ChildSafeAds Shared Task 2026: Commercial Content in Child-Facing YouTube Videos

Imagine sitting down to watch a video about building a treehouse or solving a tricky math puzzle. The creator speaks directly to you, sharing tips and stories in a friendly, familiar voice. Suddenly, without a break in the flow, they mention a specific brand of glue or a new app, explaining why it is the perfect tool for the job. For many children and teenagers, this seamless blend of entertainment and sales is hard to spot. Unlike a traditional television commercial that interrupts the show with a loud jingle, this kind of advertising hides inside the content itself. Because it looks and sounds like the rest of the video, young viewers often do not realize they are being sold something. This confusion is not just a minor annoyance; it is a legal and ethical concern. Laws in many places require that when someone is paid to promote a product, they must clearly tell the audience. Yet, in the vast, chaotic world of online video, these warnings are often missing, buried in long descriptions, or written in language that is too subtle for a child to understand.

Researchers have long worried that current systems for monitoring these videos are not good enough. They often rely on the video platform's own labels, which creators are supposed to add when they have a paid partnership. However, a new study suggests that relying on these labels alone would miss a huge number of commercial messages. To understand the true scale of the problem, a team of researchers from universities in the Netherlands created a new way to look at the data. They built a shared task, a kind of organized challenge for computer scientists, to see if machines could identify these hidden sales pitches and figure out what is wrong with them. The goal was not just to find the ads, but to understand what they were selling and whether they were breaking rules designed to protect young people.

The team started by gathering a massive collection of video clips. They did not look at every video on the internet, but instead focused on a specific tool called SponsorBlock. This is a community-driven project where regular viewers mark the exact start and end times of sponsorship segments in videos so that others can skip them. The researchers took these user-submitted markers as their starting point, knowing that if a human viewer felt the need to skip a part, it was likely a commercial message. They collected 3,360 of these segments from nearly a thousand different channels that are popular with teenagers. For each segment, they gathered everything available: the words spoken in the clip, the video title, the channel name, and the links the creator put in the description box. They even followed those links to the actual websites where the products were being sold. This created a rich dataset where a computer system could look at a short transcript, or dig deeper into the video details and the product page, to make a judgment.

The researchers then set up three specific jobs for the computer systems to perform. First, the system had to decide what kind of offer was being made. Was the creator selling a physical object like a toy, a digital service like a game, or something else entirely? Second, the system needed to categorize the product. Was it food, fashion, health advice, or perhaps something risky like gambling? Finally, and most importantly, the system had to act as a legal watchdog. It needed to flag any potential problems, such as a claim that sounded too good to be true, a failure to clearly say "this is an ad," or a direct push for a child to buy something right now. The researchers designed the task so that teams could choose how much information to use. Some systems might only read the transcript, which is cheap and easy to get. Others might crawl the linked websites, which is much more expensive and time-consuming. This allowed the researchers to see if the extra effort of gathering more data actually led to better results.

When the team analyzed the data they had collected, they found a startling reality. Nearly half of the videos in their collection, specifically 45.5 percent, did not use the official "paid promotion" label that the video platform provides. This means that if regulators or parents only looked for that specific label, they would miss almost half of the commercial content aimed at young people. The study also revealed that the task of identifying these ads is difficult for computers. While the systems were quite good at figuring out what kind of product was being sold, they struggled significantly with the legal flags. The computers often disagreed with each other on whether a disclosure was clear enough or if a claim was misleading. This disagreement highlights that understanding the nuance of language and context, especially when it comes to protecting children, is a complex challenge that machines are not yet perfect at solving.

The researchers used advanced artificial intelligence models to create the labels for their dataset, working closely with legal experts to refine the rules. They did not just guess; they iterated on their definitions, checking samples and adjusting how they asked the computer to think. For example, they narrowed the definition of "direct exhortation," which means urging a child to buy something, to ensure the computer only flagged language that truly pressured a child, rather than just a generic suggestion to click a link. Even with these careful steps, the computer models still made mistakes. In the test set, two different powerful models agreed on the exact set of legal flags for less than half of the cases. This suggests that while technology can help scan thousands of videos, it cannot yet replace human judgment in determining whether an advertisement is truly safe for a young audience.

The study concludes that monitoring child-facing content requires more than just looking for official labels. It demands a deeper look into the content itself, the links provided, and the specific language used. The researchers made their data and the rules for the task available to the public, inviting other scientists to build better systems. They emphasized that their work is a starting point for research, not a final solution. The dataset they created is a snapshot of a specific moment in time, and it has limitations. It only includes videos where someone already marked a sponsorship, and it relies on links that were still working when they were checked. It does not include videos that were purely commercial or channels that did not pass their screening for teenage appeal. Despite these limits, the work provides a clear picture of the current landscape: a vast amount of commercial content slips through the cracks of current monitoring systems, often without the clear warnings that the law requires. The path forward involves building tools that can understand context as well as they understand words, ensuring that the digital world remains a safe space for children to learn and play without being unknowingly sold to.