Radical bubbles on YouTube? Revisiting algorithmic extremism with personalised recommendations

Abstract

Radicalisation via algorithmic recommendations on social media is an ongoing concern. Our prior study, Ledwich and Zaitsev (2020), investigated the flow of recommendations presented to anonymous control users with no prior watch history. This study extends our work on the behaviour of the YouTube recommendation algorithm by introducing personalised recommendations via personas: bots with content preferences and watch history. We have extended our prior dataset to include several thousand YouTube channels via a machine learning algorithm used to identify and classify channel data. Each persona was first shown content that corresponded with their preference. A set of YouTube content was then shown to each persona. The study reveals that YouTube generates moderate filter bubbles for most personas. However, the filter bubble effect is weak for personas who engaged in niche content, such as Conspiracy and QAnon channels. Surprisingly, all political personas, excluding the mainstream media persona, are recommended less videos from the mainstream media content category than an anonymous viewer with no personalisation. The study also shows that personalization has a larger influence on the home page rather than the videos recommended in the Up Next recommendations feed.

Recommending instead of taking down: YouTube hyperpartisan content promotion amid the Brazilian general elections

“…A VPN application protects a user's browsing information by anonymizing their activity and geographic location, hiding the user's real IP address, and, in this case, simulating a different IP with a random geographic location within Brazil in each test (ExpressVPN, n.d.; Frary, 2016). No user was logged on to a Google account in any of the visits, nor did we use the account for watching videos, constructing search histories to avoid building a profile of preferences (Ledwich et al, 2022). We aimed to test YouTube's algorithm recommendation bias as a statistical dice problem (Pomeranz, 1984).…”

“…Since recommendations can direct users to videos they would not normally have selected (Alfano et al, 2021), recent studies have investigated to what extent YouTube might work as a “radicalizing instrument” (Ledwich et al, 2022; see also Tufekci, 2018; Yesilada & Lewandowsky, 2022). Thus understanding the amplification of content is key to investigating information consumption on the platform.…”

Reducing Exposure to Harmful Content via Graph Rewiring

“…and societies perceive the world and form their opinions [12,21,36,42,44]. In recent years, platforms have come under increasing scrutiny from researchers and regulators alike due to concerns and evidence that their recommendation algorithms create filter bubbles [6,26,28,45] and fuel radicalization [19,27,39,41,49]. One of the main challenges in this context is dealing with content that is considered harmful [4,7,50].…”

Exploring YouTube’s Recommendation System in the Context of COVID-19 Vaccines: Computational and Comparative Analysis of Video Trajectories

“…Earlier research has relied on anonymous, impersonal accounts without any watch history to audit YouTube’s recommendation system. However, to address the lack of user activity data, some researchers have intentionally curated user profiles to simulate personas—distinct demographics, content preferences, and watch histories—in order to generate personalized recommendations [ 26 ]. For example, to investigate the effects of personalization on the amount of misinformation in YouTube search results, Hussein et al [ 27 ] built a crawler that watches a curated subset of the videos returned by search queries in order to build the watch history of their user profiles.…”