AI Voice Cloning Scam: How It Works and How to Stop It
Short answer
In an AI voice cloning scam, criminals copy a relative's voice from short clips posted online and call with an urgent story, usually an accident, arrest or kidnapping, to get money fast. The defence is simple: hang up, call the person back on a number you already have, and agree a family safe word in advance.
The call comes at an awkward time, often late in the evening. Someone you know, your son or your granddaughter, is crying and says they have been in an accident or arrested. Someone else takes the phone: a lawyer, a police officer, a stranger who wants money. The amount is specific, the deadline is short and you are asked not to tell anyone.
Scammers ran that script long before AI. What changed is the voice. It used to be a stranger claiming to have a cold. Now it can sound like the actual person.
How common is it?
Reliable crime statistics rarely separate voice cloning from other impersonation fraud, so most numbers come from surveys. They are self-reported and should be read that way. Still, they point in the same direction.
Source: McAfee, 7,054 adults in 7 countries, April 2023; Starling Bank / Mortar Research, 3,000+ UK adults, August 2024
The FTC’s 2024 figures put the wider category in context. US consumers reported losing $12.5 billion to fraud, and imposter scams accounted for $2.95 billion of that. Phone calls were the second most common contact method after email (FTC). How much of that involved cloned voices is not broken out.
How the scam works
The method is not sophisticated, which is part of the problem.
- HarvestShort clips from public videos, voice notes or voicemail greetings.
- CloneA voice model produces new sentences in that voice. A few seconds of audio can be enough for a rough copy.
- ResearchNames, relationships and travel plans come from social media.
- CallA brief emotional clip in the cloned voice, then a second person takes over with the details.
- PressureUrgency, secrecy and a payment method that is hard to reverse.
The cloned voice usually appears only briefly. Sobbing sentences are easier to fake than a long conversation, and emotion hides artefacts. The “lawyer” or “officer” who takes over speaks in their own voice and handles the logistics. McAfee’s 2023 research found that a voice could be cloned from about three seconds of audio. Germany’s BSI gives a more detailed picture: high-quality voice models typically need several hours of training audio, helper datasets cut that to minutes, and experimental methods work with seconds at a cost in quality (BSI). For a crying voice on a bad line, rough quality is enough.
Payment requests follow a pattern. The FTC lists wire transfers, cryptocurrency and gift cards as red flags, because the money is hard to recover (FTC consumer alert). Some variants send a courier to pick up cash at the door.
Variants beyond the family emergency
The crying-relative call gets the headlines, but the same cloned voice turns up in other scripts. The FBI lists two in its 2024 warning: short clips that mimic a relative to demand ransom in a staged kidnapping, and cloned audio used to get into bank accounts that rely on voice authentication.
Fake kidnappings are the cruellest version. Callers play a scream or a plea in the cloned voice, then demand payment while keeping you on the line so you can’t check. The defence is the same: get someone else to call the supposed victim while you stay on the phone.
Voice authentication is a concern for anyone whose bank lets them log in by saying a passphrase. If your bank offers this, ask whether it can be replaced with an app confirmation or a PIN. Also ask which other checks protect a phone login. The answer tells you how much the voice alone is worth to an attacker.
Why you can’t rely on your ear
People are poor at hearing the difference. For a UCL study published in PLOS ONE in 2023, researchers played clips to 529 participants, who identified synthetic speech only 73 percent of the time, in English and in Mandarin, and training barely helped. Those listeners knew some clips were fake and were paying attention. On a real call you are surprised and frightened, and the voice belongs to someone you love.
There are audio tells, and our guide on how to detect an AI voice lists them. But the honest advice is not to bet your savings on spotting a metallic undertone. Build a procedure that works even when the voice is perfect.
What to do during the call
- Hang up. Say you will call back. A real emergency survives a two-minute delay.
- Call the person directly on the number saved in your phone. If they don’t answer, call their partner or their workplace.
- Ask for the safe word, if your family has one. Scammers cannot know it.
- Never pay on the first call. No legitimate police force or lawyer collects bail by gift card, crypto or a courier at your door.
- Tell someone. Scammers ask for secrecy because a second person usually breaks the spell.
The FBI’s December 2024 alert recommends exactly this combination: “Create a secret word or phrase with your family to verify their identity” and verify callers by hanging up and calling back (FBI IC3). Setting one up takes five minutes. Our guide to a family safe word covers how to pick one that a scammer can’t find online.
If you have already paid
Speed matters more than embarrassment. Call your bank’s fraud line straight away and ask for the payment to be recalled. Card payments and bank transfers sometimes can be stopped within hours. Crypto and gift cards rarely can, but report the gift card numbers to the issuer anyway.
Then report it. US victims should use ReportFraud.ftc.gov and ic3.gov. UK victims report to Action Fraud. Keep the call log, the number that called and any messages. Expect follow-up calls from “recovery agents” promising to get your money back for a fee. Those are the same people, or their associates.
How to make your family a harder target
Three habits cover most of the risk:
- Agree a safe word with parents and grandparents, and with teenagers, and practise it once.
- Set social media accounts that post family videos to private, as the FBI advises.
- Replace your personalised voicemail greeting with the default one. It is a free voice sample for anyone who calls.
Businesses face the same technique with higher stakes, usually a fake CEO asking for an urgent transfer. That variant is covered in deepfake CEO fraud. Companies that want verification built into their phone or chat systems can talk to a studio such as OnlineBuilders. For everyone else, the next step is the safe word. Agree it this week, before you need it. You will find more guides on phone fraud in our voice and scams section.
Questions people ask
How much audio does a scammer need to clone a voice?
McAfee's researchers reported in 2023 that three seconds of audio can be enough for a usable clone. Germany's BSI notes that convincing, high-quality clones still benefit from minutes or hours of material, but a short voice note or video can produce something good enough for a panicked phone call.
Can the scammer call from my relative's number?
Yes, caller ID can be spoofed, so the screen may show a familiar name or number. That is why the call-back matters: you dial the number from your own contacts, which reaches the real phone.
Should I stop posting videos of my family?
Making accounts private and limiting public voice clips reduces the material available, which the FBI recommends. It won't remove the risk entirely, so a safe word and the call-back habit are the more dependable protection.
Where do I report a voice cloning scam?
In the US, report to the FTC at ReportFraud.ftc.gov and to the FBI's IC3. In the UK, report to Action Fraud, or Police Scotland on 101. Tell your bank immediately if you have already paid.
