1 of 11

Exploring how YouTube recommends science videos to diverse audiences

Shiyu Yang

Assistant Professor

School of Journalism & Media

46^th Annual CCI Research Symposium

Faculty RAP

2 of 11

The algorithmically curated information environment

3 of 11

Wide education- and race-based gaps in Americans’ science knowledge

4 of 11

How does the algorithm distribute science & health information to diverse audiences?

How do we look inside the “black box” of algorithms and measure opaque algorithmic influences on different audiences?

5 of 11

Sock puppet algorithm audit of YouTube

YouTube platform

Users

Researcher sock puppets

6 of 11

Sock puppet experiment: �2 race (White vs. Black) × 2 SES (high vs. low)

2. Train sock puppet by engaging with content known to be preferred by race/SES groups

3. Search three science issues (HGE, AI, COVID-19 vaccines) and collect top 20 video recommendations for each search

4. Compare search recommendations across experimental groups

Black White

Low-SES High-SES

1. N = 840 sock puppets, 210 in each group

2. Set geolocation in Milwaukee, WI, USA

7 of 11

High-SES (White) sock puppets receive a greater diversity of video and channel recommendations

8 of 11

Overall, 47% of the unique videos are NOT recommended to all four audience groups

57% of unique AI videos
38% of unique COVID-19 vaccine videos
20% of unique HGE videos

9 of 11

Topics of videos uniquely received by �one audience group but not the others

White

Black

High-SES

Low-SES

Benefits of HGE (e.g., therapy, sustainable food supply, reforestation)
AI golf lesson
COVID vaccine data & statistics
Debunking vaccine misinfo/controversy

AI rap music
AI & basketball
AI not working well for the dark-skinned
Supporting vaccine misinfo/constroversy

Dangers of AI (e.g., privacy, misuse by government & politicians)

AI-generated audiovisuals faking politicians (e.g., Biden, Trump, Obama)

10 of 11

Why this matters?

Algorithms exposing diverse audiences to different subsets of science & health info, even when they are actively searching for the same thing

Especially for issue that is heavily discussed on social media
Implications for disparities in scientific understanding
GPT and LLM?

Social media algorithms, such as the YouTube algorithm, can indeed expose sociodemographically diverse audiences to different subsets of information even when people are actively searching for the same science issue, although the degree of information tailoring may depend on the search topic itself (e.g., how heavily the science issue is discussed by different sources on YouTube). Although low-SES audiences already tend to be informationally disadvantaged, the algorithmically infused media environment may expose them to a narrower range of science information than it does to their higher-SES, informationally resourceful peers, thus exacerbating disparities in science information and understanding. While disparities in scientific understanding between high- and low-SES social segments may be attributed to differences in learning ability or to the human tendency to select attitude-confirming information, findings from this study suggest that such disparities may also be caused by algorithms that target human preferences and expose population segments to different subsets of information without their own free will.

11 of 11

@syang364

syang53@utk.edu

Questions?

Full paper title: Biased algorithm? Exploring how YouTube recommends science videos to racially and socioeconomically diverse audiences