Deepfake voice and video scams

What it is

A deepfake scam uses generated audio or video to impersonate a specific real person. Three forms account for nearly all documented harm.

The distress call. A relative’s voice, crying, in trouble. It is the family emergency scam with the one weak point repaired — the voice no longer has to be explained away by a bad line or a broken nose.

The executive. A chief executive or finance director on a call or a video meeting, authorising an urgent transfer. This is business email compromise with a verbal confirmation attached, which defeats the control most organisations added first.

The candidate. An AI-assisted job interview to get a synthetic person hired into a remote role. The FBI’s assessment is that the objective here “generally appears to be gaining access to private computer networks” rather than immediate money.

The scale needs stating carefully, because the coverage runs ahead of the data. IC3 recorded 22,364 AI-related complaints and $893 million in losses in 2025. Within that, voice-cloned distress scams accounted for over $5 million — real, and small. AI-linked business email compromise was over $30 million; romance and confidence scams with a likely AI nexus, over $19 million.

The bigger effect is not dramatic at all. The FBI’s own advisory names generative AI being used “to limit grammatical or spelling errors for foreign criminal actors targeting US victims.” Thirty years of advice about spotting bad English described a real signal that has now been engineered away.

How it actually works

1. Reference material is gathered

   Seconds of audio from a social video, a voicemail greeting, a conference talk. Photographs from a public profile. Nothing here requires access to anything private.

2. A voice or a face is generated

   Cloned audio for a phone call, or a synthetic face for a video meeting. The FBI notes the technology is “becoming increasingly difficult to detect and easier to make”.

3. Contact, in a voice you know

   A relative in trouble. An executive with an urgent payment. A candidate in an interview. The pretext is old; only the verification barrier has changed.

4. The checks people were taught stop working

   Fluent writing, a familiar voice, a live video call. All three were reasonable proxies for identity. None of them is any longer.

   Where it could have stopped

   A callback on a number you already had defeats every version of this, because a cloned voice cannot answer a phone the operator does not control. For families, a code word agreed in advance does the same and — crucially — works when the person answering is frightened. Neither depends on your ability to detect anything.

5. The ask

   A wire transfer, bail money, a change of payment details, or system access for a newly hired remote worker.

6. Residual tells, for now

   IC3 lists lip movement that does not fully coordinate with the audio, and sounds like coughs or sneezes that do not align with what is shown. Useful, but not something to rely on.

Why it works

It attacks verification rather than persuasion. Most scams have to talk you into something. This one repairs the step where you would have checked, which makes an old and otherwise unconvincing story survive contact with scrutiny.

Voice was the last intuitive proof. People will doubt an email and still trust a phone call, because recognising a voice feels like direct knowledge rather than inference. That intuition is now wrong and has not caught up.

Distress destroys discrimination. Voice cloning is used most against people who are frightened, and a crying voice on a bad line is hard to identify even when genuine.

Inside organisations, it exploits the fix. Firms that added “call and confirm” as a BEC control have a process that a cloned voice walks straight through — unless the callback goes to a number the firm already held.

And the quiet uses are the widespread ones. Better-written phishing, plausible profiles at volume, convincing fake trading platforms. None of it makes headlines, and it accounts for most of the money.

Where it comes from

Synthetic media has no geography of its own. It is a capability, added to whatever operation already exists — which is why the FBI reports it appearing inside BEC, romance, distress and employment fraud rather than as a category of its own.

What the data does show is uneven adoption. The largest AI-linked losses attach to schemes that were already the largest: business email compromise and confidence fraud. Compound-based operations in Southeast Asia use it for profile generation and translation at scale; West African BEC networks use it for correspondence; the distress-call version turns up wherever family emergency scams already run.

The most under-discussed variant is the employment one. Synthetic candidates interviewing for remote roles is not consumer fraud at all — it is an intrusion technique wearing a job application, and it sits closer to state-linked and organised network compromise than to anything else on this site.

Real cases

Red flags

• A distressed call from a relative on an unfamiliar number. The number is the tell, not the voice.
• Any urgent payment request confirmed only by voice, however familiar the voice is.
• A video call where lip movement lags the audio, or where coughs and other sounds do not match the picture.
• An executive authorising a transfer outside the normal process, particularly while travelling.
• A candidate who avoids turning the camera on, or whose video behaves oddly during an interview.
• A public figure endorsing a specific trading platform in a social media advert.
• Pressure not to hang up and call back. This is the point of the whole exercise.
• A request that arrives with a plausible reason why the usual channel is unavailable.
• Perfect, fluent, well-structured writing from a source that used to be sloppy.

If it’s happening to you

On a call now. End it and call back on a number you already had. Not a number given to you on the call, not a number in a message — one from your own contacts, a card, or an official website. If the caller resists, that resolves it.

For a family emergency, call the person directly, then call another family member. Ask the code word if you have one.

In a business context. Do not act on a voice or video instruction alone. Confirm through a pre-existing channel with a second person, and treat any change of payment details as a stop condition regardless of who appears to be asking. See business email compromise.

If money has gone. Call your bank immediately and ask for a recall; report it the same day. See where to report. In a business case, preserve the call records, the meeting invitation and any recording.

Preventing it.

Agree a family code word. Not guessable, never posted anywhere. Tell older relatives what it is for. It is the cheapest and most effective countermeasure available for the distress variant.

Make callbacks a rule, not a suspicion. In a family and in an organisation, the framing that works is “this is simply what we do”, because it removes the social cost of appearing to doubt someone.

Reduce what is available. Public voice and video is training data. This is not a reason to disappear from the internet, but it is a reason to think about what a voicemail greeting in your own voice, or a public video, is now worth.

Stop treating fluency as evidence. Well-written, correctly branded and confidently delivered are no longer signals of legitimacy, and continuing to teach them as such is now actively harmful.

Where the money goes

Deepfakes do not change the money. A voice-cloned distress call ends in cash handed to a courier or a transfer to a mule account, exactly as the non-cloned version does. A deepfaked executive instruction ends in a wire to a receiving account that empties within hours, exactly as BEC always did.

That is worth holding onto, because it means the defences at the money end are unchanged and still work. The recall window, the second approver, the pre-existing callback number and the bank’s fraud line all operate identically whether or not synthetic media was involved.

What has changed is only the point of verification — which is why every recommendation on this page is about how you confirm a request rather than about how you detect a fake.

By the numbers

No published dataset breaks this scheme out as its own category yet, so there is no chart to show. The data page explains which agency categories exist and why some schemes are invisible in official statistics.