Northwestern University computer scientists found that TikTok users can push unwanted topics out of their For You Page, but only if they keep pushing. In experiments using 90 cloned accounts on TikTok’s real mobile app, the researchers reported that negative feedback changed recommendations at first, then weakened over time unless the accounts kept giving the same signal.
The finding matters for users who think a single tap on “not interested,” or a quick skip, should be enough to tell TikTok to stop serving a category of video. According to the Northwestern team’s paper in the Proceedings of the Twentieth International AAAI Conference on Web and Social Media, TikTok’s recommender responds to those signals, but it also keeps weighing later behavior, including renewed viewing of the same kind of content.
TikTok’s For You Page is the app’s default feed. Unlike social feeds built more visibly around follows, TikTok’s ranking system is known for relying heavily on behavior that users do not spell out, such as how long they watch a video, alongside explicit actions such as likes and follows. The Northwestern study looked at what happens when users send negative signals instead.
Piotr Sapiezynski, a co-author of the paper, told Ars Technica that his group studies “algorithm audits” to understand how online platforms operate and how they can harm users and societies. He said the TikTok project grew out of user complaints that negative feedback did not seem to keep unwanted videos away.
The team did not run a toy simulation of TikTok. Co-author Levi Kaplan told Ars Technica that the researchers used emulated devices, created accounts, and interacted with TikTok through code. The group intercepted network traffic to collect metadata, then used a large language model to decide how each account should respond. Kaplan said the model outputs were checked against human responses.
The researchers used cloned accounts because platform-provided research access did not answer the question they were asking. Sapiezynski told Ars Technica that TikTok’s official researcher API does not cover user agency in individual feeds, and that European Union researcher access provides aggregated data rather than the timeline-level view needed to study personalization.
Skipping helped less than explicit rejection
The experiments focused on three common categories: cooking, fitness, and sports betting. The researchers compared how TikTok responded when accounts skipped videos versus when they used the “not interested” control.
The explicit button worked better. The team found that “not interested” reduced unwanted videos by about 84 percent, while skipping cut them by about 48 percent. Kaplan told Ars Technica that users who want less of a topic should use the button. The paper’s authors also said the control appears to be hard for users to find, which is a design choice with consequences.
The harder part is durability. Sapiezynski told Ars Technica that TikTok initially shows fewer videos from a rejected topic, then gradually starts adding that content back. If a user watches it again, even briefly, the recommender can increase that topic in the feed again, according to the researchers.
The study does not prove how every real user’s feed behaves. The accounts were automated test accounts, not people with long histories, habits, and messy preferences. The researchers said they hope to test the same idea with real user data later. For now, their work suggests TikTok gives users some control over the For You Page, then makes them keep paying attention to preserve it.
This story draws on original reporting from Ars Technica.
