TikTok Algorithm's Negative Feedback Loop Questioned by Northwestern Study

TikTok’s For You Page (FYP) is the default home screen for users of the video-sharing platform. It’s a personalized, algorithmically driven content feed, but the approach differs from other social media in that TikTok’s algorithm relies heavily on implicit signals—such as how long users watch particular videos—as well as explicit signals such as likes or follows. And generally, that algorithm does remarkably well at predicting which videos will interest particular users.

But some users have voiced concerns that TikTok’s almighty algorithm doesn’t seem to incorporate negative feedback very well. Even when they don’t watch a suggested video or click the “not interested” feature, they keep seeing those videos on their FYP. Northwestern University computer scientists put those suspicions to the test. According to their recent paper, the engagement signals do have an effect, but only temporarily. Then the algorithm gradually relapses unless a user consistently gives the same feedback over and over again.

The research group specializes in “algorithm audits,” co-author Piotr Sapiezynski told Ars, to better understand online platforms: “how they work, how they fail, when they fail, how they harm individuals and societies.” In this case, he and his co-authors wanted to take a closer look at user agency after hearing multiple anecdotal reports from TikTok users that their negative feedback—responding to prompts by indicating they aren’t interested or want to see less of a certain kind of video—doesn’t seem to remove those posts from their FYP. “On the other hand, it’s unclear why the platforms would offer it, if it doesn’t work,” said Sapiezynski.

Their methodology did not involve computer simulations; rather, they created bot accounts on the actual TikTok mobile app, rather than studying actual users. “We used emulated devices, where we are creating accounts and automatically interfering with the TikTok algorithm through code with the sock puppet accounts,” co-author Levi Kaplan told Ars. “We’ve come up with a methodology where we get the metadata by intercepting the network traffic, and then we make a decision using an LLM. All the LLMs were validated with human responses as well.”

“We basically work from the assumption that if we want data, then we need to obtain it ourselves,” Sapiezynski said of their account cloning approach. “Even if we did, for example, want to use the official TikTok researcher API, none of the user agency is covered there. You can see what content is available, but you cannot see individual timelines that will tell you how the algorithm reacts to a particular user watching or not watching a particular video. Similarly, with the European Union’s researcher data access, all of this data can only be accessed aggregated and not from a perspective of a single user. So when you want to really study personalization, this research cannot be done on the aggregated data.”

Mind the gap

The team ran their experiments multiple times on the 90 cloned accounts and made side-by-side comparisons, using both implicit and explicit signals, to see how TikTok’s algorithm responded in terms of recommended content on the FYPs. They focused on three popular topics: cooking videos, fitness videos, and sports betting.

The “not interested” button proved most effective, reducing unwanted content by around 84 percent, compared to just a 48 percent reduction from merely skipping videos. “So if you don’t want to see something, you should be hitting that button,” said Kaplan. But the authors note that the “not interested” option seems to be deliberately hidden from users. And it was very easy for the algorithm to “relapse” into once again flooding an FYP with previously unwanted content; even a brief re-engagement by a user is sufficient.

“It turns out that it works in the beginning,” said Sapiezynski. “When you start saying, ‘I don’t want to see this particular topic,’ the platform might actually show you fewer of such pieces of content. But then the platform will slowly start putting it back in your feed. And if you don’t continue saying, ‘I really don’t want to see it,’ this may balloon back to the place where it was in the beginning. So the platform does react to your negative feedback, but then it also very much reacts to your express behavior. So if you are presented with this content again and you start watching it, the platform will again feed it to you more and more.”

In other words, be consistently very active with your feedback—constant vigilance!—when it comes to curating TikTok’s FYP. The researchers hope to test this hypothesis on real user data in the future. That said, “We can teach users how to use the platform better, but ultimately the way that you’re interfacing with the platform is going to be dictated by the design decisions that are fundamental to the platform,” said Kaplan.

Proceedings of the Twentieth International AAAI Conference on Web and Social Media, 2026. DOI: 10.1609/icwsm.v20i1.42688 (About DOIs).