Deepfake Self-Defence
Voice clones need about three seconds of audio. Voice clones need about three seconds of audio. Sourced, dated, and revised when the numbers move.
Every other room in this museum teaches you to recognise synthetic things. This one assumes you will not recognise it — because the people who lose money to voice clones are not careless, they are correct about everything except one voice. So the defenses here do not rely on your ears. They rely on a word, a callback, and a pause.
How little it takes
In 2023, Microsoft researchers demonstrated a model that could reproduce a person's voice — pitch, cadence, speaking style — from roughly three seconds of recorded audio. Not a studio session. Three seconds: a birthday clip, a voicemail greeting, a few words in someone else's livestream.Source: Microsoft Research, VALL-E — “Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers”, January 2023. Microsoft's own description: high-quality personalised speech from “only a 3-second enrolled recording of an unseen speaker”. Trained on 60,000 hours of speech from over 7,000 speakers; code not released. A 2024 successor claimed human parity.
That is the entire technical story, and it is why "just listen carefully" stopped being advice. The US Federal Trade Commission put it plainly in a consumer alert as early as March 2023: scammers are using AI to sharpen family-emergency schemes, and you should not trust the voice.Source: FTC Consumer Alert, "Scammers use AI to enhance their family emergency schemes," March 2023.
The call, beat by beat
Nearly every version runs the same five beats. Learning the shape is more useful than learning the symptoms, because the shape does not change when the technology improves.
Someone you love, in distress. An accident, an arrest, a hospital, a hostage. The voice is right, because the voice was cloned from something they posted.
A second speaker takes over: a lawyer, an officer, a doctor, a manager. Authority arrives to do the asking, so the "loved one" never has to answer questions.
Do not tell anyone. It is embarrassing, it is under investigation, it will make things worse. Secrecy exists for one reason: it prevents the one phone call that ends the scam.
It must happen now, within the hour, before the court closes. Urgency is not a detail of the story. Urgency is the attack — it exists to prevent verification.
Wire transfer, gift cards, cryptocurrency, or a courier who comes to the door for cash. Every one of them is chosen because it is hard to reverse.
It scaled to boardrooms
In January 2024, a finance employee at the engineering firm Arup joined a video call with people who looked and sounded like the company's chief financial officer and several colleagues. Every other participant was synthetic. Fifteen transfers followed, totalling about US$25.6 million (HK$200 million). The Financial Times broke the story in May 2024; Arup confirmed that fake voices and images were used and that none of its internal systems were compromised. The attackers built the impersonations from footage that was already public.Sources: Hong Kong Police briefing, February 2024 (15 transfers to five accounts, HK$200m ≈ US$25.6m); Arup identified as the victim in May 2024, confirming to the Financial Times and Fortune that fake voices and images were used and that no internal systems were compromised.
The employee did what training says to do — he was suspicious of the initial email. The video call is what dissolved the suspicion. That is the lesson worth carrying home: a face and a voice are no longer verification, at any scale, in any setting.
The pattern now has a national price tag. The FBI's 2025 Internet Crime Report — the first in the agency's 25-year history to break out artificial intelligence as its own category — logged 22,364 AI-related complaints and $893,346,472 in reported losses, including roughly $30M in business email compromise with a confirmed AI component and about $13M in deepfake job-interview schemes. Americans over 60 reported around $7.7 billion in total fraud losses — the steepest rise of any age group.Source: FBI Internet Crime Complaint Center, 2025 Internet Crime Report, published April 2026.
Four defenses that still work
None of these require you to detect anything. That is the point.
One word or short phrase, known only to your household. Anyone calling in an emergency must be able to say it. Choose something unguessable and unposted — not a pet's name, not a street you have written about. Say it out loud to each family member; never store it in a phone note, a group chat, or anywhere a breach could reach. Then tell the eldest members first, in two sentences: if anyone calls sounding like us in trouble, ask for the word, then hang up and call us back.
Any unexpected request for money or secrecy ends the same way: you hang up and call the person back on the number already saved in your phone. Never a number the caller gives you. This works in both directions — a boss, a bank, a government office. Legitimate callers have never once objected to being called back.
Stop treating panic as evidence that something is real. The script is engineered to keep you from pausing, because a five-minute pause is fatal to it. Real emergencies survive verification. If someone is working hard to stop you from checking, that effort is the finding.
Wire transfer, gift cards, crypto, or a courier for cash — the channel is chosen for irreversibility, not convenience. No hospital, court, or police department takes bail in gift cards. When the request reaches this beat, the beat itself is the answer.
A fifth, quieter defense: shrink the sample. Public video and audio is the raw material. Locking down who can see family posts will not make cloning impossible, but it makes your household a slower target than the next one.
- Do not confirm names or relationships. Answering "which grandchild?" hands the script its next line.
- Do not send anything. Do not stay on the line.
- Hang up. Call the person directly on their saved number. If they do not answer, try a text, another relative, or wherever they are supposed to be.
- If money has already moved, call your bank immediately — speed is the only lever left — and report it to the FTC at reportfraud.ftc.gov (US) or your national fraud reporting line.
Voice synthesis is not the villain here. The same technology restores speech to people who have lost it, narrates books, and crosses languages. What changed is that impersonation became cheap, and the institutions we trust to verify identity have not caught up. Until they do, the defense is a household habit, not a gadget. Set the code word tonight; it takes fifteen minutes and it outlives every model release.