Expert: AI 18+ videos appearing doesn't mean users "like to watch"
(Dan Tri) - Without actively searching or clicking "like", many people still constantly encounter AI 18+ movies. Experts say the answer lies in how the algorithm reads a series of signals from user behavior.
The algorithm doesn't just look at the "like" button
After a few times accidentally encountering AI movie ads with sensitive content, some users reported to Dan Tri reporters that similar videos seemed to appear more and more frequently.
In the article "Blush" because 18+ AI movies are rampant on social networks, the reporter noted the situation of users constantly encountering AI movie ads with 18+ elements, even though they do not actively search or monitor this type of content.
To explain why this phenomenon occurs, Dan Tri reporter talked with cybersecurity expert, Mr. Ngo Minh Hieu (Hieu PC). According to Mr. Hieu, the recommendation systems of platforms do not only rely on keywords that users actively search for.
Viewing history, time spent on a video, rewatching, commenting, sharing, or continuing to watch similar content can all become signals that help the system gauge interest.
According to Mr. Hieu, viewing time, viewing history and interactions can all become signals for the system to recommend content (Photo: provided by character).
Therefore, even without clicking "like", stopping to watch a video can provide additional data to the recommendation system.
Large platforms also openly use many different types of signals for this system. YouTube says the recommendation system can be based on watch history, search history, subscriptions, likes, dislikes and "Not interested" responses.
The system also compares viewing habits with users with similar behavior to find potentially relevant content.
This also partly explains the fact that a person who has never typed the keyword "AI 18+ movie" can still have the system try to suggest similar content.
The fact that a sensitive video appears does not mean that the algorithm has determined that users love 18+ content, because the system is based on many different signals (Illustration: Doan Thuy).
However, Mr. Hieu noted that the appearance of a few sensitive videos is not enough basis to conclude that the algorithm has determined that users "like to watch 18+ movies".
The proposed systems have to process a lot of signals at the same time. A person who pauses out of curiosity, surprise, or even discomfort may also exhibit behavior that outwardly resembles their interest in the video.
On the contrary, proactive signals such as "Not interested", "Do not recommend this channel" or dislike button are also used by some platforms to adjust recommended content.
Why can we still meet again after reporting?
Another question that arises is why users who hide or report a video can still encounter similar content later.
According to Mr. Hieu, it is necessary to distinguish between processing a specific video and identifying a whole group of content.
The same video can be reposted by another account, cut into a shorter version, change the image or sound, or use only a few scenes. With generative AI, a greater number of variations can be created at a faster rate than before.
According to experts, viewing time, viewing history and interactions can all become signals for the system to recommend content (Photo: Doan Thuy).
A video may also contain sensitive images for only a few seconds, be blurred, cropped, or continuously change details. This creates additional challenges for automated moderation systems that have to process huge amounts of content.
Therefore, according to experts, it cannot be assumed that every video that passes through censorship is because the poster intentionally "circumvented the algorithm". However, the ability to create a series of variations using AI is making the problem of detecting and handling sensitive content more complicated.
Platforms also acknowledge this issue in their policies.
TikTok stipulates that some sexual content, although not necessarily deleted, can still be removed from the For You message board or restricted by age. Meta also distinguishes between content that is allowed to exist on the platform and content that is eligible to be proactively recommended to users.
AI content can be created into many different versions, posing challenges for both the moderation and recommendation systems of platforms (Illustration: Getty Images).
According to Mr. Hieu, responsibility therefore cannot be placed solely on the user.
As platforms build systems to curate and distribute content to millions of people, they also need to assess risks, limit the amplification of sensitive content, and effectively handle user reports.
Especially for children, experts say it is necessary to combine automatic identification systems with human censorship, increase the ability to detect reposted content, strongly limit the recommendation of sexual content and apply appropriate age verification measures for adult content.
"The fact that users see a few 18+ videos is not enough to conclude that the algorithm knows they like that content. But if sensitive content continuously appears even though the user has no need and has responded without interest, the platform needs to review both the censorship mechanism and the recommendation system," Mr. Hieu commented.