Artificial intelligence has given criminals an extraordinary new arsenal. A fraudster can use AI to write convincing phishing emails, impersonate executives, manufacture photographs, clone voices, create deepfake video, automate conversations, and target thousands of potential victims simultaneously. The obvious response has been to warn Americans that AI is making scams more dangerous.
But that raises a more interesting question: If artificial intelligence can be used to perpetrate scams, why can’t artificial intelligence be tasked with defusing them?
There is something strangely defeatist about assuming that criminals will enjoy the benefits of AI while everyone else must simply become more suspicious. Technology has always created offensive and defensive capabilities simultaneously. Encryption protects legitimate communications and conceals criminal ones. Computers can attack networks and defend them. The same should be true of artificial intelligence.
Imagine an AI security agent operating on your phone, computer and financial accounts—not making decisions for you, but standing guard.
An unfamiliar number calls claiming to represent your bank. Your AI recognizes that the caller’s number does not correspond with the institution’s authenticated channels. It analyzes the conversation for patterns commonly associated with fraud: manufactured urgency, requests for verification codes, instructions to transfer money, threats that an account will be closed, or demands that the victim remain on the telephone.
Instead of expecting an elderly retiree to recognize a sophisticated scam, the device could display a warning: “This conversation exhibits characteristics associated with financial fraud. Do not provide information or transfer money.”
AI could go considerably further.
Suppose a grandmother receives a telephone call apparently from her grandson. His voice sounds perfect. He says he has been arrested, desperately needs money and begs her not to call his parents. Today, voice cloning can make such deception disturbingly believable.
But AI could fight AI.
A defensive system might detect artifacts suggesting synthetic speech. It could recognize the classic structure of a grandparent scam. With permission, it might independently contact the grandson through a previously authenticated channel. Within seconds, the grandmother could receive confirmation that the young man supposedly sitting in a jail cell is actually at work.
The scam collapses before the money moves.
Banks could deploy similar systems. Financial institutions already analyze transactions for suspicious activity, but generative AI potentially allows much richer contextual analysis. A $4,000 transfer may be perfectly normal for one customer and wildly abnormal for another. More importantly, AI could potentially recognize combinations of behavior: a customer suddenly adding an unfamiliar recipient, liquidating savings, purchasing gift cards, moving cryptocurrency, or wiring money immediately after receiving suspicious communications.
The challenge is doing this without turning fraud prevention into financial surveillance.
That distinction matters enormously. Conservatives should be especially skeptical of a system in which government or corporations continuously inspect everyone’s private communications under the justification of protecting consumers. Security technology should serve the individual, not become an excuse for creating another enormous database recording Americans’ conversations and financial lives.
The better model is personal control.
Where technically possible, defensive AI should operate locally on the user’s device, with clearly defined permissions. Consumers should decide what their security agent can monitor. Sensitive information should not automatically be transmitted to centralized servers. A fraud-detection system does not need to become a surveillance system simply because both are technologically possible.
There is also a danger in allowing AI to become an unquestioned authority. Fraud detection will never be perfect. Legitimate transactions will sometimes look suspicious. Authentic voices may occasionally resemble synthetic ones. Criminals will deliberately search for ways to manipulate defensive systems.
The AI should therefore function like a smoke detector, not a police officer.
It should warn, verify, slow suspicious transactions and give users better information. Except in narrowly defined circumstances authorized by the customer or existing law, it should not possess unlimited authority to freeze accounts, terminate communications or report private activity to government agencies.
This suggests an enormous opportunity for technology companies, telecommunications providers, banks and cybersecurity firms. Americans may eventually expect an AI fraud shield just as they now expect spam filtering, antivirus protection and credit-card fraud monitoring.
The government has a role, but it should be a limited one. Law enforcement should aggressively pursue organized fraud networks. Regulators can establish reasonable disclosure and security requirements. Congress should modernize laws addressing synthetic impersonation and financial fraud where genuine gaps exist.
But Washington should resist creating a centralized federal AI guardian monitoring everyone’s communications. That cure could ultimately become more dangerous than the disease.
The larger principle is straightforward: Artificial intelligence should not be surrendered to the offense.
Every technological capability criminals acquire should create incentives to develop a defensive counterpart. Deepfake generators should inspire deepfake detectors. AI phishing should produce AI phishing filters. Voice cloning should accelerate voice authentication. Automated fraud should encounter automated fraud detection.
Criminals have historically depended upon asymmetry. They need only fool the victim once. The victim must recognize the deception before that one mistake occurs.
AI could begin reversing that advantage.
The future of fraud prevention may therefore involve something far more powerful than teaching Americans to recognize suspicious emails. It could mean giving every citizen a tireless digital sentry capable of asking, before money moves or information is surrendered, one simple question:
Does this make sense?
If artificial intelligence is intelligent enough to manufacture the lie, we should demand that it become intelligent enough to recognize one.