Almost six out of every ten videos that TikTok recommends to a newly created account are low-quality content generated with artificial intelligence. This is according to a study by the video editing platform Kapwing, which has analyzed 10,742 publications belonging to twenty categories and has separately examined the first 500 videos shown in the “For you” section of a new user.
Of those initial 500 recommendations, 294 were classified as “AI slop,” an expression used to describe content created automatically, produced with little care and distributed primarily to accumulate views, subscriptions or interactions. The result puts the proportion at 59%, almost three times higher than the 21% that Kapwing detected through an equivalent exercise with a new YouTube Shorts account.
The analysis focuses on the first impression offered by the platform before the recommendation system has sufficient information.
The data does not mean that 59% of all the content available on TikTok belongs to this category. The analysis focuses on the first impression that the platform offers before its recommendation system has sufficient information about the user’s interests, interactions and consumption habits.
Until it gets those signals, TikTok offers popular content aimed at a broad audience and adjusted for, among other factors, language and location. According to the study, this initial configuration is currently dominated by AI-generated videos that reproduce easily recognizable visual, narrative and sound formulas, but provide little value and, in some cases, spread incorrect information.
Children’s content concentrates the greatest presence of automated videos
The research also analyzed a sample of 10,742 videos associated with popular tags from twenty categories. Children’s content presented the highest proportion of “AI slop”: 1,147 of the 2,000 pieces examined, 57.4% of the total.
The presence is even higher in some specific labels. In #CartoonKids, 97 of the 100 videos reviewed were generated with artificial intelligence. In #cartoons and #babysong, the percentage reached 83%, while in #forkids it stood at 79%.
These videos often reuse well-known children’s characters in incoherent situations, use synthetic voices, and present animations in which objects and figures change shape without continuity. They also include songs, stories or educational content constructed quickly and without sufficient supervision.
The problem is especially delicate when these pieces have errors that an adult can recognize, but that a small child does not yet have enough knowledge to detect. Among the examples pointed out by Kapwing is a preschool learning video that uses “Sesame Street” characters to teach counting cookies, although the numbers spoken do not match the objects shown.
“I think of it as children’s misinformation generated with AI on an industrial scale. It’s very risky for a developing brain”said Dana Suskind, Professor of Pediatrics at the University of Chicago, in statements cited in the report.
The specialist warns that the experiences received during the first years intervene in the formation of neuronal connections. From this perspective, the risk reaches the repeated transmission of incorrect concepts during a decisive stage for learning.
More than 70% of babies and children under two years old use devices with screens
The phenomenon also coincides with early and widespread exposure to screens. Kapwing collects data from the British organization 1001 Critical Days Foundation according to which more than 70% of babies and children under two years of age use devices with screens, and one in ten regularly falls asleep in front of one.
Science, health and history also present a high saturation
After children’s content, the categories with the greatest presence of low-quality AI videos were Science and Education, with 35%; Health, with 33.8%; and History, with 33.5%.
Concentration on these areas adds an informational dimension to the problem. A poorly generated entertainment video may simply provoke rejection or indifference, but a health, historical or scientific explanation with invented data can be misleading and compete directly with sources prepared by specialists.
Automated generation makes it easy to create images, presenters, voices and scripts at a speed that is difficult to match through human production. However, this ability also allows large volumes of content to be published without verifying the accuracy of each statement.
At the opposite extreme are categories in which human creation continues to be clearly the majority. Only 1.6% of the analyzed fitness videos, 1.5% of the musical videos and 1.3% of the fashion videos were classified as low-quality automated content.
TikTok’s first impression before the algorithm learns
TikTok’s recommendation system adjusts based on signals such as watch time, skipped videos, likes, followed accounts, and topic preferences. But none of those variables are available when a person first opens an account.
Kapwing’s experiment sought to observe precisely that initial phase. The first post shown to the account created for the investigation was already a low-quality AI-generated video that, according to the researchers, was later deleted.
After reviewing the first 500 recommendations, the team determined that 294 fell into that category. Kapwing interprets the result as a consequence of the general saturation of the platform: in November 2025, TikTok had already labeled 1.3 billion videos as generated by artificial intelligence.
The initial experience can influence the perception that the user develops about the platform and the signals it provides to the algorithm. If automated content occupies a majority of the first contact, those aesthetics and those narrative formulas may end up being presented as the cultural normality of TikTok. Researcher Roland Meyer uses the concept of “platform realism” to describe how generative models do not directly reproduce reality, but rather an aggregation of the patterns already present in the human content of platforms. The result tends to homogenize styles, structures and themes that have already demonstrated the ability to attract attention.
It should be noted that TikTok introduced a tool in November 2025 that allows users to increase or reduce the amount of artificial intelligence-generated content that appears in their recommendations. The company then defended that some people enjoy this type of publications and that they should have greater capacity to adjust their presence.
The platform also announced a $2 million educational fund to support content on literacy and safety related to artificial intelligence. It has also reduced the use of some automatic summaries after users detected notable errors in the generated descriptions.
The report also reflects the negative reaction of many users. In some AI-generated posts
For Kapwing, these measures “passive” They do not solve the underlying problem. Research data suggests that automated content continues to occupy a central space in the experience of new accounts and in categories aimed at especially sensitive audiences.
The report also reflects the negative reaction of many users. In some posts generated with AI, comments that reject this type of content accumulate more interactions than the video itself, a sign that the high volume of production does not necessarily imply good reception.
Norway bans generative AI in primary education
The study appears in a context of growing institutional concern about the relationship between minors, learning and generative tools. From the end of August, Norway will prohibit the use of generative artificial intelligence for students from first to seventh grade, corresponding approximately to ages between six and thirteen years old.
The Norwegian Prime Minister, Jonas Gahr Støre, has justified the decision by the risk that these tools allow minors to skip fundamental phases of learning. “The most important thing in school is that our children learn to read, write and do mathematics”he pointed out.
Students between fourteen and sixteen years old will only be able to use generative AI under the direct supervision of a teacher. From the age of seventeen, educational authorities will encourage autonomous and appropriate use of these tools.
Although the school ban does not prevent minors from accessing generative tools from their homes, it does establish an institutional limit in a controlled space. And artificial intelligence can expand the narrative, visual and educational resources available to creators. The problem arises when automation replaces human review, precision, and judgment, and when recommendation algorithms reward that mass production before assessing its usefulness.
