What “synthetic media” means
A deepfake is an image, video or audio clip generated or altered by software so that it appears to show a real person saying or doing something. Voice cloning is the audio version: a system that produces speech in a voice that resembles a particular person. The same family of methods also produces entirely invented people who never existed.
These tools are not magic and not perfect, but they have become accessible enough that sounding or looking like someone is no longer reliable evidence that it is them. That changes how verification needs to work.
How a synthetic voice is made
Most voice cloning systems are a form of machine learning. In broad terms, a model is trained on recordings of speech so it learns the patterns that make up how people sound: pitch, rhythm, accent, and the way sounds blend together. Some systems are trained on many speakers and then adapted to a specific voice using a smaller sample of that person’s speech. Others work from text: you provide words and the model reads them aloud in the chosen voice.
Two practical points follow from this:
- Samples matter. The more clean audio there is of a person, the easier it generally is to imitate them. Public videos, voicemail greetings and recorded talks can all serve as source material, though how much is needed varies by system and is hard to pin down.
- Quality varies. Some cloned voices are convincing in a short phone call and fall apart in longer conversation, with odd pacing, flat emotion or strange handling of unusual words. Others are good enough to fool family members. Do not count on hearing the flaws.
Real-time versions also exist, in which one person speaks and the output is converted into another voice as they talk. Whether a given call uses that or a pre-recorded clip, the listener’s experience is similar.
How fake faces and videos are made
Image and video fakes use related techniques. Some replace one person’s face with another in existing footage. Others generate new footage from scratch. Many rely on the same families of neural networks that power image generators, which learn from large collections of pictures what faces look like and how they move.
Typical weaknesses in lower-quality fakes include blurring around the edges of the face, inconsistent lighting, unnatural blinking or lip movement, and odd detail in teeth, hair, ears or jewellery. The caution is that these cues fade as the technology improves, so they are weak evidence in either direction. A clean-looking video is not proof of authenticity, and a slightly odd real video can look suspicious for innocent reasons such as compression.
Where it is misused
Misuse tends to fall into several categories.
- Family emergency calls. A caller who sounds like a relative claims to be in trouble and asks for money quickly and quietly. This is a classic scam script that a cloned voice makes more persuasive.
- Impersonating executives or colleagues. A voice message or video call appears to come from a senior person and asks for an urgent payment or a change of bank details.
- Bypassing voice checks. Some services have used voiceprints as an identity check, and cloned audio can be used to try to defeat them. This is one reason voice alone is a weak proof of identity.
- Fake endorsements and misleading clips. Fabricated videos can promote fraudulent investments or spread false claims, which overlaps with how online misinformation travels.
- Harassment and non-consensual imagery. This is a serious harm, and many places have been introducing or updating laws about it. If you or someone you know is affected, reporting to the platform and to local authorities or support organisations is the sensible route.
Underneath most of these is the same pattern as other fraud: the technology provides the credibility, and ordinary social engineering supplies the urgency.
Why detection tools are not the answer on their own
It is tempting to look for software that tells real from fake. Such tools exist, but they can be wrong in both directions, and the problem resembles the one with AI detectors in text: results are probabilistic and can be fooled. Treat any automatic verdict as one input rather than a ruling. For everyday situations, a procedure you control is more dependable than a score.
Practical ways to verify a caller
The most useful defences do not depend on spotting flaws in the audio.
- Hang up and call back on a known number. Use the number saved in your contacts, not one the caller gives you. This single step defeats most impersonation, because the attacker cannot receive calls to the real person’s number.
- Agree a family code word. A shared word or question that only close family know can be used when someone calls with an emergency. Choose something not posted online, and do not treat it as a guarantee, since people can be pressured into revealing it.
- Ask a question only the real person could answer, and avoid ones answerable from social media.
- Slow down when money or codes are involved. Requests for urgent transfers, gift cards, crypto wallets or one-time codes deserve a second channel of confirmation regardless of how familiar the voice is.
- Use a second channel inside organisations. Payment changes should need confirmation through a separate, established process, not just a message or call.
- Reduce what is publicly available. Limiting public posts of your voice and your children’s voices is a modest step, but it is not a complete protection. Consider the digital footprint you and your family leave.
Keeping a sense of proportion
Most scam calls still use no cloning at all, and plain scripts work well enough for criminals. At the same time, it is reasonable to assume that a familiar voice or face can be imitated. The healthy response is not to distrust everyone but to attach important decisions to verified channels rather than to how real something sounds.