TikTok’s “not interested” button is supposed to give users agency over their For You Page (FYP). A Northwestern University computer science team now reports a harsher reality: negative feedback can work, but only temporarily. After that initial adjustment, the algorithm gradually “relapses” unless a user keeps giving the same feedback over and over again.

That finding matters because TikTok’s FYP is the default home screen for most users, and it is the core product experience the company optimizes. The feed is personalized and algorithmically driven, relying heavily on implicit signals like how long users watch particular videos, plus explicit signals like likes or follows. In other words, the system is designed to infer what you will engage with next. So when users say the “not interested” tool does not remove the kind of videos they dislike, researchers decided to test the claim instead of arguing with anecdotes.

To be clear, the study is not saying TikTok is ignoring negative feedback entirely. The engagement signals do have an effect. The problem is durability. Northwestern’s paper finds that the algorithm’s response to negative signals declines over time. If you stop repeating the same “I do not want this” signal, the recommendations drift back toward the content pattern it would have selected without that sustained pushback.

The research group specializes in “algorithm audits,” a niche but increasingly important practice in online platforms. The auditors focus on how recommendation systems work, how they fail, when they fail, and how those failures can harm individuals and societies. In this case, they were motivated by “multiple anecdotal reports” from TikTok users who described the same pattern: even when they do not watch a suggested video, or they click “not interested,” those posts keep appearing on their FYP.

That tension is the core of the debate: what does “agency” mean in a system that is primarily driven by implicit behavior? TikTok’s approach is not just about what you say you want. It is about what you do. If you skip a video, how long do you watch before skipping? If you interact with an adjacent clip, does that still create engagement signals? If the recommendation system is weighted toward implicit feedback like viewing time, a one-time explicit negative signal might be treated like a temporary bump in your profile rather than a durable redefinition of your preferences.

The study also sharpens the question of why the platform provides negative feedback controls at all. Co-author Piotr Sapiezynski told Ars that the team wanted to understand why the “not interested” option exists if it does not reliably remove the disliked content. His comment highlights an uncomfortable mismatch decision-makers should recognize: users may interpret product features as levers that give them control, while the underlying model may treat those levers as weak or short-lived inputs.

For executives, this is more than a UI debate. Recommendation systems sit at the intersection of growth incentives and user trust, and the FYP is the engine that converts attention into ad value, retention, and engagement. If users feel trapped seeing unwanted content, it can translate into reputational risk, regulatory attention, and potentially changes to how platforms communicate user controls. And in the broader regulatory landscape, regulators across jurisdictions have increasingly scrutinized how platforms handle user consent, preference signals, and transparency around algorithmic curation. Even when regulators do not prescribe exact model behavior, they often target user understanding: can people meaningfully steer the experience they are being shown?

Second-order implications are easy to underestimate. A “temporarily effective” negative feedback mechanism can create a hidden cost for users. Instead of steering once, they may need to curate continuously, watching for recurrence and repeating the same negative action until the model stabilizes. That is a different kind of burden, and it becomes especially relevant for groups trying to avoid sensitive content categories. It also changes how community managers and product teams should think about feedback loops. If negative signals fade, moderation tools and content ranking safeguards may require different design patterns, and policy discussions may increasingly focus on persistence and efficacy, not just the existence of a button.

The strategic stakes are clear for boards, investors, and peer operators: TikTok’s FYP is demonstrating something many recommendation systems struggle with, preference drift and feedback discounting over time. If users believe they can control the feed, but the algorithm gradually returns to earlier patterns, that gap between perception and reality can become a trust problem. Northwestern’s audit suggests the gap is not imagined. And as algorithmic feeds become the default interface for millions of people, the real question for leaders is whether “agency” is a genuine feature of the system, or a short-lived suggestion that fades as soon as users stop fighting the model.