A call in your child's voice: what to do about it
A voice can be faked from a short recording taken off social media. Telling the difference by ear during the call is impossible — and that is the whole point of the problem. Below is the one defence that works, and it is set up in one evening with no technology at all.
A voice can be faked from a short recording taken off social media. Telling the difference by ear during the call is impossible — and that is the whole point of the problem. Below is the one defence that works, and it is set up in one evening with no technology at all.
If there is no time to read: five points
- Agree a code word with your family. One word only your own people know.
- Ask for it on any call requesting money — even if the voice is definitely your relative's.
- The word must not be guessable: not the dog's name, not a maiden name, not the street you grew up on.
- Never write it down — not in messages, not in notes. Only aloud and only to your own people.
- No answer to the word — hang up and ring back yourself on the number you know.
Why the old ways of checking stopped working
Protection used to rest on faking a voice being difficult. That is no longer so.
A short recording is enough. A voice message in a messenger, a story, any video where the person speaks. Most people have recordings like that posted publicly.
Telling the difference by ear is impossible. Not because you are inattentive. The timbre, the manner and the intonation match — and under stress you are not listening, you are reacting.
And most of all: the call is built so that you do not think. The scenario is always the same — trouble, urgency, tell nobody. A person in a panic does not analyse a voice.
Hence a conclusion that changes the approach. Learning to recognise a fake is pointless. What is needed is a check that does not depend on how similar the voice is.
Who I am and what I know about voice deepfakes
I have been working with AI since 2019. I know how speech synthesis works — and that is exactly why I am not suggesting you "listen carefully to the intonation".
The technology cannot be stopped or undone. But it can be got round by a method older than any technology.
The only thing that works: a code word
Agree with your family on one word only your own people know.
On any call requesting money or urgent help you ask for that word. No answer and the conversation is over.
Why this works when nothing else does. A voice can be faked. A number can be spoofed. Circumstances can be found on social media. A word that is written nowhere and has never been said publicly cannot be obtained from anywhere.
It is not a technical defence, and that is its strength: it cannot be got round by a better model.
How to choose the word
Four rules, all because the other side knows how to search.
Not publicly connected to you. Not the dog's name, not your mother's maiden name, not the street you grew up on, not the name of your school. All of that is found in ten minutes.
Not logical. A good code word is one that cannot be guessed from you. "Stool", "watermelon", "November" are better than the pet's name.
Simple, so it can be recalled in a panic. That is the key point. During the call the person is frightened — a complex phrase will not come to mind. One short word.
One for the whole family. Different words for different people will not be remembered, which means they will not work.
How to agree it. Gather the family, say the word aloud, ask them to repeat it. Five minutes over dinner.
How to use the code word
Ask for it always, not only when something feels suspicious. The habit matters more than vigilance: things feel suspicious at times other than when there is danger.
The phrasing is simple: "say the code word". Do not hint, do not prompt, do not ask "remember what we agreed?"
Do not prompt with options. "It is our watermelon, isn't it?" — and you have said it all yourself.
No answer, or they start explaining why they cannot — hang up. A real relative will not take offence, and if they do, you can explain afterwards.
And an obligatory second action: ring back yourself on the number you have saved. Not the one that called.
What must not be done
Do not write the word down. Not in phone notes, not in messages, not on a piece of paper by the monitor. Only in your heads.
Do not say it in front of others and do not send it in a messenger — even to your own people.
Do not use it as a password for anything.
Do not change it often. A forgotten code word is worse than none: at the moment it is needed it will not work for either side.
Who to do this with first
With elderly relatives. They are the ones called most often, and the ones least likely to have heard that voices can be faked.
Build the conversation around something other than technology. Not "Mum, AI can fake voices now" — that sounds like something distant. But rather: "if somebody rings you in my name and asks for money, ask for this word. I will always answer."
With children. They get called less often but believe more readily. And explaining it to them is easier: children take a code word as a game rather than as a security measure.
What to do if the call has already happened
If you sent nothing — simply hang up and ring your relative yourself. And tell the family there was an attempt: that is the best reason to finally agree a word.
If money has gone — ring the bank immediately and block the transfer, then the police to report it. Speed matters more here than anything else.
And do not blame yourself. Calls like that are built professionally, on panic and lack of time. People of any age and any caution fall for them.
What to do today to protect your family
One action, five minutes.
Gather the family — over dinner or in a group chat, but say the word aloud rather than in text. Agree on one word.
Check everybody remembered: ask each of them to say it the next day.
That is all. No technology, no applications, no settings. The only defence that works against a faked voice costs five minutes of conversation.
Why a code word is more reliable than "I would recognise the voice". A voice is faked from a short recording — a few sentences from a voice message or a video is enough. The technology stopped being rare long ago: it is openly available and costs pennies. Recognising a fake by ear in a stressful situation works for almost nobody — which is exactly why what works is not recognition but a check on something that was not in the recording.
What to read next
Where does what I write to an AI go — what happens to your data and what not to upload.
Where to start: one evening, one task — if after this subject you want to find out what AI is useful for.
How to verify a model's answers — the other side of trust: accuracy rather than safety.
A voice clone instead of a narrator — how a voice clone is made in practice, so you understand what exactly is being faked.
Try it yourself. Registration is free; three days of the Basic plan switch on with one button, no card needed. Basic includes 200+ AI assistants and the programme's courses.