How to spot a deepfake: count the hands, shadows and faces

Nearly every giveaway that still works is a wrong number. We found that out the hard way — by trying to make fakes a child could actually spot.

Published October 22, 2026 · Wisesoft.ai · 5 min read

A deepfake is a fake photo, video or voice recording, made by AI, that shows a real person doing or saying something they never did. It used to take skill and expensive software. Now anyone can make one in minutes, just by typing a description or uploading a few photos of the person.

The short answer: count things

A deepfake’s giveaway — the small mistake that shows a picture is fake — is almost always a wrong number now: three hands in a two-person handshake, two shadows under one person, the same stranger four times in one crowd. The giveaways that have stopped working are the ones people still teach: squinting at skin texture, hunting for a soft hairline, asking whether the light “feels right”.

What happened when we tried to make a fake

To build the Deepfake Detective level in FutureMinds, we needed photos with a deliberate mistake in them — made the way anyone makes a fake now, by typing a description and letting an AI draw it.

Bodies were easy: ask for a third hand and you get one, first try. But ask for a shadow falling the wrong way and the AI quietly fixed it, putting the shadow exactly where the sun would. Same with a mirror that didn’t match the room.

So we stopped asking for things that were physically wrong, and asked for things there were too many of:

The five things to count

• Hands, and fingers on hands. The most reliable of the lot. The AI has learned what a hand looks like without ever learning how many a body gets.
• Arms and legs where two people touch. A handshake, an arm round a shoulder. Where two bodies meet, the AI has to work out which limb belongs to whom — and often just draws another.
• Shadows. Not the direction — the number. One person under one sun gets one shadow.
• Faces that repeat. Backgrounds are the part an AI takes least care over, so it fills them by copying. Four identical people mid-stride in matching clothes is a crowd made out of one person.
• Any writing in the picture. The exception to the counting rule, and still the best giveaway there is. A shop sign, a name badge, a page of a book: the AI copies the shape of letters without knowing what a word is. Zoom in on writing before trusting a photo.

Don’t turn it into a checklist

Ten of the thirty-six pictures in our level have nothing wrong with them, and “nothing wrong here” is the right answer. That is deliberate. A list of giveaways quietly teaches you to always find something.

A child who decides every picture is fake hasn’t become harder to fool — they’ve just lost the ability to tell.

When the fake is of someone real

Everything above answers one question: is this picture real? There is a second one — has this person been faked? — and it is the one that actually turns up, usually in a group chat. (The first question gets the full treatment in how to tell if a photo is AI-generated.)

Here, spotting stops mattering. Your child knows a doctored picture of them isn’t real — they weren’t there. What matters is that it’s spreading. And it is more common than most parents think:

• Thorn surveyed 1,200 American teenagers and young adults in autumn 2024: one in eight personally knew someone targeted by fake nude pictures made while they were under 18, and one in seventeen had it done to them.
• Common Sense Media surveyed 1,314 American teenagers in autumn 2025: nearly half had seen AI-made explicit pictures, and of that half, one in four had seen one of themselves or someone they know.

Four things to do, and the first is the one almost everybody gets wrong:

• Do not forward it. Not even to warn someone. Every forward is another copy somewhere new. Screenshot the conversation — who sent it, when, to which group — not the picture.
• Report it to the platform, and use the right words. Since May 19, 2026, US law gives platforms 48 hours to remove a reported fake nude — AI-made ones included. Report it as “an intimate image shared without consent”; that goes down a faster route than a general complaint.
• If a child is involved, start at Take It Down. It’s free, run by the National Center for Missing & Exploited Children, and the picture never leaves your phone — only a short code that identifies it is sent, so platforms can block copies without anyone seeing it.
• Say out loud whose fault it is. A targeted teenager usually expects to be blamed for a photo they never took. Say the opposite, first.

Faked voices

None of this helps with a voice on the phone, and listening for clues barely works anymore. What does work is a phrase only your family knows, asked for rather than offered. Our short lesson shows how: create your family password. AI scams covers the phone call itself.

The catch

These tricks won’t last. To build the level we had to use an older AI on purpose — newer ones make fewer of these mistakes. So teach the counting, which works today, but also the question that never goes out of date:

A picture with no source, built to make you feel something strongly, deserves suspicion no matter how clean the fingers look.

Two minutes, tonight

Open image search for something completely ordinary — “people in a coffee shop” will do — and scroll it with your child under one instruction: don’t decide whether any of them are fake. Just count hands.

That’s it. It usually ends with a teenager finding one and showing you.

Read next

• How to tell if a photo is AI-generated: 5 things to check
• What your child’s photo gives away