Researchers tracked 900 users’ viewing to match YouTube recommendations with actual choices

Evidence

US researchers from Dartmouth College in New Hampshire, and Northeastern University in Massachusetts, recently sought to solve this issue in a novel way. Through tracking people’s online behaviour they matched the recommendations YouTube made to each user with an individual’s actual viewing choices. They sourced a panel of 4,000 US citizens, including ‘frequent’ YouTube viewers (who were more likely to be younger and male) and people who, on a pre-test, had already expressed negative views on race and gender. Overall, about 900 of these individuals agreed to take part in the study. They asked participants to install a special browser plug-in extension that recorded data from websites they visited, including their browser history. The team then looked at visits to about 200 well-known ‘alternative’ YouTube channels – which often provide a gateway to more extreme content – as well as various outright ‘extremist’ or white supremacist channels. Overall, the researchers were able to extract about 800,000 YouTube visits which resulted in over 4 million recommendations that were made by the YouTube algorithm for each user. — Sander van der Linden, Foolproof, 410

Categorization

• SEED: A usable method note for any piece on algorithm-audit research design: sampling only ‘frequent viewers’ with prior expressed bias risks preselection caveats worth interrogating.
• XREF: connects to: YouTube’s recommender drives up to 70 percent of content watched. — Sander van der Linden, Foolproof: Why Misinformation Infects…, 401