Fake Voices and Faces: What the Technology Can Do Now

Two checks that people have relied on their whole lives stopped being reliable in the space of about three years: recognising a familiar voice, and seeing someone's face on a video call.

This is not a reason for alarm, and it is a reason to change one specific habit. Understanding what the technology can and cannot do is most of the defence.

For a practical workplace perspective, Monitask explains employee PC activity tracking and how it is used.

What it can do

Clone a voice from a few seconds of audio. AI can clone anyone's voice from just a few seconds of audio found on social media, voicemail, or video. The result is convincing over a phone line, particularly when the person is supposedly upset.

Put a face on a live video call. Real-time deepfake software allows a criminal to appear on a video call as a bank executive, financial advisor, or romantic interest, and the technology now runs on a standard laptop.

Write flawless messages. AI-generated phishing eliminated the grammar mistakes that once helped people spot fakes.

Generate a face that does not exist, for a profile photograph that cannot be reverse-searched because it has never appeared anywhere.

What it is used for

FBI data for 2025 recorded 22,364 complaints citing AI as a tool in fraud, with losses exceeding $893 million across all ages. For people aged 60 and over, AI-related fraud accounted for $352 million — and only in cases where the victim realised AI was involved.

The main uses:

Family emergency calls, where the voice is the whole persuasion. See the family emergency call.

Bank and authority impersonation, where a video call now backs up the phone call.

Romance fraud, where a video call was previously the standard proof of identity. See romance fraud.

Fake endorsements — a familiar public figure appearing to recommend an investment. These are entirely fabricated and are among the most-viewed fraudulent advertisements online.

What it still cannot do well

Useful to know, and worth treating as a secondary check rather than a defence.

Handle the unexpected. Real-time face replacement struggles with a hand passing in front of the face, a full profile turn, standing up and moving, or an object held close to the camera. Asking for any of these on a video call frequently breaks it visibly.

Sustain a long unscripted conversation. Voice clones handle short emotional exchanges. Extended back-and-forth, with interruptions and specific questions, degrades.

Know things. This is the real weakness. The technology can reproduce a voice; it cannot supply what only the real person knows.

Be verified at a second location. It cannot answer the phone when you call back on a number you already have.

Why the answer is a procedure, not a keener ear

The instinct is to listen harder. That is the wrong response, and it will fail — the technology is improving and human detection is not.

Replace the judgment with a procedure. Something that does not require you to assess anything.

A family word. Agreed in advance, never written online, never used casually. Anyone claiming to be family in an emergency is asked for it. A cloned voice does not have it.

A callback, always. Any request involving money or details gets ended and returned on a number you already had. This defeats every version of this simultaneously, regardless of how good the fake was.

A second channel. If a video call asks for something significant, confirm by a different route — a text to their usual number, a call to their office switchboard.

These work because they do not depend on detection. You do not have to be right about whether the voice was real. You just have to follow the step.

Practical points

Reduce what is available, without going into hiding. Public videos with clear audio are the raw material. This is not a reason to delete a life online — it is a reason to have the family word, which costs nothing and works regardless.

Voicemail greetings are a common source of sample audio. A generic greeting rather than your own voice is a small, free adjustment.

Talk to family about it. Especially with grandchildren, whose audio is most available and who are most often impersonated. The conversation is short: here is what it is, here is our word.

Assume a video call proves nothing when money is involved.

Recognising a fake endorsement

If a well-known person appears to be recommending an investment or a health product:

Check the source. A genuine endorsement appears on the person's own verified accounts and in mainstream coverage. A fabricated one appears only in the advertisement.

Look at the destination. Fabricated endorsements lead to sign-up forms, trading platforms, or a "limited opportunity".

Apply the general rule. No legitimate investment is promoted through a social media advertisement featuring a celebrity and a countdown.

If you have been targeted

Report it, and say specifically that AI was involved — the agencies are tracking this category separately and the data is incomplete.

FTC: reportfraud.ftc.gov · FBI: ic3.gov · National Elder Fraud Hotline: the AARP Fraud Watch Network

See the first two hours after you realise for what to do immediately.

The thing worth carrying

The old checks — a familiar voice, a face on a screen, a well-written message — no longer prove what they used to. That is a genuine change and it is worth knowing.

But the replacement is simpler than the problem: hang up and call back on a number you already have. It requires no expertise, works against every version of this, and will keep working however good the technology gets.

For additional public information on consumer fraud and scam prevention, see Federal Trade Commission.