How Engagement Signals Often Fail to Identify Valuable Sensitive Content

Raw text

Transcribed text as recorded, including headers and formatting fragments.

Case 4:22-md-03047-YGR Document 2757-21 Filed 02/20/26 Page 1 of 17 AMENDED Exhibit 229 PLAINTIFFS' OMNIBUS OPPOSITION TO DEFENDANTS' MOTIONS FOR SUMMARY JUDGMENT Case No.: 4:22-md-03047-YGR MDL No. 3047 In Re: Social Media Adolescent Addiction/Personal Injury Products Liability Litigation Case 4:22-md-03047-YGR Document 2757-21 Filed 02/20/26 Page 2 of 17 Document Provided in Native Format HIGHLY CONFIDENTIAL (COMPETITOR) META3047MDL-047-01030576 Case 4:22-md-03047-YGR Document 2757-21 Filed 02/20/26 Page 3 of 17 A How Engagement Signals Often Fail to Identify Valuable Sensitive Content FB App Integrity October 2021 1 Case 4:22-md-03047-YGR Document 2757-21 Filed 02/20/26 Page 4 of 17 Case 4:22-md-03047-YGR Document 2757-21 Filed 02/20/26 Page 5 of 17 Case 4:22-md-03047-YGR Document 2757-21 Filed 02/20/26 Page 6 of 17 Ranking & Human Psychology 4 Case 4:22-md-03047-YGR Document 2757-21 Filed 02/20/26 Page 7 of 17 Weight for 2020H1 Too Brief, Simple Description of Ranking Feed (and comment threads) are ranked, generally, to show users what we think they will find most valuable Based on research, we try to predict what content users will find valuable Much of this is based on user engagement with content, where different types of engagement confer different amounts of value If content gets lots of engagement, our models generally assume that is because users like, want, and value it [1] "The algorithm" and how it works [2] Table from Lada, Mocanu, & Bower, 2020 Component Likes 1 Reaction 1.5 Reshares 1.5 RSVP maybe 4 Sticker Comment 12 Non-significant Comment / Message / Avatar Sticker 15 Significant Comment / Message 20 RSVP going 20 5 Case 4:22-md-03047-YGR Document 2757-21 Filed 02/20/26 Page 8 of 17 • • • Generally, engagement signals value, but on sensitive topics it may mislead In some domains, like civic, politics, and even in some health domains (e.g., vaccines, COVID-19) people tend to form social networks comprised of people who disproportionately share their beliefs and do not fact-check their personal beliefs In these echo chambers, people polarize, and are more likely to embrace conspiracy theories, hate speech, and false news (Del Vicario et al., 2016; Ramchandani et al., 2020) Ο This stems, in part, because highly similar people love and support ideas that they believe, which may be hateful of a minority group or political party, or may be skeptical of vaccine science Consistent with this, we see that user agreement with the beliefs stated in comments predicts their perception of that comment being “high quality” New Confirm Expand SOURCE Echo Familiaychamber Familiar Explove New CONTENT https://arxiv.org/abs/2007.08078 http://www.cs.toronto.edu/~ashton/pubs/alg-effects-spotify-www2020.pdf Case 4:22-md-03047-YGR Document 2757-21 Filed 02/20/26 Page 9 of 17 Case 4:22-md-03047-YGR Document 2757-21 Filed 02/20/26 Page 10 of 17 Case 4:22-md-03047-YGR Document 2757-21 Filed 02/20/26 Page 11 of 17 Case 4:22-md-03047-YGR Document 2757-21 Filed 02/20/26 Page 12 of 17 Case 4:22-md-03047-YGR Document 2757-21 Filed 02/20/26 Page 13 of 17 Case 4:22-md-03047-YGR Document 2757-21 Filed 02/20/26 Page 14 of 17 Case 4:22-md-03047-YGR Document 2757-21 Filed 02/20/26 Page 15 of 17 Case 4:22-md-03047-YGR Document 2757-21 Filed 02/20/26 Page 16 of 17 Case 4:22-md-03047-YGR Document 2757-21 Filed 02/20/26 Page 17 of 17