mass content and mass listening. On the content side, he allegedly uploaded hundreds of thousands of AI-generated songs (bought from an accomplice) to major streaming platforms. On the listening side, he allegedly used automated bot accounts to stream those tracks billions of times.
That combination matters. Bots are easier to spot when they hammer one track into the ground. It’s harder when there’s an endless inventory of tracks to rotate through.
The mechanics prosecutors pointed to
Here’s the core workflow described in the court-doc summary:
- Upload at scale: A huge catalog of AI-generated tracks gets distributed across platforms.
- Automate “listeners”: Bots run streams around the clock, creating fake engagement that looks like real demand.
- Hide the fingerprints: Bots connect through VPNs to dodge anti-fraud systems that flag suspicious IP behavior.
Then comes the part that explains “why it worked for so long”: the operation wasn’t a handful of accounts doing obviously impossible things. Prosecutors cite scale and distribution. At the scheme’s peak, Smith allegedly used more than 1,000 bot accounts. He even emailed himself an internal breakdown describing 52 cloud service accounts, each running 20 bot accounts—a neat, modular way to spin capacity up or down without putting all the risk in one basket.
The “spread thin” strategy (the part most teams miss)
In a 2018 email, Smith spelled out the evasion logic in plain language: “we need to get a TON of songs fast to make this work around the anti fraud policies” and they needed “a TON of content with small amounts of Streams” to avoid raising issues.
That’s the tell.
If your detection is tuned to catch spikes, a fraudster can just manufacture flatline volume across a massive catalog. Each track stays under typical alert thresholds. The total payout still stacks.
And that’s why “AI-generated music” isn’t just a content moderation headache. It can be the fuel source for streaming manipulation—because when supply is basically unlimited, fraudsters don’t have to protect any single song.
The numbers that should scare every platform team
The “spread it out” tactic only works because the math stays profitable at volume.
In one internal breakdown prosecutors cited, Smith estimated each bot could stream ~636 songs per day. Multiply that across his operation and you get ~661,440 streams/day.
That’s the kind of number that quietly wrecks a royalty pool. No single track needs to look like a breakout hit. The payout still arrives.
The per-stream economics that add up fast
Prosecutors say the same breakdown assumed an average royalty rate of half a cent per stream. At that rate, the estimate was:
- $3,307.20 per day
- $99,216 per month
- $1.2M+ per year
Two uncomfortable takeaways for trust & safety teams:
- Fraud doesn’t need perfect evasion. It needs “good enough” accounts that survive long enough to print volume.
- Small payout rates don’t protect you. Low per-stream value just pushes attackers to automate harder.
The DOJ comparison that should reset your baselines
Platforms often sanity-check listening by comparing it to “normal” demand. DOJ did that here using YouTube Music family plan streams.
They cited April 2023, where Taylor Swift’s entire catalog got 9.3 million family plan streams. In the same month, DOJ says Smith’s bot accounts used family plans to fraudulently stream his AI-generated music 80.9 million times.
That’s not “slightly above normal.” That’s a fraud signal screaming through your metrics.
And it lines up with what DOJ says Smith later boasted in a February 2024 email: “over 4 billion streams and $12 million in royalties since 2019.”
Why AI content changes the fraud model (it’s not just ‘more bots’)
Those stream counts only pencil out if you can keep feeding the machine. That’s where AI shifts the threat model.
In the Smith case, prosecutors say the catalog itself was the accelerant: hundreds of thousands of AI-generated songs uploaded across platforms. When content creation becomes cheap and fast, fraudsters don’t need a “hit.” They need inventory.
AI turns fraud into a supply-chain problem
Traditional streaming fraud was often constrained by content scarcity:
- You had a small set of tracks.
- Bots hammered them.
- Detection had clear targets (repetitive playback, obvious concentration, unnatural velocity).
With AI-generated music, that scarcity disappears. A bad actor can spin up thousands of “new” tracks, each one disposable, each one capable of absorbing a little fraudulent listening.
Prosecutors captured this mindset in Smith’s own words. In a 2018 email, he wrote: “we need to get a TON of songs fast to make this work around the anti fraud policies” and called for “a TON of content with small amounts of Streams” to avoid raising issues.
That’s not an artist mindset. That’s an evasion strategy.
Why older detection struggles (plain English)
Most anti-bot systems are strong at catching loud patterns. AI-assisted streaming fraud can stay quiet:
- Rotation beats repetition: Bots can rotate across huge catalogs, so no single track looks “botted.”
- Distribution hides intent: Streams spread thin across many tracks can look like long-tail discovery.
- Freshness breaks blocklists: If you’re keying off known bad track IDs or known bad audio, “fresh” uploads keep resetting the clock.
The hard part isn’t spotting bots in isolation. It’s spotting the content farm + listener farm working together—especially when the operator is intentionally staying under the usual tripwires.
What platforms must fix now: controls that make fraud expensive again
If AI makes “infinite songs” cheap, the only way to win is to make fake listening expensive. Not annoying. Expensive.
This case description has the usual ingredients—bots, VPN masking, family plan abuse, and cloud-scale operations—but the fix isn’t one magic model. It’s a control stack that ties behavior together across accounts, devices, networks, and payout flows.
1) Family plan + subscription abuse controls (treat plans like risk objects)
DOJ explicitly called out fraudulent streaming via family plans. That should push plan-level monitoring higher up your roadmap.
Tactical controls:
- Household plausibility checks: reasonable limits on device churn, geo spread, and concurrent usage per plan.
- Plan graphing: link accounts under the same plan to shared signals (payment instrument, device clusters, session timing).
- Tiered friction: if a plan looks “industrial,” step up checks before it can generate meaningful royalty weight.
2) Bot mitigation that targets automation, not just volume
Fraud farms don’t act like enthusiastic fans. They act like software.
What works in practice:
- Session integrity signals: impossible play timing, always-on playback, repetitive interaction patterns.
- Rate controls that attackers can’t easily tune around: per-device and per-account caps on “royalty-eligible” behavior (not just total plays).
- Kill switches for suspected farms: fast containment when a cluster is confirmed, not days later after payout cycles.
3) VPN/proxy risk scoring (because “hide behind VPNs” is a feature, not an accident)
Prosecutors say VPNs were used to avoid detection. A blunt “block all VPNs” stance will hurt legitimate users, so treat it as a risk input.
Practical approach:
- Reputation + velocity scoring: new account + datacenter/VPN egress + high playback regularity = higher risk.
- Geo-consistency checks: frequent country hops without travel signals should downgrade trust.
- Step-up challenges only when needed: keep baseline UX clean, add friction when the score demands it.
4) Stronger identity + device signals (link “new” accounts back to the same operator)
The operation described relied on lots of accounts and cloud scale. Your job is to make “new account” not equal “fresh start.”
Controls that move the needle:
- Device fingerprinting (privacy-aware): stable device signals to cluster repeat operators.
- Behavioral linking: shared playback schedules, identical playlist paths, similar session structure across accounts.
- Account lifecycle scoring: brand-new accounts shouldn’t immediately have full influence on payouts.
A privacy note platforms can’t ignore
Yes, better verification helps. No, you don’t need to hoard personal data to do it.
One clean principle: let real users prove continuity while limiting exposure. Tools like Cloaked (masked emails/phone numbers and compartmentalized identities) are a good example of how consumers reduce account takeover and limit identity re-use across services. That’s not “anti-fraud” by itself, but it supports a healthier ecosystem where account creation and recovery don’t depend on oversharing—while platforms focus their strongest checks on high-risk behavior patterns.
