Deepfakes: what to actually teach a 10-year-old

The advice you were going to give does not work

Start here, because it saves you teaching the wrong thing well. Köbis and colleagues ran a pre-registered experiment with 210 people and reported four findings in one sentence: "people cannot reliably detect deepfakes", "neither raising awareness nor introducing financial incentives improves their detection accuracy", people are "biased toward mistaking deepfakes as authentic videos (rather than vice versa)", and they "overestimate their own detection abilities".

Read the middle two again, because they are the ones that matter for parenting. Telling someone to be careful did not help. Paying them to get it right did not help. And the errors are not symmetrical — the default failure is believing a fake, not doubting something real. Combine that with overconfidence and you get a child who has been warned, feels equipped, and is slightly more exposed than before.

On still images it is worse again. Nightingale and Farid found that synthesis engines "have passed through the uncanny valley and are capable of creating faces that are indistinguishable — and more trustworthy — than real faces." Not merely as believable: rated as more trustworthy. These are adults, specific image and video sets, and the engines of their publication dates — none of which softens the conclusion for a child, because the direction of travel only ever makes the viewer's job harder.

And a detector app is not the backstop

The natural next move is to reach for a tool. Groh and colleagues compared ordinary people against the leading machine learning model on a large set of short videos, and found people "perform in the range of the leading machine learning model" — already a less impressive endorsement of the model than it sounds. A combined human-plus-model system did beat either alone.

But the sting is in their next sentence: "inaccurate model predictions often lead humans to incorrectly update their responses." A detector that is wrong does not merely fail to help; it takes people who had it right and moves them to wrong. That is worse than no tool, and it is why this page recommends none. Judging the artefact — by eye or by app — is a dead end for a household, so the rest of this page is about the other question.

Teach the four questions instead

• "Who put this here first?" — Not who sent it to you — who originally posted it. Almost everything a child sees arrives third-hand, and the trail usually ends somewhere that either is or very obviously is not a real source. This one question resolves more cases than any visual check, and it works on text, images and video identically.
• "Is anyone else saying it?" — Leave the thing and search for the claim itself. If something genuinely happened, more than one outlet that does not share a parent company will have it. If the only place it exists is the place you found it, that is the finding.
• "What does it want me to feel?" — Synthetic media that spreads is engineered for an emotion — outrage, disgust, delight, alarm. Noticing the feeling arriving, as a signal rather than a verdict, buys half a second before the share. The question is not "is this fake" but "why did this make me want to send it to someone?"
• "What would I expect to see if it were real?" — The one check that still works, because it is about the world rather than the pixels. If a famous person said that, it would be on their own account. If it happened at your child's school, somebody in the house would know. Absence of the thing you would expect beats any artefact in the image.

Leave the page — the move is sideways

Questions one and two have a measurement behind them. Wineburg and McGrew put 45 experienced internet users — ten PhD historians, ten professional fact checkers, twenty-five Stanford undergraduates — in front of live websites. "Historians and students often fell victim to easily manipulated features of websites, such as official-looking logos and domain names. They read vertically, staying within a website to evaluate its reliability. In contrast, fact checkers read laterally, leaving a site after a quick scan and opening up new browser tabs… Compared to the other groups, fact checkers arrived at more warranted conclusions in a fraction of the time."

Two things there are worth saying to a child directly. The experts who stayed on the page were fooled, and by exactly the features that look authoritative. And the people who did it right were faster — which matters enormously when you are asking a ten-year-old to do something before they share. Checking is not the slow, worthy option; done laterally, it is the quick one. This is 45 adults judging websites rather than children judging a video, so what transfers is the strategy, not a number.

It is also worth knowing that this does not arrive on its own with screen time. Breakstone and colleagues gave 3,446 US high-school students six tasks with a live internet connection and reported plainly that "the students struggled on all of them" — 96% never discovered that a site presenting itself as publishing factual climate reports had ties to the fossil fuel industry. The reason is the useful part: students "were often duped by weak signs of credibility: a website's 'look', its top-level domain, the content on its About page, and the sheer quantity of information it provided." Those are surface properties, and surface properties are precisely what a generative model produces perfectly.

The same habit, applied to what the AI tells them in words · The wider safety picture this sits inside

A ten-minute drill you can run tonight

1. 1. Find two images of the same event and let them be confident Pick any topic your child cares about and find two images or clips — you do not need to know in advance which, if either, is genuine. Ask which one is real and, crucially, how sure they are out of ten. Write the number down. The confidence number is the part that teaches: it is the thing the research says is miscalibrated, and no child believes that about themselves until they have watched it happen once.
2. 2. Now do it the other way, out loud Together, run questions one and two: who posted it first, and is anybody unconnected saying the same thing? Narrate your own uncertainty while you do it — "I do not know yet" is the most useful sentence a parent can model here. Then revisit the confidence number. The gap between the first guess and what you found is the lesson, and it lands far harder than being told.
3. 3. Agree the household rule before you need it One sentence, agreed while nothing is wrong: in this house we do not forward a picture or video of a real person unless we know where it came from. Make it a rule about forwarding rather than about believing — forwarding is the behaviour you can actually see, and the one that does the damage. Rules agreed in calm are the only ones that survive a group chat at nine at night.

The part about pictures of people they know

There is a version of this that is not about celebrities or politics: a fake image or video of a classmate, made as a joke or to be cruel. Your child does not need a technical explanation of how that works, and this page is not going to give them one. They need one rule and one reassurance.

The rule: if you see a picture or a video of someone you know that they would not want shared, you do not send it on, and you tell an adult. The reassurance, said in advance and said plainly: you will not be in trouble for showing it to me. Children sit on these things for days because they are worried about being blamed for having seen it, and the delay is usually the thing that causes the most harm. If it involves your child or a child you know, your school and your local reporting routes are the next step, not a group chat.

How you find out whether any of it stuck

The only real test of this habit is whether your child uses it when you are not there, which means you need to hear their own words rather than their test score.

The research under this page

Five sources. The first three are here because they refute the advice most pages on this topic give, which is the only reason this page exists.

1. Köbis, Doležalová & Soraperra, 2021 — Fooled twice: people cannot detect deepfakes but think they can. iScience, 24(11), 103364. (A pre-registered behavioural experiment, N = 210: "(1) people cannot reliably detect deepfakes and (2) neither raising awareness nor introducing financial incentives improves their detection accuracy… (3) people are biased toward mistaking deepfakes as authentic videos (rather than vice versa) and (4) they overestimate their own detection abilities.") The core finding: detection is unreliable, warnings and money did not help, errors skew towards believing fakes, and confidence outruns ability. Adults, one video set — evidence against "look harder", not a measurement of children. What it does not show: Adults, one video set, one study. It is strong evidence that warning someone to look harder does not work — which is the claim this page makes — and it is not a measurement of what any child can or cannot spot.
2. Nightingale & Farid, 2022 — AI-synthesized faces are indistinguishable from real faces and more trustworthy. Proceedings of the National Academy of Sciences, 119(8), e2120481119. ("Our evaluation of the photorealism of AI-synthesized faces indicates that synthesis engines have passed through the uncanny valley and are capable of creating faces that are indistinguishable — and more trustworthy — than real faces.") Synthetic faces rated indistinguishable from real ones, and more trustworthy. Still images and the engines of its date, so treat it as a floor on the problem rather than a current reading. What it does not show: Still portrait faces, not video, and the synthesis engines of its publication date rather than today’s. The direction only ever gets worse for the viewer, so the page treats it as a floor on the problem rather than a current measurement.
3. Groh, Epstein, Firestone & Picard, 2022 — Deepfake detection by human crowds, machines, and machine-informed crowds. Proceedings of the National Academy of Sciences, 119(1), e2110013119 (print issue 5 January 2022). ("Ordinary humans perform in the range of the leading machine learning model on a large set of minimal context videos… While we find that a system integrating human and model predictions is more accurate than either humans or the model alone, we show inaccurate model predictions often lead humans to incorrectly update their responses.") Why this page recommends no detector app. Cited as much for its second half — a wrong model prediction moves correct humans to incorrect answers — as for the headline comparison. What it does not show: Short, minimal-context videos and adult crowdworkers. Cited for the second half as much as the first: a detection tool that is wrong does not merely fail, it drags a person who was right towards being wrong — which is why this page does not recommend one.
4. Wineburg & McGrew, 2019 — Lateral reading and the nature of expertise: reading less and learning more when evaluating digital information. Teachers College Record, 121(11), 1–40. (45 experienced internet users — 10 PhD historians, 10 professional fact checkers, 25 Stanford undergraduates — thinking aloud as they judged live websites. "Historians and students often fell victim to easily manipulated features of websites, such as official-looking logos and domain names. They read vertically, staying within a website to evaluate its reliability. In contrast, fact checkers read laterally, leaving a site after a quick scan and opening up new browser tabs in order to judge the credibility of the original site. Compared to the other groups, fact checkers arrived at more warranted conclusions in a fraction of the time.") The basis for "leave the page and open a tab", and for the detail that subject expertise did not protect the historians. What it does not show: A small expert/novice study of 45 adults evaluating websites, not children evaluating a chatbot. What transfers is the STRATEGY, and the finding that subject expertise did not protect the historians; the study says nothing about how well a nine-year-old can execute it.
5. Breakstone, Smith, Wineburg, Rapaport, Carle, Garland & Saavedra, 2021 — Students’ civic online reasoning: a national portrait. Educational Researcher, 50(8), 505–515. (3,446 US high-school students with a live internet connection, on six constructed-response tasks. "The students struggled on all of them." 96% never learned that a site claiming to disseminate factual climate reports had ties to the fossil fuel industry; two thirds could not distinguish news stories from ads; more than half believed an anonymously posted video shot in Russia was "strong evidence" of US voter fraud. Students "were often duped by weak signs of credibility: a website’s ‘look’, its top-level domain, the content on its About page, and the sheer quantity of information it provided.") The evidence that this does not develop on its own. High-school students on websites, so the numbers do not transfer — but the mechanism does: they were duped by surface signals of credibility, which is what a generative model produces best. What it does not show: High-school students, and about websites rather than chatbots. It is evidence that this skill does not arrive on its own with age or screen time — not a measurement of what a seven-to-twelve-year-old can do, and we quote no figure as if it were.

Questions parents ask

How do you spot a deepfake?

Mostly, you do not, and that is the honest answer rather than a defeatist one. In a pre-registered study of 210 people, detection was unreliable, and neither warning participants nor paying them improved it; separately, AI-synthesised faces have been rated as indistinguishable from real ones and more trustworthy than real ones. Visual tells also expire faster than any article about them. The durable move is provenance: who published this first, and is anyone unconnected to them saying the same thing?

What age should I talk to my child about fake images?

Around the age they start seeing media you did not hand them, which is usually well before the topic gets raised. You do not need the word "deepfake" or any explanation of the technology. "Some pictures and videos are made by a computer, so the way we check is to find out where it came from" is complete and accurate for a seven-year-old, and it is the same lesson you will still be giving at fourteen.

Should I install an app that detects AI images?

We would not, on the evidence. When ordinary people were compared against the leading detection model on short videos, people performed in the range of the model — and the same study found that inaccurate model predictions often led humans to change correct answers to incorrect ones. A tool that is confidently wrong is worse than no tool, because it replaces a healthy "I do not know" with a false certainty.

What if there is a fake image of my child or their friend?

Two things, in this order. Make sure your child knows in advance that they will not be in trouble for showing you — children delay telling an adult because they fear being blamed for having seen it, and the delay usually does more harm than the image. Then do not forward it, keep the evidence, and go to the school and your local reporting route rather than handling it in a group chat.

Practise the checking, not the spotting

This habit is hard to drill at home because most tools answer instantly and confidently every time. Luchi is built to ask a kid what they think before it tells them, and to make them explain what they made — and it writes you a short note in their own words, so you can hear whether the checking question is becoming theirs.

Start your first 7 days for $5

Is AI safe for kids?

Part of the series: Is AI safe for kids?

Related guides in this series

• Is ChatGPT safe for kids? — A straight answer by age — the risks that actually come up, and the five habits that make AI safe when you do let kids use it.
• Screen time and AI — Not all screen time is equal. Why active, creative AI use is different from passive scrolling — and how to keep your child on the right side of the line.
• What is my kid doing on AI? — You can’t read every transcript. Four checks that take under a minute each, what each one tells you, and the one question a ten-year-old actually answers.
• AI apps a parent can actually monitor — Alert, transcript, or record — three different things sold under one word. What seven products say they show a parent, quoted from their own sites.
• Is Google Gemini safe for kids? — Google’s own page for parents is more candid than the reviews — imperfect filters, stated hallucination, "not a person to confide in". What it says, what it leaves out, and the four decisions that are yours.
• The school Chromebook already has AI on it — A school account is managed by the district, not by your Family Link. Who actually holds the switch, what Google’s own documentation says a student account gets, and five questions worth one email.
• What to ask at a parent-teacher conference about AI — The list you found ends with “ask how you can help with homework” — the one form of involvement a 50-study meta-analysis found was not associated with achievement. Six questions that are, and what each answer tells you.
• Is Siri the kind of AI parents worry about? — Apple has quietly split Siri in two — its own page now says Apple Intelligence “powers an entirely new version of Siri: Siri AI”. How to tell which one your child is talking to, what Apple keeps, and the settings worth changing tonight.
• A family AI agreement you will actually keep — Twenty clauses, signed on Sunday, dead by Thursday. A four-clause version built on what a 94-test meta-analysis found actually makes a plan start — naming the when and the where — plus the one question to ask each week.
• An AI said something it should not have — Google’s own parent page says the filters "are not perfect". The order to do things in — keep the conversation, ask the one question, report from inside the chat, decide one thing — and the point at which this hands you to somebody qualified.
• My child believes everything the AI says — The gap is usually confidence rather than method, and a check is still a looking-up — which across nine experiments left people MORE sure of what they understood. The calibration habit that moves it, in four turns.