Can Kids Tell AI-Generated Art From Human Art? Mostly Not

Can kids tell AI-generated content from the real thing? Mostly not. In a study published in March 2026, 37 children aged 6 to 10 tried to sort texts, pictures, artwork and faces into “made by a person” or “made by a SmartBot.” On average they did no better than flipping a coin, and sometimes worse. The kids who spent more time on devices at home were the least accurate.

The instinct is to teach kids the tells: weird hands, glossy skin, too-perfect text. That’s the wrong lesson. Adults can’t reliably spot AI by eye either, and the tells keep disappearing. What children need is a different question, and it’s one you can teach at the kitchen table.

Key Takeaways

• Langer, Martinez, Marshall and Chein (Technology in Society, 2026) found 37 children aged 6 to 10 “performed at or below chance” at telling human-made from AI-generated content, well below 49 adults.
• Children’s accuracy was negatively associated with parent-reported home technology use.
• 95% of the children understood that the AI agent isn’t alive. Knowing what AI is didn’t help them spot what AI made.
• Adults struggle too: a 2022 PNAS study found AI-synthesized faces were indistinguishable from real ones and rated more trustworthy.
• Teach provenance (“where did this come from, and how could we check?”) rather than visual guessing.

Can Kids Tell AI-Generated Content From Human Work? The 2026 Study

AI-generated content means text, images, audio or video produced by a generative AI model rather than made directly by a person. Kids meet it constantly in apps, videos, games and school tools, often unlabeled.

Allison Langer, Steven Martinez, Peter Marshall and Jason Chein designed one of the first studies of children’s ability to tell the difference. Their paper, “Children’s susceptibility to content generated by artificial intelligence,” was published online March 14, 2026, in Technology in Society.

Here is how it worked, according to the paper’s abstract:

• Setup: 37 children aged 6 to 10 heard a story about a “SmartBot,” a stand-in for AI, that could write text, invent photos and create art.
• Task: they looked at a series of short texts and images of objects, art and faces and decided whether each came from the SmartBot or from a human teacher.
• Extra measures: children completed tests of abstract reasoning and vocabulary; parents filled out a home technology use survey.
• Comparison: 49 adults aged 18 to 34 did the same sorting task.

The results:

• “On average, children performed at or below chance in discriminating human-created from AI-generated content across modalities.”
• Their accuracy was “significantly below that of adults.”
• Discernment “was significantly and negatively associated with parent reports of children’s at-home technology use.”
• Children “broadly understood that an AI agent (the SmartBot) is not alive (95%).”

The authors conclude there is “risk for misattributing AI content as human” and call for “early AI literacy.”

Limits worth knowing

Thirty-seven kids is a small sample. The tech-use link is a correlation from parent reports, so it doesn’t show that screen time causes worse discernment; kids who use more tech might simply see more polished media and treat it as normal. And we read the abstract and metadata, not the full text, so we don’t report scores by age or by type of content. Note also the adults were better than the children, not necessarily good. The abstract doesn’t say adults were highly accurate.

Why Looking Harder Doesn’t Work

Here’s the mechanism behind the problem. A generative image model is trained on huge numbers of real images and learns to produce new ones with the same statistical patterns. Early models left artifacts, mistakes a careful viewer could catch, like extra fingers or melted text. Each new generation reduces them, because reducing artifacts is literally what the models are optimized to do.

That means any checklist of visual tells has a short shelf life. By the time a child memorizes “look at the hands,” the next model draws hands fine.

Adults hit this wall years ago. Sophie Nightingale and Hany Farid, in a 2022 PNAS paper, reported that AI-synthesized faces “are indistinguishable” from real faces “and more trustworthy.” If professionals studying the problem conclude that eyes alone aren’t enough, it’s unfair to expect a 7-year-old to manage.

A September 2026 scoping review of 22 studies in Frontiers in Psychology (Niu et al.) also noted that children struggle to understand where AI’s information comes from. The 95% finding in the Langer study adds an important twist. Children knew the SmartBot wasn’t alive. They had the concept. What they lacked was a method for judging origin. Knowing about AI isn’t the same as knowing where something came from.

The tool that actually answers “who made this”

The engineering answer to provenance isn’t detection by eye. It’s records. The Coalition for Content Provenance and Authenticity (C2PA) publishes an open standard called Content Credentials: “tamper-evident, cryptographically signed data structures that travel with the asset” describing how a piece of content was created and edited. Any modification “will break this cryptographic linkage, signalling tampering.”

It has limits. The C2PA’s own FAQ notes that metadata “can be removed inadvertently or intentionally,” which is why it also supports watermarking and fingerprinting to recover credentials. And many images online carry no credentials at all. So a missing label doesn’t prove a human made something. But the principle is the one to teach: origin is a question of evidence, not appearance.

How to Teach Your Kid to Judge Where Content Came From

UNICEF’s updated Guidance on AI and children (version 3.0) lists “Prepare and skill children for present and future developments in AI” among its ten requirements for child-centered AI. Here’s what that looks like at home.

Ages 5–8: The “Who made it?” sorting game

Gather 10 pictures: five drawn or photographed by your family, five made with an AI image tool (do this part yourself). Mix them up. Your child sorts them into “person” and “computer” piles and explains each choice. Then reveal the answers together. Most kids will get several wrong, which is the lesson: “See? We can’t always tell by looking. So how could we find out?” Land on the answer: ask who made it, or look for where it came from. Keep it playful; the goal is curiosity, not suspicion.

Ages 9–12: The provenance detective

Pick three images or short texts your child has seen online this week. For each, they answer three questions on paper:

1. Where did it come from? Who posted it first, if we can tell?
2. What’s the evidence? A caption, a credit, a source link, an artist’s name?
3. Can we check? Try a reverse image search, look for a Content Credentials label, or find the original post.

They mark each item “known origin,” “probable” or “unknown.” The win is a child who’s comfortable writing “unknown” instead of guessing.

Ages 13+: Replicate the experiment

Teens can run a version of the Langer study. Build a set of 20 items, half AI-generated, half human-made, with a sibling or friend helping so the tester doesn’t know the answers. Have family members sort them and calculate each person’s accuracy. Chance is 50%. Discuss: Who beat chance? By how much? Is that good enough to trust? Then add a twist: give testers 30 seconds to check one source per item and compare scores. The gap between guessing and checking is the finding.

The question to ask: “If you can’t tell by looking, what would you need to know to be sure who made it?”

Guessing vs. Checking: What Works for Kids

Approach	What it relies on	Why it fails or works	Shelf life
Look for visual glitches (hands, text)	Model artifacts	Models are trained to remove them	Months
”It looks too perfect”	Gut feeling	Human professionals also make polished work; kids in the 2026 study were at chance	Already unreliable
Know what AI is	Concept	95% of kids knew AI isn’t alive and still couldn’t sort content	Doesn’t transfer
Ask who made it	Source	Works if the source is honest and known	Durable
Check for Content Credentials	Cryptographic record	Reliable when present; absence proves nothing	Growing
Reverse search for the original	Publication trail	Often finds the first poster or artist	Durable

What to Do at Home This Month

Label your own AI use

When you use AI to make a birthday invitation or a picture, tell your kids: “I made this with an AI tool.” Kids learn provenance habits by hearing adults state them. It also shows that using AI isn’t shameful, while hiding it is a choice.

Swap “Is it fake?” for “Where’s it from?”

“Is it fake?” invites a guess. “Where’s it from?” invites a search. Use the second question every time a surprising image or video comes up, especially in family group chats.

Watch for the screen-time link, without panicking

The Langer study found kids with more home tech use were worse at discernment. That’s correlational, but it’s a reason to talk about AI content with the kids who see the most of it, rather than assuming they’re the experts. Our guide to teaching kids to spot AI-generated content has more conversation starters.

Connect it to deepfakes for older kids

For tweens and teens, the stakes rise with realistic fake photos and videos of real people. The same skill applies: check the source before believing or sharing. Our explainer on kids, media literacy and deepfakes covers that territory.

What not to do

Don’t run “spot the fake” quizzes as a test your child can pass. They’ll either fail and feel stupid, or pass by luck and feel falsely confident. Both are worse than learning that eyes aren’t enough. Frame every exercise around how to find out, not how to guess.

What to Watch For Over the Next 3 Months

• Week 4: Your child says “I can’t tell, where’s it from?” at least once about something online. That’s the target behavior.
• Month 2 red flags: Confident claims like “I can always tell AI.” Sharing surprising images without checking. Treating anything polished as AI, or anything rough as human.
• Month 3 self-check: Show your child two new images, one AI and one human. Ask how they’d decide. If the answer is about source and evidence rather than “the hands look weird,” the lesson stuck.

Frequently Asked Questions

Can kids tell AI-generated images from real ones?

Usually not reliably. In a 2026 study, children aged 6 to 10 performed at or below chance at telling AI-generated texts, images, art and faces from human-made ones, and well below adults. Even adults struggle; a 2022 PNAS study found AI faces indistinguishable from real ones. Visual guessing isn’t a dependable skill at any age.

At what age can children understand what AI is?

The 2026 study found 95% of children aged 6 to 10 understood that an AI agent isn’t alive. But understanding the concept didn’t help them sort content. Kids can grasp that computers make pictures and stories quite young; judging where a specific piece came from is a separate skill that has to be taught.

Does screen time make kids worse at spotting AI?

The study found children with more parent-reported home tech use were less accurate. That’s a correlation from a small sample, so it doesn’t prove screen time causes the problem. It does suggest that heavy device users aren’t automatically savvier, and they may benefit most from explicit lessons.

What are Content Credentials?

Content Credentials are an open technical standard from the C2PA that attaches a cryptographically signed record to an image or video showing how it was made and edited. Any tampering breaks the signature. They’re useful when present, but many images lack them and metadata can be stripped, so no label doesn’t mean human-made.

Should I teach my child to look for weird hands and other AI tells?

Briefly, as a curiosity, but don’t rely on it. Generative models are trained to remove those artifacts, so visual tells go out of date fast. A more durable habit is asking who made something and checking the original source. That works no matter how good the models get.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Langer, A., Martinez, S., Marshall, P. J., & Chein, J. (2026). “Children’s susceptibility to content generated by artificial intelligence.” Technology in Society, 86, 103303. https://doi.org/10.1016/j.techsoc.2026.103303
2. Nightingale, S. J., & Farid, H. (2022). “AI-synthesized faces are indistinguishable from real faces and more trustworthy.” Proceedings of the National Academy of Sciences, 119(8). https://doi.org/10.1073/pnas.2120481119
3. Coalition for Content Provenance and Authenticity (C2PA). “FAQs.” https://c2pa.org/faqs/
4. UNICEF Innocenti. (2025). “Guidance on AI and children, Version 3.0.” https://www.unicef.org/innocenti/reports/policy-guidance-ai-children
5. Niu, T., Liu, H., Pang, P., Luo, Y. T., & Liu, T. (2026). “The role of generative AI in facilitating children’s creative thinking and cognitive development: a systematic scoping review.” Frontiers in Psychology, 17. https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2026.1880052/full