TikTok FYP “not interested” only works briefly, then the feed relapses

TikTok users who hit “not interested” may be getting a very specific kind of placebo. A recent paper from Northeastern University computer scientists found that engagement signals do affect what shows up on a For You Page (FYP), but only temporarily. After that short window, the algorithm gradually relapses unless users consistently give the same negative feedback over and over again.

That matters because TikTok’s FYP is not just another feed. It is the default home screen, personalized through an algorithmically driven content stream. TikTok’s approach depends heavily on both implicit signals, like how long a user watches a particular video, and explicit signals, like likes or follows. In general, the algorithm does a remarkably good job predicting which videos will interest particular users. But the new research targets a pain point TikTok users have been voicing: even when they do not watch suggested videos, and even when they click “not interested,” those same types of content can keep resurfacing.

The paper’s core claim is essentially a timing problem. Negative feedback is not ignored. Instead, it seems to be overwritten by the system’s ongoing assessment of what keeps people engaged. The algorithm audit suggests that the impact of the feedback decays. If a user gives a single “less of this, please” response, the feed behavior drifts back. Only repeated, consistent signaling appears to prevent that drift.

This is where the research framing gets especially sharp. The group at Northeastern University specializes in “algorithm audits,” an approach focused on understanding online platforms at the level of how they work, how they fail, when they fail, and how they harm individuals and societies. In other words, this is not just a UX complaint study. It is an attempt to test whether the interface creates meaningful agency for users, or whether it provides a control that users expect to steer outcomes, even when it does not.

The co-author Piotr Sapiezynski told Ars Technica that their motivation came from multiple anecdotal reports from TikTok users. Those reports described a mismatch between what the interface promised and what users experienced. On the surface, “not interested” is a clear, user-driven negative preference signal. The puzzling part, as Sapiezynski put it, is that if negative feedback does not remove posts from the FYP reliably, then it is unclear why the platforms would offer the option at all.

From an executive perspective, this is a second-order governance and risk issue, not just an algorithm detail. If users believe they can correct recommendations but the system only responds temporarily, then “control” becomes part of the user experience, not the reality. That can shape user trust, retention, and reputational risk. And it can also complicate compliance conversations. Regulators increasingly care about the mechanics of how platforms curate feeds and how user controls function. Even when a platform claims personalization, the question becomes whether “negative feedback” is operationally effective in a stable way, or whether it fades because other signals dominate.

There is also a product incentive tension hidden inside the findings. TikTok’s success depends on keeping people watching. When the algorithm treats watch time as a primary implicit signal, negative feedback can only do so much against the volume of engagement data coming in continuously. The audit suggests that engagement signals shift recommendations, but the system then returns to its prior equilibrium once the temporary effect wears off. For decision-makers, that is the uncomfortable implication: user agency might exist, but it can be fragile, conditional, and cumulative. In practice, “agency” becomes a task users must keep doing.

For peers across social platforms, creators, and investors, the strategic stakes are straightforward. Personalization engines are becoming a core layer of how people discover information, entertainment, and community. When user controls behave like short-term nudges rather than durable preferences, it can influence exposure patterns long enough to matter. And in a world where algorithmic feeds are scrutinized for both consumer harms and societal impacts, a “gradually relapses” behavior pattern is exactly the kind of technical nuance that later becomes a policy headline.