A study by Mozilla using YouTube recommendation data from 20,000+ users and 500M+ videos finds the “dislike” and “not interested” buttons have little impact
Even when users tell YouTube they aren't interested in certain types of videos, similar recommendations keep coming, a new study by Mozilla found.
Context & Ripple Effects
YouTube introduced topic controls and a channel-level opt-out in 2019, but Mozilla’s earlier crowdsourced work found that participants’ objectionable videos were frequently recommended by YouTube itself. The new dataset tests whether the feedback tools themselves meaningfully alter that recommendation path.
That matters because prior coverage described recommendations as driving more than 70% of viewing time and as capable of surfacing divisive or misleading material to users who did not seek it out through recommendations.
First-order effects
- YouTube users who use dislike or “not interested” signals receive little practical control over the similar videos that follow, according to Mozilla’s analysis.
- Mozilla gains a large-scale evidentiary basis to challenge YouTube’s claim that its visible feedback controls give users meaningful influence over recommendations.
Second-order effects
- YouTube’s 2019 topic and channel controls become a weaker trust-and-safety defense if the underlying recommender does not materially honor adjacent negative feedback.
- Creators and advertisers remain exposed to a distribution system where recommendation placement, rather than a viewer’s stated preferences, has outsized influence over what audiences encounter.
Third-order effects
- If repeated audits produce the same result, video platforms’ user-control features risk being treated as interface-level consent rather than substantive governance of algorithmic distribution.
- The pattern points toward greater scrutiny of whether platforms can demonstrate that feedback mechanisms change recommendation outcomes, not merely that the controls exist.
The trend: Recommendation platforms are facing a shift from judging user-control tools by their availability to judging them by their measurable effect on automated distribution.