Mozilla's new browser extension RegretsReporter aims to crowdsource research into harmful YouTube recommendations
RegretsReporter will let users send in reports about recommended harmful videos — YouTube's recommendations algorithm can lead you down some very weird rabbit holes …
Context & Ripple Effects
When The Verge covered RegretsReporter's launch, Mozilla was answering years of outside reporting rather than inside data: the Wall Street Journal had already documented that YouTube's engine, driving over 70% of viewing time, surfaces divisive and false content even to users who never asked for it, and YouTube's own product chief Neal Mohan had discussed the rabbit-hole effect in an interview about radicalization. Mozilla's move turns Firefox users into the data-collection layer the company itself lacks — a reversal of its earlier Advance experiment, which used machine learning to recommend pages rather than audit someone else's recommendations.
First-order effects
- Users gain a one-click way to file regret reports directly to researchers, giving Mozilla a labeled dataset of harmful recommendations that YouTube's internal metrics never surface.
- YouTube faces, for the first time, a systematic external record of which recommended videos users themselves judge harmful.
Second-order effects
- The crowdsourced data lets Mozilla test YouTube's own controls against reality — its later study of 20,000+ users and 500M+ videos found the 'dislike' and 'not interested' buttons barely change what gets recommended, undermining the platform's preferred account of user control.
- A follow-up 10-month Mozilla study reported more than 71% of participant-flagged objectionable videos were recommended by YouTube itself, shifting the debate from 'do users seek this out' to 'does the algorithm supply it' — directly challenging the framing of the earlier Mohan interview.
Third-order effects
- If the pattern holds, recommender-system accountability moves to independent, browser-mediated audits: civil-society researchers build the measurement infrastructure that platforms and regulators cannot or will not provide, and platforms' self-reported moderation claims lose their monopoly on evidence.
The trend: Platform algorithms are increasingly held to account not by regulators with subpoena power but by nonprofits crowdsourcing audit data through browser extensions.