AI Didn't Fire the Scammers. It Gave Them a Thousand Hands.
Fraud always needed humans to keep lies alive for weeks at a time. Language models, voice clones, and cheap video just removed that bottleneck, and crypto users are the easiest marks in the room. The defense is improving too, but the economics still favor the offense.
AI didn't replace the scammers. It gave them a staff of a thousand. That's the uncomfortable part of the automation story nobody puts on the conference slide, and it's going to cost crypto users more money this year than any onchain exploit you'll read about.
For anyone who figured AI would eliminate every job except the one texting you about your frozen account, fraud just got an automation budget. And fraud, unlike most industries, already knew exactly how to spend it.
Fraud's labor problem, solved
Scamming has always been a job. That's the part people miss. Romance scams need someone to keep a conversation alive for six weeks, remembering a fake job, a fake dog, a fake mother's surgery. Investment fraud needs a person who can answer questions as a fund manager for months without contradicting themselves. Admittedly, that's tedious, repetitive, human work, and it doesn't scale.
So they scaled it with machines. A language model can hold a thousand conversations at once, in fluent English, in whatever tone the mark responds to. Voice cloning turns two minutes of someone's podcast into a phone call from their boss. Video tools produce a plausible executive on a live call just long enough to approve a wire. Fake exchange dashboards come off a template now.
The losses aren't hypothetical. The FBI's Internet Crime Complaint Center logged roughly $5.6 billion in crypto-related fraud losses in 2023, up about 45% from 2022, with investment fraud taking the biggest share. Not all of it was AI-assisted. But the marginal cost of one more victim is collapsing, and that changes the math for everyone.
When another attempt costs nearly nothing, you make a lot more attempts.
The other side of the arms race
Here's the counterpoint, and it isn't weak. Defense teams use the same models. Exchanges run behavioral analytics that flag a withdrawal pattern before the money leaves. Chain-analytics firms, banks, and phone carriers all deploy synthetic-text and voice-clone detection. Some of it works well.
And AI slop has tells. Odd phrasing. Overexplaining. A politeness no human bothers with. Skeptics argue people will retrain themselves to spot machines the way an earlier generation learned to spot a phishing email.
The question worth asking: does that instinct scale faster than the models improve? That's a bad bet, based on the record. Spam filters got better and spam got worse. Every filter trained a smarter sender.
My verdict
The offense is winning because it has the cheaper feedback loop. A scammer can test ten thousand messages tonight and keep whatever converts. A bank can't A/B test its customers' trust. Regulators move in years. Model releases move in weeks, and open weights put the good stuff in anyone's hands for free.
Color me skeptical, but detection alone just sets the training data for the next model.
Watch two things from here. First, whether exchanges start holding first-time withdrawals for 24 hours, which would be annoying and probably effective. Second, whether anyone manages to shift liability onto the platforms whose tools generate the fakes. That fight, not the model race, decides how bad this gets.
Time will tell, though the early returns aren't kind.
Related Articles
Explore More
Key Terms Explained
An approval term meaning authentic, bold, or worthy of respect.
A marketplace where cryptocurrencies are bought and sold.
A social engineering attack where scammers create fake websites, emails, or messages that look legitimate to steal your credentials or trick you into signing malicious transactions.