It was a Wednesday afternoon, the particular kind of low-energy Wednesday where I’d already done the playground and the library and lunch felt like a lifetime ago. My daughter was standing at my knee pulling at my sleeve, and I was exhausted in that specific way that makes the television remote look very reasonable. I didn’t turn it on. But I sat there and thought about it for a long moment, and that moment is why I ended up spending three hours that night reading research I could not stop reading.
What I found wasn’t the ambient worry I expected — screens bad, less is more, we know, we know. It was something more specific and more urgent: a category of video that is being manufactured at a scale that has no precedent in the history of children’s media, served algorithmically to children my daughter’s age, and almost completely invisible to parents who assume that if something is animated and cheerful it has been made by someone who thought about a child for even one second. It hasn’t.
This is the post I wish I’d found that Wednesday night instead of having to piece it together myself. Everything factual here comes from named sources. I’ll tell you what I actually do in our house at the end, including the parts where I fall short.
What we mean by AI-generated kids’ videos
The first major reporting came from WIRED on March 12, 2024, when deepfake-detection firm Reality Defender found evidence of generative AI in videos from kids’ channels including “Yes! Neo,” a Cocomelon-style 3D toddler channel. Alongside that, WIRED documented an entire tutorial cottage industry dedicated to teaching people how to churn out kids’ videos using ChatGPT and AI voice and video tools — not as a creative project, but as a revenue stream. The process is automated enough that a single operator can publish dozens of videos a day without ever thinking about what a two-year-old needs from a video, because no one is thinking about what a two-year-old needs from a video.
The scale is the thing that stopped me cold. A channel called Jo Jo Funland posted more than 10,000 videos in approximately seven months starting in August 2025 — roughly 50 videos a day. Sesame Street, which has been producing children’s television with researchers, educators, and child development consultants for decades, has posted about 3,900 videos across 20 years. Undark’s March 2026 investigation put both of those numbers in the same sentence and let them sit there. That contrast is the whole story.
What the New York Times investigation actually found
In early 2026, the New York Times ran an investigation whose findings were reported widely in March 2026. Researchers created a fresh account, had it watch kids’ channels, and tracked what YouTube recommended next. Of more than 1,000 Shorts recommended to that account, nearly half had AI-generated visuals. After a single CoComelon video, more than 40% of Shorts recommended within 15 minutes were synthetic. That’s not a marginal contamination of a mostly-safe feed. That is the feed.
More disturbing than the volume was what was in the videos. Documented examples, as reported by Futurism, included: a baby swallowing whole grapes; an infant eating honey; children walking in traffic; and garbled “educational” text rendered as things like “Conmecticut.” The videos presented themselves as educational. Let me be plain about why two of those examples are not just wrong but dangerous: whole grapes are a genuine choking hazard for babies and toddlers — they’re the exact size and shape to completely block a small airway — and honey is genuinely unsafe for children under one year old because it can contain bacterial spores that cause infant botulism, a serious and potentially fatal illness. A toddler watching a cheerful animated video that shows a baby eating these things is receiving instruction, not entertainment.
If your child is watching alone, what is she being taught right now, and by whom?
This has happened before — and it got worse
Parents who were paying attention in 2017 will remember Elsagate: disturbing videos wearing the costumes of familiar kids’ characters that flooded YouTube and YouTube Kids before the platform moved to address them. That precedent matters because it shows the pattern — the platform’s openness gets exploited, children are exposed to genuinely harmful content before anyone acts, and then the fix is partial and slow. In May 2025, WIRED found an AI-era echo of Elsagate, with dozens of channels using AI to show familiar kids’ characters being harmed, all set to nursery-rhyme music. The characters changed. The mechanism is the same.
The money, the loophole, and where policy stands as of July 2026
The reason this is happening at scale is simple: it is profitable. Fortune reported on April 1, 2026 that top AI-slop kids’ channels earn more than $4.25 million per year. A Kapwing analysis from November 2025 estimated that approximately 21% of YouTube’s feed is low-quality AI video. These are not hobbyists experimenting. This is an industry.
YouTube has taken some steps. Since March 2024, creators have been required to label “altered or synthetic” content — but only realistic content. Clearly animated content is exempt. That exemption is a loophole large enough to drive the entire AI kids’-content industry through, because nearly all AI-generated children’s cartoons are, by definition, animated. Since July 15, 2025, mass-produced “inauthentic content” has been ineligible for ad revenue, which sounds meaningful but enforcement depends on YouTube’s own classification. In April 2026, YouTube said it is developing dedicated AI labels for YouTube Kids. As of July 2026, no timeline has been given. YouTube’s own disclosure policy page explains the current labeling framework.
In March 2026, the child advocacy organization Fairplay organized an open letter to the CEOs of YouTube and Google, signed by more than 200 organizations and experts, demanding that YouTube ban AI-generated content on YouTube Kids entirely, label all AI content, stop recommending it to minors, and add an off-by-default parental AI toggle. Kathy Hirsh-Pasek, a researcher at Temple University and the Brookings Institution, told Undark in March 2026: “We’re at the beginning of a monster problem.” Dana Suskind, a pediatrics researcher at the University of Chicago, described toddler AI misinformation happening “at an industrial scale” and called it “very risky for the developing brain.” You can read the full Fairplay letter here.
How to spot it: my detection checklist
These are the markers researchers and journalists have consistently flagged. None is definitive alone, but several together should end the video:
- Malformed hands, fingers, and teeth. Current AI video generation handles extremities and mouths poorly. Extra fingers, fused hands, and teeth that multiply or disappear are common artifacts.
- Garbled on-screen text. Letters that don’t form words, or words that are misspelled in ways that suggest a model hallucinating language rather than a person typing it.
- Details that morph between frames. A character’s hair color shifting, a pattern on clothing changing shot to shot, objects appearing and disappearing without edit cuts.
- Robotic narration and off lip-sync. AI voice generation is improving but still tends toward flat affect; mouth movements frequently don’t match the audio track.
- Physically impossible scenes. Gravity behaving wrong, objects passing through each other, environments that are incoherent.
- Channel upload velocity and template thumbnails. A channel posting dozens of videos a day, all with near-identical thumbnail layouts, is a reliable signal.
If a channel raises flags, you can block it through YouTube’s parental controls. That’s a one-tap fix for channels you’ve already found. It does nothing about the ones in the autoplay queue.
What we actually do in our house
We don’t use autoplay. That’s the single rule I feel most confident about — the autoplay queue is where this problem lives, because it is the mechanism that hands control from a parent to an algorithm. Our daughter doesn’t have her own watch time; when she watches something, one of us is next to her. I know that’s not always possible for every family and I’m not holding it up as a moral standard, just describing what we do.
We choose programs in advance and open them directly. In practice that means I have a short list of channels I’ve vetted — I’ve watched several videos from each, looked at the channel’s upload history, and checked for the markers above. When she watches something, it’s from that list, opened intentionally. We don’t browse.
Here’s what’s hard: I have not been perfectly consistent. There have been sick days. There have been work emergencies. There have been moments when I handed over the tablet without checking what was already open. I don’t think the goal is a perfect record. The goal is a default posture that makes intentional viewing the path of least resistance in our house, so that even on bad days the starting point is better than the open algorithm. That’s all I’m actually managing to do, and I think it’s worth something.
Questions I keep getting
Isn’t YouTube Kids supposed to be safe?
YouTube Kids is a filtered version of YouTube, but filtering is not curation. The New York Times investigation found that after a single CoComelon video, more than 40% of Shorts recommended within 15 minutes on a fresh account were synthetic AI content. The filtering removes some categories of content, but it does not exclude AI-generated animation — in part because YouTube’s own labeling policy explicitly exempts clearly animated content from its synthetic-content disclosure requirements. As of July 2026, YouTube has said it is developing dedicated AI labels for YouTube Kids but has given no timeline for when they’ll arrive.
Does my child actually learn anything from these videos?
The concern from researchers isn’t only what children fail to learn — it’s what they might learn incorrectly. Dana Suskind at the University of Chicago has described the misinformation in these videos as happening “at an industrial scale” and called it “very risky for the developing brain.” Documented examples from the New York Times investigation included videos presenting whole grapes and honey as normal infant foods. Whole grapes are a choking hazard; honey under age one carries a risk of infant botulism. “Educational” framing makes incorrect modeling more dangerous, not less, because it signals to a watching child that this is how things are done.
What if I can’t co-view every time?
Then the most protective single thing you can do is turn off autoplay and pre-select what’s playing before you step away. The autoplay queue is the delivery mechanism for most of this content — a child who starts on a vetted video and can’t tap to the next one is dramatically less exposed than one whose feed is running freely. It’s not a perfect solution, but it changes the risk profile significantly without requiring you to sit down for every minute of screen time.
Is anyone making YouTube actually do something about this?
Fairplay’s March 2026 open letter, signed by more than 200 organizations and experts, made four specific demands: ban AI-generated content from YouTube Kids, label all AI content, stop recommending it to minors, and create an off-by-default parental AI toggle. YouTube responded in April 2026 by saying it is developing dedicated AI labels for YouTube Kids — which addresses one of the four demands, partially, with no timeline. Kathy Hirsh-Pasek, speaking to Undark in March 2026, described the current situation as “the beginning of a monster problem.” That framing — the beginning — is what I think about. The advocacy is real and worth following, but the platform’s response, as of now, has not matched the scale of what researchers are documenting.
