Skip to content
Back to Blog
5 min read

AI Voice Cloning Can Now Fool Human Listeners

AI-enabled fraud grew 1,210% in 2025. How voice cloning scams work, and the verification habits that protect families and organisations.

  • AI security
  • voice cloning
  • deepfakes
  • cybersecurity
  • fraud prevention

A scammer needs about three seconds of your voice to clone it with 85% accuracy, according to McAfee research. Researchers say voice cloning has now crossed the "indistinguishable threshold": ordinary listeners can no longer reliably tell a cloned voice from a real one.

AI-generated deepfakes and voice cloning scam illustration
AI-generated deepfakes and voice cloning scam illustration

The fraud numbers have moved with the technology. Cybersecurity firm Vectra AI reports that AI-enabled fraud grew 1,210% in 2025, against 195% growth in traditional fraud. Major retailers say they receive more than 1,000 AI-generated scam calls a day, and projected losses from these attacks could reach $40 billion by 2027.

How the scam works

Scammers collect audio from videos you post on social media, or they call you posing as a survey company and record your answers. Five to ten seconds of normal conversation is enough to build a convincing clone.

The clones are good. Researcher Siwei Lyu says today's tools reproduce a voice "complete with natural intonation, rhythm, emphasis, emotion, pauses and breathing noise." The small flaws that used to give synthetic audio away have mostly gone.

The most common attack targets families. You get a call from what sounds exactly like your son or your mother, saying they have been arrested or had an accident and need money now. The urgency is designed to stop you checking. One in four people has now come across an AI voice scam, and 77% of victims lost money.

Organisations face the business version. An attacker clones a senior executive's voice and calls the finance team with an urgent transfer request. If the voice sounds right, staff tend to comply.

Why the usual warning signs no longer help

Most fraud awareness training taught people to spot mistakes: poor grammar in a phishing email, a call from an unknown number. AI writes clean text, and the caller now sounds like your own manager. Training built on catching obvious tells is losing its value.

The volume is also rising fast. Cybersecurity firm DeepStrike estimates that deepfakes online grew from about 500,000 in 2023 to 8 million in 2025, an annual growth rate of nearly 900%. Free tools can produce a convincing clone in minutes.

Lyu expects the next step to be real-time synthesis, with interactive AI actors whose "faces, voices and mannerisms adapt instantly to a prompt." Video calls, which many people currently treat as safer than a voice call, will face the same problem. Longer-term defences will rely more on infrastructure, such as cryptographic signing of genuine media and forensic tools that detect synthetic audio and video.

Protecting your family

The most effective defence costs nothing. Agree a family safe word, something that cannot be guessed, so avoid street names, pet names and school names. If someone calls claiming to be a relative in trouble, ask for the word before doing anything.

Two other habits help. Be careful about how much video of yourself you post publicly, since that is where scammers find voice samples. And when a call asks for money, hang up and call the person back on a number you already have saved. Never call back the number that called you, and never send money on the strength of a phone call alone.

In the Gulf, family ties are strong and people trust a familiar voice. These scams are built to exploit exactly that sense of obligation in an emergency, which is a good reason to set up the safe word this week rather than later.

Protecting your organisation

For organisations, the controls matter more than awareness sessions. Any significant payment should need two approvals and written confirmation through a verified channel. Voice recognition should never be enough on its own.

Sensitive requests, such as granting access or authorising a transfer, should trigger a callback to a contact method your team already holds, never one supplied in the request. Executive teams can agree pre-shared code phrases for emergencies, the business equivalent of the family safe word.

Technical tools are improving. Voice biometric anomaly detection can flag synthetic audio, and Network Detection and Response tools can spot infrastructure linked to AI fraud campaigns. They add a layer on top of staff judgement, and staff still make the final call.

If you do one thing after reading this, ask your finance team a direct question: if the CEO called tomorrow asking for an urgent transfer, what written check would they require before sending it? If there is no clear answer, write that procedure first.

Sources: