Concerned parents are raising alarms over the presence of sexually explicit and violent AI-generated videos allegedly being suggested to their children by YouTube.

A mother named Liz Guilar revealed to the Daily Mail that she encountered an overwhelming amount of disturbing content after creating an account for her young children, aged one and three.

Guilar described the unnerving visuals, stating, “The imagery can be terrifying, featuring melting faces, vanishing limbs, and inadvertently created monsters reminiscent of something from Silent Hill. The unsettling part is that toddlers could be exposed to this if parents depend on YouTube for content moderation. It’s quite chilling.”

AI specialist Jeremy Carrasco suggests that Guilar’s experience might not be an isolated incident and that YouTube could be systematically recommending alarming AI-generated videos to young audiences.

In an experiment, Carrasco claims to have uncovered evidence that these disturbing videos are being promoted by algorithms typically intended for children’s content.

These videos reportedly include cartoon-like portrayals of child abuse, AI figures undergoing grotesque surgeries, handling weapons, committing violent acts, or engaging in sexual activities.

Carrasco said he drew his conclusion by creating a fake account and simulating the streaming habits of a child on YouTube to test what kinds of videos would be recommended. 

He told the Daily Mail: ‘YouTube could tell I was a kid interested in Bluey or Roblox and yet it still was suggesting these on the homepage. Even the thumbnails on the homepage were sometimes egregiously awful.’

Video producer and AI consultant Jeremy Carrasco said he created an account pretending to be a child to see what kind of content was pushed to him

Disturbing AI videos on YouTube often include humanoid animal characters. In this video, a cat-mother smothers her baby with a pillow before throwing it off a cliff

He claimed that the homepage prominently recommended disturbing AI-generated content from the site’s YouTube Shorts feed, from channels with names such as ‘MeowBoom’ and ‘MeowKitten007.’

YouTube told the Daily Mail that none of the videos Carrasco’s experiment flagged appeared in the YouTube Kids app, but it did disable several channels highlighted by this outlet after finding they violated community guidelines.

Jack Malon, a spokesman for the platform, told the Daily Mail: ‘To be clear, this content is not in the YouTube Kids app. Parents are in control of what their kids see on YouTube, and must choose a specific content restriction level for their child’s account during setup.’ 

Carrasco said that even if these disturbing AI videos do not appear on the YouTube Kids app, the regular version of the platform should still be able to tell if a child is watching based on the content they consume. 

‘I think that they should be able to, with some age tools, be able to determine, “Oh, hey, we have a seven-year-old watching, we should probably just not put shorts on the homepage because we know that’s the healthy thing to do,”‘ he said. 

‘They already have the tools to do this. They could do that tomorrow. I’m not even suggesting that they block those users from seeing the short-form videos if they search them. Just don’t algorithmically push them the most harmful sh*t.’

AI-generated shorts channels began cropping up in late last year as the technology became more sophisticated. 

In this AI video, an angry mother with a sheep’s head beats a cat with a bat and pours what looks like tar over it

Since then, dozens of channels have been cashing in with easy-to-make AI videos that have violent, disturbing and sexual themes.

‘My immediate reaction was to think about how terrifying it would be if my children had been watching on their own,’ Guilar told the Daily Mail.

‘I don’t think young kids have the ability to process what they’re seeing the same way adults do – so what’s a hilarious AI fail for us, could potentially confuse them, frighten them or even normalize disturbing themes.’

Guilar made a post on Reddit earlier this year stating she was, ‘genuinely horrified to discover how much brain-melting AI slop there is, like straight-up, horrifying stuff.’

In the thread, dozens of other parents shared similar experiences to the Australian mom and lamented the harmful and disturbing nature of the content they found.

One parent wrote: ‘I went to a hairdresser that has a private studio recently and she brought her 4/5yo because school was canceled for weather and I didn’t mind.

‘But then she put YouTube on the TV for her and it went from educational songs/videos to the weirdest AI brain-melting crap I had ever seen SO quickly.’ 

Another parent said: ‘Some of them are genuinely freaky. I wish I took screenshots of some of the nightmare fuel that I blocked.’

The AI videos can include human characters as well. In this one, a panicked mother throws her baby out of a crashing plane’s window

Liz Guilar was ‘horrified’ by the disturbing AI-generated videos she found on YouTube after screening an account she had set up for her children

Guilar also said it was ‘terrifying’ to imagine her children watching the AI videos she found

This grotesque video featuring an AI-generated version of the famous soccer player Ronaldo did particularly well, receiving more than 200,000 likes

Andrew Koepp, a developmental psychologist and assistant professor of applied psychology at NYU, explained to the Daily Mail the harm this kind of content could have on children. 

He said the ‘videos’ surface-level features are clearly designed to appeal to children with bright colors and exaggerated, cartoon-like characters… Yet the content is clearly very inappropriate for children.’

‘Before the age of eight, children have difficulty distinguishing between fiction and reality,’ Koepp said, adding that the videos sent to him by the Daily Mail for review, ‘touch on abandonment and limb amputation, topics that deserve a great deal of sensitivity.’

The controversy follows the YouTube ‘Elsagate’ scandal of the late 2010s, in which creators use likenesses of children’s characters, such as Frozen’s Elsa or Spiderman, to act out inappropriate scenarios with violent and sexual themes.

The videos performed well because their creators put the names of the characters in the title and tags, which led the algorithm to promote them to children. 

Doctor Kostantinos Papadamou, a computer scientist and researcher, studied the Elsagate phenomenon and published a paper about it in 2020. 

Physical abuse of children is a common theme in these AI-generated videos

Some of the videos include blood and disturbing imagery of injuries

The current problem of disturbing AI videos is the latest iteration of a longer-lasting issue of YouTube recommending inappropriate videos to children. These screenshots are from videos that were part of the late 2010s ‘Elsagate’ controversy

In the study, he and his team simulated the behavior of a child on YouTube and found that there was a ‘3.5 percent chance that a toddler following YouTube’s recommendations will encounter an inappropriate video within ten [clicks] if she starts from a video that appears among the top ten results of a toddler-appropriate keyword search randomly.’

YouTube made changes to the platform in an attempt to address the Elsagate problem. 

It added a feature that required anyone posting to the website to click a box stating whether the video was meant for children, which helped alleviate the issue. 

Papadamou drew connections between Elsagate and the current problem with inappropriate AI content, and told the Daily Mail he was concerned about the ease of creating these disturbing new AI videos. 

‘The problem was big enough by 2020 with the volume of videos we used to have, and they couldn’t keep up because most of the measures were binded [sic] to a human,’ he said.

‘You can imagine how these measures cannot keep up with the volume nowadays.’