What is being faked
A deepfake is audio, video or an image that has been generated or altered by software so that it appears to show a real person saying or doing something. Voice cloning is the audio version: software produces speech that sounds like a specific person.
In scams these tools do not create a new kind of fraud. They add a convincing face or voice to familiar stories, such as “I’m in trouble and need money” or “this is your manager, please arrange an urgent payment”. The underlying scheme still depends on urgency and an unusual request.
How a synthetic voice is made
Voice cloning systems are trained on recordings of speech. They learn the patterns of a voice (pitch, rhythm, accent, the way sounds are shaped) and can then produce new sentences in a similar voice from typed text, or convert one person’s speech into another’s. This is a form of generative AI, built on the same general principle as other tools that learn from training data.
How much audio is needed varies by system. Some products advertise that a short clip is enough for a rough imitation, and quality generally improves with more and cleaner material. Public videos, voicemail greetings and social media posts are all possible sources. The result can range from obviously robotic to quite convincing, particularly over a phone line where audio is already compressed and imperfect.
How video fakes are made
Video deepfakes typically work in one of two ways. One swaps a face onto another person’s footage. The other animates a still image or a face so it appears to speak particular words. Similar technology underlies many image generators, which create pictures from descriptions.
Quality differs widely. Pre-recorded fakes can be refined and edited. Real-time fakes during a video call are harder to do well and often show glitches, but the technology is improving and it is unwise to depend on spotting flaws.
How scammers use them
Reported uses generally fall into a few categories, though how common each is can be hard to measure:
- Family emergency calls. A caller who sounds like a child, grandchild or friend claims to be in trouble and asks for money or a code, often insisting on secrecy.
- Workplace payment requests. A voice message or call that appears to come from a senior colleague asks for an urgent transfer or for gift cards.
- Fake endorsements. Video of a well-known person seemingly promoting an investment or giveaway. These are common in investment scam patterns.
- Identity checks. Fakes used to get past voice-based or video-based verification at a service.
A voice does not need to be perfect. A distressed, crackly, hurried call hides many flaws, and the emotion of the moment reduces suspicion.
Why spotting flaws is not enough
Advice on detecting fakes often lists clues: unnatural blinking, odd mouth movement, flat emotion, strange background sounds, unusual phrasing. These can help sometimes. But they are unreliable. Tools improve, poor phone audio hides problems, and a stressed listener is not in a good state to analyse a voice. Automatic AI detection tools exist, but they are also imperfect and can give false confidence in either direction.
A better approach is to treat the content of the request, not the sound of the voice, as the thing to check. If someone who sounds like a person you know asks for money, secrecy or a code, the voice is not evidence.
Practical ways to verify
None of these is foolproof, but together they make a convincing fake much less useful to a criminal.
- Hang up and call back. End the call, then ring the person on a number you already have saved. If the call was a trick, the real person will not know about it.
- Use another channel. Message them, or call someone close to them, to confirm the situation.
- Agree a code word. Families and small teams can pick a word or question that only they would know, and use it for any urgent money request. Choose something that cannot be found on social media.
- Ask something specific. Questions about recent shared events can reveal an impostor, though a scammer with personal information from social media or data breaches may answer some correctly.
- Slow the process. For payments at work, require a second person or a different channel to approve any unusual request, however senior the apparent sender.
- Be cautious with unusual payment methods. Requests for gift cards, cryptocurrency or transfers to new accounts deserve extra suspicion whatever the voice says.
A caller who becomes angry or more panicked when you propose hanging up to check is itself a warning sign.
Reducing the material available
You cannot fully prevent your voice or face being copied, since many people appear in public videos. Still, a few steps may reduce exposure. Consider who can see videos and voice notes you post, and review privacy settings on social media. Avoid posting long, clear recordings publicly if you do not need to. Avoid answering unknown callers with a long conversation if you suspect a recording attempt, although the extent to which this is a real risk is difficult to judge.
If you think you have been targeted
If you already sent money or shared a code, act quickly: contact your bank, change affected passwords and report it. The broader steps are covered under scam recovery steps. If you only received a suspicious call, it is still worth warning relatives or colleagues, since the same approach is often repeated against several people.
Synthetic media will probably keep improving. The reliable defence is a habit rather than a detection skill: verify through a separate, known route before acting on any urgent request, no matter how real it sounds.