Study finds TikTok feedback only briefly reshapes For You recommendations
Northwestern researchers found TikTok’s “not interested” tool cut unwanted videos, but the effect faded unless users kept giving the same signal.
By James Whitfield · Staff Writer
3 min read
TikTok users can influence what appears on their For You Page, but a new study found that control may be short-lived. Northwestern University computer scientists reported that the app’s recommendation system responds to negative feedback at first, then may drift back toward the same kinds of videos unless users keep objecting.
The findings come from a paper published in the Proceedings of the Twentieth International AAAI Conference on Web and Social Media. The research examined TikTok’s default feed, the For You Page, which uses personalized recommendations based on user behavior.
According to the paper, TikTok’s system weighs both direct actions, such as likes, follows and “not interested” responses, and indirect signals, including whether a person watches or skips a video. The study focused on complaints from users who said unwanted topics continued to appear even after they had tried to reject them.
Researchers tested TikTok with cloned accounts
The Northwestern team did not study real users or run a simulation. Instead, co-author Levi Kaplan told Ars Technica that the researchers created automated accounts on the actual TikTok mobile app using emulated devices.
Kaplan said the team collected metadata by intercepting network traffic and used a large language model to help decide how the accounts should respond to videos. He said the model’s decisions were checked against human responses.
Co-author Piotr Sapiezynski told Ars Technica that the group used that approach because official research access does not show how recommendations change for a specific user after particular actions. He said aggregated data would not allow the team to study personalization from the viewpoint of one account.
The researchers ran repeated experiments across 90 cloned accounts. They compared how TikTok responded when accounts sent direct negative feedback or gave indirect signals, such as skipping videos.
The study examined three categories of content: cooking, fitness and sports betting. The researchers then measured how much those topics appeared in the accounts’ For You feeds after different kinds of feedback.
“Not interested” worked better than skipping
The paper found that TikTok’s “not interested” button reduced unwanted topic recommendations by about 84 percent. Skipping videos reduced them by about 48 percent.
Kaplan told Ars Technica that users who want to reduce a topic should use the “not interested” option. The researchers also said the feature appears to be difficult to find in the app’s design.
The study found that negative feedback did not produce a lasting reset on its own. Sapiezynski told Ars Technica that the platform initially shows fewer videos from a rejected topic, but may gradually restore them to the feed if the user stops giving negative feedback.
The researchers also found that renewed engagement can quickly undo earlier signals. If an account watches content from a topic it had previously rejected, TikTok may begin recommending more of it again, according to the study.
Kaplan told Ars Technica that users can be taught better ways to use the platform, but the choices available to them are still shaped by TikTok’s design. The researchers said they hope to test their findings with real user data in future work.
This story draws on original reporting from Ars Technica.