They cloned her voice in 3 seconds. Then called her mother.

Voice cloning: the family emergency scam

A grandmother receives a phone call. It's her grandson's voice — panicked, crying. "Grandma, I'm in jail. I need bail money. Please don't tell mom and dad." She wires $15,000. It wasn't her grandson. It was an AI clone of his voice, generated from a 3-second clip pulled from his Instagram story.

This scenario is happening thousands of times. AI voice cloning has become commodity technology — available to anyone for under $5/month. The voice is near-perfect, capturing tone, accent, and speech patterns.

Deepfake video calls

In February 2024, a finance employee at a multinational firm joined a video call with the company's CFO and several colleagues. The CFO instructed a $25 million wire transfer. Every person on the call was a deepfake — AI-generated video puppets mimicking real executives in real-time. This is increasingly a business fraud threat, not just a consumer one.

Deepfake video is rapidly improving. Current limitations — slightly unnatural blinking, audio-video sync issues, difficulty with hand gestures — are being solved by each new model generation. Within 12-18 months, real-time deepfake video will be indistinguishable from reality.

AI-generated phishing at scale

Traditional phishing relied on volume — send millions of bad emails and hope someone clicks. AI changes the equation entirely:

• ▸

  The FBI logged more than 22,000 AI-related fraud complaints in 2025, with losses above $893 million.

• ▸

  AI writes grammatically perfect, personalized emails in seconds.

• ▸

  It scrapes your LinkedIn to reference your real job, company, and recent activity.

• ▸

  Polymorphic AI generates unique variations of each email, defeating pattern-based spam filters.

• ▸

  AI chatbots can carry on convincing real-time conversations, replacing human scammers.

The "Nigerian prince" era is over — this is industrial-scale, AI-augmented fraud.

How to protect yourself from AI scams

The defenses are behavioral, not technological:

1. 1.

   Create a family safe word for verifying phone calls — no AI can know it.

2. 2.

   Never authorize financial transactions based solely on phone or video calls — verify through a separate channel.

3. 3.

   Be skeptical of any urgent request for money, regardless of who appears to be asking.

4. 4.

   Limit personal audio and video on public social media — scammers mine these for cloning material.

5. 5.

   For businesses: implement verbal verification protocols for wire transfers. No amount should be authorized by voice or video alone.

6. 6.

   Keep your voicemail greeting generic — detailed greetings provide more audio for cloning.

Why the check matters more than the tell

Every scam on this page is the same event described three ways. The artefact became perfect. The voice is right, because it was cloned from three seconds of a story. The face is right, because it is rendered in real time. The email is right, because the model wrote it and read your LinkedIn first. This is the one category of fraud where the tells were not sloppy and then improved by accident. They were engineered out on purpose.

Which is why the detection advice on this page carries a stated expiry date, and you should notice that it does. The blinking, the audio sync, the strange hands: we say above that each new model generation is solving them. Advice built on those tells stops working on someone else's release schedule. The safe word and the separate channel do not, and the reason is worth being precise about.

So the answer is not whether they pass the check. It is what they do when you ask for it. A real grandson is not insulted by a safe word; he is relieved his family has one. A real finance director does not object to a twenty-five million dollar instruction being confirmed on a number you chose, because that is what the control is for and it protects them too. Anyone who argues against the verification has told you something the voice and the face cannot tell you, and it is the only thing on the call that is genuinely theirs.

Listen specifically for the reason you cannot verify. Do not tell mum and dad. I am calling from a different number, do not use my usual one. There is no time, it has to be now, stay on the line. Every one of those exists for a single purpose: to stop you opening a second channel. That instruction is the one part of an AI attack that cannot be synthesised away, because the operation fails outright the moment you use a route it does not control. The model can generate a perfect voice. It cannot generate a good reason for you not to hang up and call back.

None of this means stop looking. The tells above are still worth knowing, plenty of operations are still using last year's models, and recognising one saves you the trouble. The claim is narrower. On this page more than any other, inspecting what you are shown is no longer sufficient on its own to decide, because what you are shown is exactly the part that got automated. Use the tells to recognise it. Use the callback, and the reaction to the callback, to decide. The longer version of this argument is in why scammers sound official.

If an AI scam has already cost you money, act fast with our recovery guide, then see where to report it in your country.