The Person Video Calling You About Your Wallet Does Not Exist
TRM Labs reports AI adoption in crypto crime rose 40% year over year. Deepfake and chatbot scams have climbed sharply since 2022. The significance is not that the tools got better. It is that the expensive part of a scam became free.
Every scam in this section has the same bottleneck, and it is not technical. It is labour.
Pig butchering works because somebody spends weeks messaging a target, remembering what they said, building something that feels like a relationship. Romance fraud requires the same investment. Fake support desks need a person who can improvise when the victim asks an unexpected question. Impersonation needs someone who sounds convincing on a call. The reason these operations concentrated in compounds in Southeast Asia, staffed in many cases by trafficked workers, is that the work was unavoidably human and therefore had to be done by humans somewhere cheap.
TRM Labs reports that AI adoption in crypto crime rose 40% year over year, with deepfake and AI-chatbot scams increasing sharply since 2022. Read alongside everything else in this section, that number describes the removal of the bottleneck.
The pattern breaks down into a few concrete substitutions.
The conversation is now automated. A large language model can hold thousands of simultaneous relationships, each with consistent recall of what the target said three weeks ago, in fluent idiomatic English or Mandarin or German, without fatigue and without a supervisor. The single greatest constraint on pig butchering was the ratio of operators to victims. That ratio is no longer a constraint.
The face is now synthetic. Video verification was for several years a genuinely useful defence: asking someone to join a call defeated most impersonation, because faking a live human face in real time was hard. It is no longer hard. Real-time deepfakes are good enough that a video call now proves less than a phone call did a decade ago.
The voice is now cloned. A few seconds of audio, which almost anyone with a public presence has produced, yields a convincing voice model. The specific application that matters for crypto is the urgent call from a colleague, a family member, or an exchange's security team, arriving at a moment engineered to remove time for reflection.
The documents are now generated. Fake identification good enough to pass a cursory know-your-customer check no longer requires a forger.
What none of this changes is the mechanism of loss. The money still leaves because the victim moves it. Every AI capability above is applied to the same target it has always been applied to, which is the moment a person decides to trust an instruction. CMZ has documented recovery scams that re-victimise people already defrauded, a DPRK operative who wrote code inside MetaMask for a month before anyone noticed, and a hijacked X account that had a fake token trading within minutes. None of those needed AI. AI makes each of them cheaper to run and possible to run at a scale that was previously limited by how many people you could pay to sit at a desk.
There is one countervailing detail worth recording, because it cuts against the alarm. Automated conversation is cheap but it is also detectable in aggregate: thousands of accounts running the same model produce statistically similar behaviour, and exchanges and messaging platforms can look for that pattern in ways they could never look for a human in a compound. The advantage AI hands to fraud at the individual interaction is partly given back at the population level. That is cold comfort if you are the individual.
The defensive implication is the part worth taking seriously, because most published advice has not caught up. Guidance built around spotting bad grammar, awkward phrasing, or a reluctance to appear on camera is now advice for detecting a class of scammer that has been automated away. The reliable checks left are the ones that do not depend on judging authenticity at all: verifying through a channel the other party did not choose, refusing to act inside an artificial deadline, and treating any unsolicited contact about your holdings as hostile by default regardless of how credible it appears.
The uncomfortable version of that is simpler. You can no longer tell. So stop trying to tell, and change what you do instead.
The Aftermath
AI-enabled crime is now tracked as a distinct category by blockchain intelligence firms rather than treated as an emerging concern. No single incident defines the trend; the change is in the unit economics of every scam type already documented in this section. Conventional detection advice built on spotting language errors, refusal to appear on video, or inconsistent storytelling has been substantially undermined. The countermeasures that still work are procedural rather than perceptual: independent verification through a channel the other party did not select, and refusing to act within a deadline somebody else imposed.
COMMENTS