YouTube’s recommender still steers viewers toward same-category harmful videos despite algorithm changes

Evidence

Interestingly, results revealed that, in general, people followed a recommendation about 20 per cent of the time. In terms of average content exposure, about one in ten (9 per cent) people viewed at least one video from an extremist channel and two in ten (22 per cent) a video from an alternative channel. Now consider that YouTube has over 2 billion monthly users worldwide. This finding points to the risky nature of YouTube’s architecture; anyone can now upload extremist content and compete with the mainstream media. Moreover, when people watched such content, almost 40 per cent of recommendations from alternative and about 30 per cent of the recommendations from extremist channels went to other videos of the same category – despite the fact that YouTube claimed to have made over thirty changes to their algorithm to reduce recommendations of harmful content. — Sander van der Linden, Foolproof, 411

Categorization

• XREF: Dovetails with prior chapters’ material on algorithmic filtering and echo chambers as amplifiers rather than root causes. — connects to: Automated recommendation feeds can expose teenagers to thousands of radicalizing videos — Sander van der Linden, Foolproof: Why Misinformation Infects…, 381; A majority of studies find YouTube’s recommender is a gateway to problematic content — Sander van der Linden, Foolproof: Why Misinformation Infects…, 415; Radicalization stories overstate the causal role of recommendation algorithms — Sander van der Linden, Foolproof: Why Misinformation Infects…, 387 (+1 more)

Broader Connections

• A majority of studies find YouTube’s recommender is a gateway to problematic content #corroborates (0.95) — Sander van der Linden, Foolproof, 415
• Radicalization stories overstate the causal role of recommendation algorithms #complicates (0.80) — Sander van der Linden, Foolproof, 387
• Creators hijack recommendation algorithms to expose radicalized users to rebuttals #corroborates (0.72) — Sander van der Linden, Foolproof, 385