How AI Fraud Actually Works
Seven specimens of AI-assisted fraud, pinned and annotated: the cloned executive, the deepfake video call, the fake job…
Seven specimens, pinned and labelled. None of these are new crimes — they are old crimes wearing better masks. What artificial intelligence changed is not the scheme but the cost of being convincing: fluent language, a familiar face, a trusted voice, produced at volume. Read each one for its shape, not its details. The details will be different when it reaches you.
THE CLONED EXECUTIVE
An email arrives from a senior colleague about a confidential transaction. You hesitate — and then the phone rings, and it is unmistakably them, confirming it. The call is what closes the gap that email alone could not.
- Confidentiality is requested before the money is.
- The request bypasses a process the organisation already has.
- The voice confirms but never invites a callback on a known line.
- The transaction is time-boxed to before someone senior can be reached.
THE MEETING OF GHOSTS
The most expensive documented specimen. In January 2024 a finance employee at the engineering firm Arup joined a video call with the company's CFO and several colleagues. Every other face on the call was synthetic, built from footage that was already public. Fifteen transfers followed, approximately US$25.6 million (HK$200 million), and no internal system was ever breached.
- The meeting is convened by the request, rather than the request arising in a normal meeting.
- Participants perform authority but avoid unscripted exchange.
- Verification is offered only through the channel the attacker controls.
- Structural tell: nothing about the call can be confirmed outside the call.
THE CANDIDATE WHO DOES NOT EXIST
The inversion: the fraud applies to you. A remote candidate interviews well on video, is hired, and is issued credentials and access. The FBI notes that in many of these cases money is not the objective at all — access is.
- Camera artefacts at the edges of the face during fast movement, or reluctance to move at all.
- Audio and lip timing drift when the connection degrades.
- Resistance to spontaneous requests — turning the head, holding up an ID, standing up.
- Documentation that is internally perfect and externally unverifiable.
THE ENDORSEMENT
A recognisable public figure appears in a short video recommending a trading platform. The platform shows early gains, which are numbers on a screen and nothing else. Investment fraud was the single largest AI-flagged loss category in the FBI's 2025 accounting — larger than every other AI-related category combined.
- The figure is famous but the platform is not.
- Guaranteed or unusually specific returns; genuine advisers cannot promise outcomes.
- The video lives on an ad, a reposted clip, or a message — never on the person's own verified channel.
- Withdrawals require an additional payment. That is the moment the machine reveals itself.
THE PATIENT ONE
Weeks of warm, attentive conversation before money is ever mentioned — and language models made that patience nearly free. One operator can now hold many conversations at once, each fluent, each remembering yesterday. The investment ask arrives only after the relationship is established, which is why victims describe the loss as betrayal rather than theft.
- Consistently unavailable for live video, or video that is brief and oddly static.
- The relationship moves quickly off the platform where it began.
- Wealth is displayed early; a financial opportunity is introduced late.
- Requests for secrecy from your family — the same secrecy beat as every other specimen.
THE FLAWLESS LETTER
For twenty years the public was taught to spot fraud by its bad grammar. That lesson is now actively dangerous. Generative models produce fluent, correctly formatted, contextually plausible messages in any language at no cost — and they can personalise each one from information the target published themselves.
- Judge the request, not the prose: credentials, payment, or urgency delivered by message.
- The link text and the actual destination differ — hover before trusting.
- The sender domain is close to, but not exactly, the real one.
- It arrives at a plausible moment. Plausibility is now cheap; treat timing as no evidence at all.
THE PARCEL
The most common specimen, and the least dramatic: an undelivered package, an unpaid toll, a bank alert. Individually trivial, collectively enormous — phishing and spoofing were among the most frequently reported complaint types in the FBI's 2025 report, with over a million complaints filed in total.
- You are asked to pay or log in from a message you did not initiate.
- A small, believable sum — designed to be paid without thought.
- Open the app or type the company's address yourself; never follow the link.
- If it truly is your delivery, it will still be there when you arrive the long way round.
The shape under all seven
Strip the technology away and every specimen runs the same three beats: a trusted identity, a reason not to check, and an irreversible payment channel. AI made the first beat cheap. It did not touch the second or the third — which is precisely where you still have power.
- Verify on a channel the other party does not control. Callback on a saved number, the app instead of the link, the official address typed by hand.
- Treat urgency and secrecy as findings, not context. They are the only two ingredients no legitimate request needs.
- Slow the payment. Wire, crypto, gift cards, and cash couriers are chosen for irreversibility. A method that can be undone is a method a scammer avoids.
Fraud is the harm on this site with the hardest numbers attached, and it lands hardest on the oldest people among us — Americans over 60 reported roughly $7.7 billion in losses in 2025 — the steepest rise of any age group. That deserves anger, but not at the technology in the abstract: the same models write these letters and also help someone read a legal notice they could not otherwise parse. The correct response to a cheaper mask is not fear of every face — it is a household that verifies. Send this page to the person in your family most likely to answer an unknown number.