Reels Integrity: Sizing the Challenges

Raw text

Transcribed text as recorded, including headers and formatting fragments.

Case 4:22-md-03047-YGR Document 2757-14 Filed 02/20/26 Page 1 of 25 AMENDED Exhibit 208 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 2757-14 Filed 02/20/26 Page 2 of 25 Reels Integrity: Sizing the Challenges Alison Lee, Ph.D (RES) DS) depobook.com EXHIBIT 5 1 Case 4:22-md-03047-YGR Document 2757-14 Filed 02/20/26 Page 3 of 25 For internal use only - do not distribute! 2 Case 4:22-md-03047-YGR Document 2757-14 Filed 02/20/26 Page 4 of 25 I reviewed the related body of Reels, integrity, and well-being (equity, teens) research (link) to identify key priority areas for Reels. TLDR: Part 1 of this review highlighted the following challenges for integrity on Reels: 1 Fast-Moving and High Volume development 2 Entertainment, Mimicry and Youth 3 Demographic Variations of Integrity Case 4:22-md-03047-YGR Document 2757-14 Filed 02/20/26 Page 5 of 25 Timeline - Reels Integrity Review Scoping and Prioritization Part 1 Review Existing Literature Part 2 Extant Data Investigations: use existing, internal data (FRX and prevalence data) coupled with account demographic data (gender, zip code, age, country) to size and scope integrity challenges April 2021 May 2021 June 2021 July 2021 Aug 2021 Part 1 & 2 to inform key product roadmap priorities for RI in H2 2021 4 Case 4:22-md-03047-YGR Document 2757-14 Filed 02/20/26 Page 6 of 25 There are multiple (15+) teams working on Reels launching a high volume of product improvements (200+ on 2021 H1). This product pace and complexity is leading to greater volatility in integrity. Case 4:22-md-03047-YGR Document 2757-14 Filed 02/20/26 Page 7 of 25 Reels is a quickly evolving surface, with lots of We released a high volume (200+ in H1 2021) of product launches, 30+ leading to integrity regressions. The pace at which these launches occur make it difficult for us to identify the additive, interaction or long-term integrity effects of each launch. Multiple Teams simultaneously work on Reels. Including the Reels Recommendation Relevance team (3 sub-teams), Reels & Camera team (4 sub-teams), Reels Sharing (3 sub-teams), Reels Engagement, Family Reels, as well as other non-surface specific teams such as Integrity, Well-being, and Equity. moving parts. Gaps on recommended surface experimentation cause unexpected integrity impacts. Example: biases between CEPd and PD can cause disparities between QE pretests and post-launch topline impact. 6 Case 4:22-md-03047-YGR Document 2757-14 Filed 02/20/26 Page 8 of 25 Subsurfaces on Reels have different integrity challenges. Example: Reels Tab has higher prevalence of regulated goods, Cover of disturbing content, and RIFU of sexually suggestive content. Reels presents new and unique System for identifying harmful content in Reels is harder than other content review systems. integrity system challenges. Systems for identifying harmful content in short form videos works less well compared to images and text, takes longer, and is harder to judge/less accurate. Features of Reels (see next section) create unexpected challenges for integrity models. For example, the fast paced nature of viral trends require faster recurring training for models to keep up with non-recommendable spikes. 7 Case 4:22-md-03047-YGR Document 2757-14 Filed 02/20/26 Page 9 of 25 1 So what? Now what? 2 3 How can teams more proactively consider integrity impacts in their product decisions and strategy, both in early design phases as well as in process (i.e., using guardrails)? How can teams communicate and launch new product features strategically so that we can move fast and in sync without compromising safety? How might Reels product teams understand where (and for whom) disproportionate harm is limiting engagement? What improvements would lead to reduction of harm and more compelling Reels experiences? 8 Case 4:22-md-03047-YGR Document 2757-14 Filed 02/20/26 Page 10 of 25 Trends, Entertainment, and Youth: great for engagement, not so great for integrity. Who's at greatest risk? Case 4:22-md-03047-YGR Document 2757-14 Filed 02/20/26 Page 11 of 25 Part 1: Trends, mimicry, and entertainment are leading to increased risk of harm. Ranking models cause integrity regressions by increasing the prevalence of non-recommendable content. Increasing topic diversity increased all objectionable content, and recommending based on P(click), Mimicry, and downranking watermarked media increased the prevalence of sexually suggestive content. Trends, Entertainment, and Mimicry can lead to harm. Risky trends (i.e., the TikTok WAP trend for teens or trends mocking disabilities) may leave people open to harm, ranging from seeing harmful content to bullying or predatory behaviors. Offensive content (hate speech, violence, bullying) may also be delivered under the guise of "entertainment" at higher rates on Reels. Unconnected dancing videos featuring young girls raise safety concerns. Studies and internal feedback suggests that people are uncomfortable with seeing and concerned about the safety of unconnected Reels content of tween and young girls dancing and posing. 10 Case 4:22-md-03047-YGR Document 2757-14 Filed 02/20/26 Page 12 of 25 Part 2 Data Insights: Objectionable and Sexually suggestive content are higher on Reels. Sexually suggestive (non-rec), hate (violating), and violent (violating) content are both reported and detected at higher rates on Reels. (cit 1; cit 2) There are also significant discrepancies between what we detect, what users report, and what reports are actioned on, suggesting over/under-enforcement challenges. Meaningful Interactions are highly related to harmful content. Interactions such as clicks, likes, private shares, and comments are associated with sexually suggestive and hate speech content across IG. (link) Trends contribute significantly to integrity challenges. 12-23 of Top 50 sexually suggestive media are part of trends (via weekly review). However, boosting trends overall seems to result in integrity wins (i.e., reduction of prevalence). Sexually Suggestive Content featuring Minors (SSCM) is 2-3x higher on Reels than Explore*. Female and unknown gender teens, as well as unknown gender non-teens, create the most SSCM. However, we're also over-enforcing by labeling birth announcements and parenting videos under this category. (DM me for quip on Protecting Youth!) 11 "for data pulled from 04/01-04/30 A Case 4:22-md-03047-YGR Document 2757-14 Filed 02/20/26 Page 13 of 25 Part 2 Data Insights (cont'd): Users perceive and respond to different kinds of harm differently on Reels. Profanity, Misinfo, and Civic Misinfo continued to be the most prevalent issues, but they also made people feel least bad compared to other integrity issues... ...While witnessing Hate, Bullying, and Violence on Reels were among the most upsetting. Integrity issues may cost us engagement- people most often report scrolling past, exiting Reels, and exiting Instagram when having a negative experience on Reels rather than using other remediation tools. (1, 2) The fourth most used remediation behavior, selecting "not interested", provides both users agency over their experience and gives us signals about how to improve. We should invest in these personalization tools to improve engagement. The highest perceived reach integrity types are the least detected, and vice versa. People didn't often report the highest labeled prevalence integrity types (sexually suggestive, violence, graphic content). Similarly, hate speech and bullying are some of the lowest detected prevalence on Reels, but participants reported witnessing those more often and report them as being more upsetting. 12 Case 4:22-md-03047-YGR Document 2757-14 Filed 02/20/26 Page 14 of 25 1 How might we leverage multiple ML approaches (relevance, non- recommendable content, age modeling) to identify when particular trends may leave people - especially young girls - particularly vulnerable to harm? So what? Now what? 2 What expansions in non-recommendable policies might we need to implement to ensure ranking efforts are not pushing unpleasant or uncomfortable trends or content to users? 3 What design (self-remediation) and engineering approaches might better support safer experiences? (e.g., giving people more control over recommended content, limiting who can see content featuring young people, etc.) 13 Case 4:22-md-03047-YGR Document 2757-14 Filed 02/20/26 Page 15 of 25 Teens, women, LGBT people, and communities of color are more vulnerable to harm and marginalization. How do we solve this in Reels? 14 Case 4:22-md-03047-YGR Document 2757-14 Filed 02/20/26 Page 16 of 25 Teens are the target audience for Reels, but we have limited knowledge, tech, and tools to keep them safe. But, here's what we do know. Teens hear about and see integrity issues at higher rates on Reels, but they're less bothered by them than adults. Teens especially see and hear about more profanity and nudity, as well as (civically-related and other) bullying and hate. Teens don't think deeply about safety risks until something bad happens. Consider this with 1) how frequently content featuring minors shows up on Reels recommendations (see slide 9), 2) teens are fueled to produce by likes and comments, and 3) teens are more willing to talk to people who are "friends of friends". Teens want more control over who sees their content on Reels While this finding focused on audience control out of social risk (i.e., being embarrassed by content), it aligns with research on teens' preference for privacy, control, and awareness to address harm. 15 Case 4:22-md-03047-YGR Document 2757-14 Filed 02/20/26 Page 17 of 25 Part 2 Data Insights: Teens on Reels... Teens report (23-60%) more objectionable content across all violation types... compared to their DAU (Except for misinfo) ...but don't consume more objectionable and sexually suggestive content. (Except Regulated Goods content on Reels cover, which they view more than non-teens). Female and "Unknown Gender" Teens create 73% of "SSCM" content created by teens... ...While male teens report the most. "Custom Gender" Teens both create (1.2%) and report (1.9%) "SSCM" content at higher rates than their DAU (.4%). But the group who creates the most SSCM, creating 32% of ALL SSCM content, are "unknown gender non-teens". 16 Case 4:22-md-03047-YGR Document 2757-14 Filed 02/20/26 Page 18 of 25 How do we tackle the multiple Youth Safety issues at play here? 1 "True" SSCM: Improving our detection of minors, as well as getting ahead of problematic trends (i.e., upcoming dance trends) that may put children at risk of producing suggestive content can help us keep them safe. 2 Benign but still risky content featuring minors: Children dancing/posing unto itself is not problematic, but it becomes an issue if bad actors see it. Tolerance for this content varies by country and is not preferred for many non-teen users. Preventing IIC actors from seeing/interacting with minors, more effective / child content personalization (SFPLT) or creating "teen- centered" ranking approaches for teens may help solve this problem. 3 Baby nudity. Much of what we're proactively detecting are baby and parenting videos that users do not find problematic. We should examine whether we need to revisit policies around baby nudity. 17 Case 4:22-md-03047-YGR Document 2757-14 Filed 02/20/26 Page 19 of 25 Inequities in Integrity: We must start by naming that integrity work is inherently equity work. Anytime we talk about harm, we cannot avoid its explicit connection to broader societal patterns of marginalization, representation, and power and therefore, our role in addressing inequity of safety, agency, and belonging on our platform. 18 Case 4:22-md-03047-YGR Document 2757-14 Filed 02/20/26 Page 20 of 25 Part 1: Inequities in safety, belonging, and engagement are replicated in Reels. Harm not effectively covered by our policies has led to users perceiving IG as unsupportive, exclusionary, and silencing. Users perceive that harmful content targeting marginalized groups are under-enforced and voices of marginalized groups are over- enforced across IG. Lack of Representation and Social Comparison are barriers to engagement on Reels. Low distribution, lack of representation, and no credit to original creators lead to under-engagement and people believing that "Reels is not for people like me." A culture of "non-relatability" and social comparison on Reels is discouraging for everyone, but may especially impact teen girls and non-binary teens. Global differences in integrity needs require inclusive solutions. Different countries vary on how they (esp. young women) experience and respond to harm. Empowering global users would require that integrity tools are designed for varying levels of language fluency and global needs. 19 Case 4:22-md-03047-YGR Document 2757-14 Filed 02/20/26 Page 21 of 25 Case 4:22-md-03047-YGR Document 2757-14 Filed 02/20/26 Page 22 of 25 Case 4:22-md-03047-YGR Document 2757-14 Filed 02/20/26 Page 23 of 25 Case 4:22-md-03047-YGR Document 2757-14 Filed 02/20/26 Page 24 of 25 Case 4:22-md-03047-YGR Document 2757-14 Filed 02/20/26 Page 25 of 25