Problematic Use On-Platform Measurement

Raw text

Transcribed text as recorded, including headers and formatting fragments.

Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 1 of 33 AMENDED Exhibit 109 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 social impact well-being Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 2 of 33 Problematic Use On-Platform Measurement Concentration survey results for IG Dec 2022-Jan 2023 Elena Davis and X-Meta Well-being 1 Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 3 of 33 TL;DR - IG To advance our understanding and measurement of potential problematic use (PU) on our platforms, this workstream: 1) Developed and validated new survey measurement of users' perceived control of their use of Facebook and Instagram 2) 3) Surveyed ~30K youth and ~30K adult users of IG in 12 countries Identified on-platform behaviors associated with perceived low control Findings 1) Users' perceived control can be validly measured across age cohorts with questions assessing low attentiveness around platform use (i.e., spending more time, checking, and consuming content without realizing it) and low capability to change platform use (i.e., low confidence in ability to reduce time, checking, or interactions) 2) Rates of perceived low control are similar for youth and adults, with low attentiveness (~34-40%) more common than low capability (13-20%) 3) 1 Perceived low control nearly fully explains the small association between on-platform behaviors and perceiving that one's platform use prevents fulfilling other responsibilities in life (‘life interference'). 4) The ability to predict perceived control using logged platform behaviors is modest, but improved over past PU models: 0.77 AUC for attentiveness (precision/recall ~0.5/0.7) and 0.80 for capability (~0.4/0.4). 5) Global behaviors associated with perceived low control reflect high usage, evening + late night time spent, and specific behaviors, like no engagement repeat sessions and video chaining. 6) Specific product usage associated with perceived low control includes content consumption surfaces (Explore, Reels Tab, Home) and active engagement (Stories creation, Direct). Takeaways These survey items can be adapted for use in product evaluations (designed on a case-specific basis) to understand perceived impact of products. While there are no objective cut-offs for defining 'low' control, we see that this is a general issue experienced by both youth and adults. Youth-first products can be explored to extend to gen pop. Focusing on addressing areas on our platforms where users are likely to experience perceived low control is a viable strategy for making progress in this space. We have improved predictive precision by focusing on control, but on-platform behaviors likely cannot be used to directly measure (or proxy) perceived control. Our next step is to iterate with the team on appropriate cut-offs (e.g., p90 usage) and prioritized behaviors (e.g., evening + late night time) to use for heuristics and goaling, noting that these are not definitional of PU. Existing or novel control solutions may be provided in-context within these products, such as nudges within Reels or extensions of Quiet Mode to support limits and boundaries around active engagement. 2 1 Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 4 of 33 Approach 4 Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 5 of 33 Context: Success for problematic use workstream Overall success for PU is: People have the tools that they need and are empowered to control their time and experiences on our platforms. Measurement will encompass: Signals of product efficacy + user value (e.g., behavior change metrics; product engagement; perceived impact) and positive perception (e.g., perceived usefulness; product awareness) [to be aligned with measurement framework for all Youth + Well-being efforts across the company] 1) Priority Detail Status Meet youth industry practice requirements Ensure we are providing PU-related product solutions that are on par with the industry, based on regulatory expectations and competitor feature sets IG feature set has largely met this bar, pending confirmation that products provide educational information in context (details TBD) 2) Measure success of products 3) Identify and close additional gaps Ensure that the features are a) effective; b) bringing user value; c) reaching the relevant users Develop updated behavioral metrics to use in goaling Identify product usage and/or behaviors that are associated with perceived low control for additional investment H1 2023 priority through behavior change goal, product evaluations and sizing, and product awareness research This report + next steps 6 Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 6 of 33 Goal of this workstream: Identify on-platform behaviors that are associated* with user perception** of low control Outline of results 1) Measurement approach and rates of perceived low control 2) General behavior patterns (time spent, sessions, time of day) associated with perceived low control 3) Specific product usage associated with perceived low control *All associations in this report are non-causal, i.e., there can be no conclusions drawn about the direction of how behavior and perception may impact each other. **All user perception measures come from self-report surveys on the participants' subjective self-assessment of their experiences with sense of control and their perceptions of the platform. 7 Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 7 of 33 Approach: Updated survey measurement of potential problematic use Focus is on measuring users' perception of control over their time and experiences on our platforms (see note) Measurement review of PU constructs & existing survey items Pilot test and cognitive test potential survey items covering key control constructs Refine survey items and field on-platform measurement survey Specify constructs for 'sense of control': low attentiveness and low capability Identify on- platform correlates of low control, and identify low control heuristics August Lit review Past surveys 1, 2 Measurement audit + takeaways List of items September - October Research plan Cog testing results Pilot results + notebook All items tested November - December Research plan Simon FB Simon IG December Results Analysis notebook This report December - February Analysis readout IG Heuristics 8 Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 8 of 33 We identified two constructs of perceived control and measured them in an on-platform survey Low attentiveness around use (possible unintended, unaware, and/or habitual use) Perceived control capability (perceived self-efficacy to exert control, if desired) These two areas of control are consistent with literature suggesting lack of control over social media use may come in two forms - habitual or low attention use; and lack of ability to control use (1, 2, 3). These also map to models of self-regulation of behavior that define two systems, an automatic one (akin to low attentiveness) and a conscious, goal-monitoring one (akin to exerting deliberate control) (4). See Appendix for details on survey development and item selection. 9 Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 9 of 33 Details of survey sample, fielding, and analysis Sample N=30,699 adults and 30,346 youth (13-17) on Facebook N=30,810 adults and 30,756 youth (13-17) on Instagram Across 12 countries: AU, BR, EG, GB, KR*, ID, IN, MX, NG, PH, TR, US (*KR was IG only) Data collection Sampling was stratified to be 50% users who exhibited 1 or more behaviors previously associated with PU (no engagement late night sessions, late night product switching, late night long sessions, no engagement repeat sessions, high frequency notification checking, or high volume of short sessions) and 50% users who exhibited none of those behaviors L7 Data was collected late November - early December 2022 (11/18-12/14 on FB; 11/21-12/8 on IG). Configs targeting less active users took longer to reach quotas. Analysis 1 We verified that survey items formed 2 factors and created composite scores by averaging across the 3 attentiveness items and the 3 capability items. Survey responses were weighted using Graviton weights to reflect the Facebook and Instagram user population across the 12 countries during the time frame of the survey, based on recommended indices from SWAMP: gender, country, age, indices of use (IG: followers, followings, creator, business, private account; FB: L30 + mobile and web, days since confirmed, friend count, subscription count profile picture), and device. Associations between on-platform signals and survey measures were analyzed in 3 ways to confirm converging signals: unique predictors in a multivariate regression, strongest contributors in a predictive decision trees (xgboost), and simple correlation 10 Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 10 of 33 Results: Instagram 12 Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 11 of 33 Summary of users' report of perceived low control How common are experiences of perceived low control on Instagram? Rates of reporting the lowest scores (<=2 on 5 pt composite scale; or 1 or more lowest responses on any item): Low attentiveness: Low capability: youth, adults youth; adults How do these experiences differ across ages and markets? Adults were slightly more likely to report low attentiveness; youth were slightly more likely to report low capability (small significant differences) Markets differ: higher perceived control in western/high income (US, GB, AU); lower in some emerging markets (NG, EG, ID) and also relatively low in KR - possibly due to cross-cultural differences in perceptions of control and role of social media in people's lives How does perceived low control relate to other perceptions of the app? Users reporting low control also report higher 'life interference' from their app use (i.e., perceiving that use of Instagram prevents the fulfillment of other responsibilities). Low attentiveness: higher life interference; low capability higher life interference. Low control is not strongly related to perceiving value from the app or views on whether the app is trustworthy or not (correlations < .14) Descriptive details starting here 13 Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 12 of 33 Perceived control explains how platform use may be related to perceived 'life interference' Previous measurement work has included a focus on ‘life interference' as an important outcome related to problematic use; i.e., perceiving that one's platform use is preventing the fulfillment of other life responsibilities. Here we see that perceived life interference is almost entirely related to actual platform use through perceived low control. We don't see an association between platform use and life interference except through perceived control. Platform use (all behaviors) Perceived [lack of] control : Low attentiveness Low capability For teens: 96% of the relationship is explained by low control For adults: 91% of the relationship is explained by low control .. ............ Beyond the effects of perceived low control, platform signals only explain 0.7% of the variance in teen users reported life interference and 1.8% of the variance for adults. Perceived life interference 14 Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 13 of 33 Focusing on control increases our measurement precision This supports our approach to focus on control: from both a conceptual and measurement standpoint, perceived control over platform use is more proximal to the actual on-platform experience, including behaviors that we can measure and potentially change. Our overall models predicting perceived control are stronger than those that include life interference or other life impact measures, likely because there is less noise. Our updated heuristics will likely be better as well. Problematic use Focus is here Perceived lack of control over social media perceived to contribute to Negative life impacts use Control over time spent Control over experiences Precision Recall See 1-pager here AUC Former PU model on FB 0.17 all ages 0.73 all ages 0.66 all ages Outcome: low control + negative life impact Current PU model on IG 0.48 youth 0.51 adults 0.70 youth 0.69 adults 0.77 youth 0.78 adults Outcome: low attentiveness Current PU model on IG 0.36 youth 0.50 adults 0.45 youth 0.36 adults 0.81 youth 0.80 adults Outcome: low capability 15 Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 14 of 33 Specific behaviors associated with low perceived control on IG Explanatory power for low attentiveness overall is 2-3x as strong as low capability. Predictors largely consistent across youth and adults. Stack Behavior rank Attentiveness Capability Detail 1 Total time spent X X Time spent and sessions (higher overall use) are top predictors 2 Number of sessions X X 3 Evening + early night time spent X X (7-10 pm; 10 pm-12 am) 4 Passive evening + early night X X time spent (7-10 pm; 10 pm-12 am) Whereas past PU research implicated late night time spent (1, 2), this study also identifies evening time spent as a predictor of perceived low control, slightly stronger than late night time spent. Whereas past PU research found no association with sessions being more active or passive on average (2), certain no engagement sessions (repeat and late night) were found (1). Here we see passive sessions in evening and early night associated with perceived low control. 5 Late night time spent (12am- 4am) X X Overall late night time spent, as well as passive time spent (slightly less strong), emerging here, consistent with past work. (overall and passive) 6 No engagement repeat sessions X X Consistent with past heuristic work (1) and possibly related to checking behavior. Relatively stronger as a predictor for adults 7 Video chaining X New finding, a behavior specifically associated with perceived low attentiveness. Particularly related to frequency of chaining, rather than chaining depth. 8 Short sessions X New finding, a behavior specifically associated with perceived low capability. Associations were evaluated 3 ways to highlight converging signals: unique predictors in a multivariate regression, strongest contributors in a predictive decision trees (xgboost), and correlation 16 Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 15 of 33 Specific product usage associated with low perceived control on IG Stack Product rank 1 2 3 Attentiveness Capability Detail (Youth and/or Adults) (Youth and/or Adults) Y+A Y+A Y Y+A Y+A 4 Y Y+A 5 Y+A Y+A 6 7 Y+A Y A Y Associations were evaluated 3 ways to highlight converging signals: unique predictors in a multivariate regression, strongest contributors in a predictive decision trees (xgboost), and correlation 17 Takeaways Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 16 of 33 Next steps 1) Many behaviors are associated with both forms of perceived low control, and predictors are largely similar for youth and adults. 1) Findings are consistent with the stance around focusing on late night use as important to support user control; extending to evening use may be a next step. 1) Expanding solutions to adult cohorts may also be beneficial. 1) Specific behaviors like video chaining, particularly on passive consumption surfaces like Reels, might be opportunities for attentiveness nudges or adjusting autoplay settings. 1) New product opportunities may exist in-context for specific surfaces or targeting specific behavior patterns - such as extending Quiet Mode to support limits and boundaries around active engagement. 1) 1) Focus this broad list to an abbreviated, non-overlapping list of 3-5 specific behaviors with relatively highest precision/recall to form updated PU heuristics • Leverage team priority area around late night use Present alternatives and iterate on final set Define thresholds Evaluate sensitivity of heuristics: have we seen movement of these in experiments to date? 1) Opportunity size the potential impact of using these behaviors for additional targeting for Quiet Mode and Take a Break (e.g., expansion to evening hours) • • Given a certain targeting threshold, what % of TAM is reached? Update behavior change goals based on change in % of TAM reached and measured efficacy of the feature 18 Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 17 of 33 IG product details 46 Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 18 of 33 Theme: Continuous content consumption 4Specific product usage associated with perceived low control 5 Product Teens Low Attentiveness Adults Teens Low Capability Adults *Explore tab also includes behaviors like tapping into full screen + chaining. **Frequency of video chaining at any depth is associated with perceived low control. Frequently getting a few videos deep is twice as important in association than daily chaining depth (i.e., frequently getting a few videos deep matters more than occasionally getting 10-20+ videos deep). 3 Associations were evaluated 3 ways to highlight converging signals: unique predictors in a multivariate regression, strongest contributors in a predictive decision trees (xgboost), and correlation Tier of strength prediction is relative: Tier 1 = top 3 strongest predictor from regression and/or decision tree; Tier 2 = top 3 strongest predictor from one cohort or one form of low control; Tier 3 = top 10 predictor 47 Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 19 of 33 Theme: Active engagement Specific product usage associated with perceived low control 11 oduct Teens Low Attentiveness Adults Teens ***Direct time spent still a predictor even controlling for notifications. Low Capability Adults *Associations were evaluated 3 ways to highlight converging signals: unique predictors in a multivariate regression, strongest contributors in a predictive decision trees (xgboost), and correlation ** Tier of strength prediction is relative: Tier 1 = top 3 strongest predictor from regression and/or decision tree; Tier 2 = top 3 strongest predictor from one cohort or one form of low control; Tier 3 = top 10 predictor 49 Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 20 of 33 Theme: Other Specific product usage associated with perceived low control Product Teens Low Attentiveness Adults Teens Low Capability Adults *Associations were evaluated 3 ways to highlight converging signals: unique predictors in a multivariate regression, strongest contributors in a predictive decision trees (xgboost), and correlation ** Tier of strength prediction is relative: Tier 1 = top 3 strongest predictor from regression and/or decision tree; Tier 2 = top 3 strongest predictor from one cohort or one form of low control; Tier 3 = top 10 predictor 51 Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 21 of 33 Theme: Other Specific product usage associated with perceived high control Product High Attentiveness Greater usage of these products associated* with less frequently spending more time, checking more, and consuming more content than intended, without realizing it High Capability Greater usage of these products associated* with higher perceived ability to reduce or limit time spent, checking, and interactions, if wanting to Teens Home Share of time spent Adults Teens Adults *Associations were evaluated 3 ways to highlight converging signals: unique predictors in a multivariate regression, strongest contributors in a predictive decision trees (xgboost), and correlation ** Tier of strength prediction is relative: Tier 1 = top 3 strongest predictor from regression and/or decision tree; Tier 2 = top 3 strongest predictor from one cohort or one form of low control; Tier 3 = top 10 predictor 52 Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 22 of 33 Usage of products associated with perceived low control 7 Continuous content 52onsumption 3 plore tab 6 Reels tab Video chaining Home esp long passive Active engagement Stories creation Direct Notifications Other Profile Home share of time Content consumption in Explore, Reels, and Home - especially passive - is an experience within which we can offer control features. For example, Take a Break with more nuanced time spent targeting. Frequently viewing chained content, regardless of which surface, is associated particularly with low attentiveness. This makes sense as a product affordance that may contribute to people spending time or consuming content without realizing it (expert report). Importantly, depth of chaining is less associated. This would support enacting more limits on autoplay of next videos that occurs across surfaces, and/or adding pauses or friction. Occasional longer immersive viewing sessions may be more intentional especially if they are paired with features that support awareness and limits around that experience. Stories creation is the strongest predictor of perceived low capability to reduce use, and also associated with low attentiveness - though has been found to be a high intent behavior (1). Content creation was also identified in previous PU research (2), as highly active users are also more likely to create content. Creation likely leads to more responses and thus likelihood of returning to use the app, all in a time-bound ephemeral way. Direct messages have been previously found associated with PU heuristics and related survey outcomes, but also are related to high session satisfaction (3). Control opportunities here will have to find a balance. Notifications have been specifically pointed out as a possible trigger of or contributor to low control usage (4, 5, 6, 7); this may not only be due to initiating new sessions, but diversion within a session. Profile time spent was also identified in previous PU research (1), but unclear why. Past work found no differences in browsing own vs. other profiles (2). Profile use may be related to revisiting own past content; social curation, editing/updating (3; 4; 5). Low control experiences especially with passive profile time spent could be a particular form of content consumption. Share of time spent on Home associated with higher control might be due to trading off with time spent on other surfaces associated with low control. 53 Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 23 of 33 How well can we predict user self-reported perceived control using on-platform signals? Overall our predictive power for perceived low control is modest: precision of ~0.5 and recall of ~0.7 for attentiveness and ~0.4 on both precision and recall for capability. We can predict ~30% of the variance in capability and 37% of the variance in attention. 15 10 Attentiveness Precision Recall Capability AUC Precision Recall AUC All data - Youth .48 .70 0.766 .36 .45 0.807 All data - Adults .51 .69 0.775 .50 .36 0.801 Top 10 - Youth .40 .71 0.696 .23 .33 0.693 Top 10 - Adults .42 .71 0.695 .20 .29 0.674 55 Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 24 of 33 Global behavior correlations - FB youth Correlations surface_visits Attentiveness Capability -0.12 -0.05 engagements -0.06 -0.04 time_spent -0.14 -0.06 push_notif_session time_to_sess_from_push late_night_ts ts 47 -0.14 -0.07 0.06 0.05 -0.10 -0.02 -0.12 -0.05 ts_7_10 -0.12 -0.05 ts_10_12 -0.11 -0.04 late_night_sessions -0.12 -0.04 no_eng_sessions -0.11 -0.04 late_night_high_surface_switching_sessi ons -0.09 -0.03 notif_checks -0.10 -0.02 short_sessions -0.09 -0.03 repeat_no_eng_session -0.12 -0.06 late_night_no_eng_sessions -0.09 -0.03 no_engagement_4_7 -0.10 -0.04 no_engagement_7_10 -0.10 -0.04 no_engagement_10_12 -0.10 -0.03 57 Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 25 of 33 Boosted Decision Tree for Attentiveness - FB youth Feature (Top 10) Gain (marginal effect of predictive accuracy) Frequency (of behavior) Correlation photo_ts notif_ts_share nf_ts_share story_ts_share feedback_ts_share notif_checks photo_ts_share feedback ts ts_7_10 nf ts reels_ts_share engagements ts_4_7 push_notif_session ts_10_12 feedback_engagement_per_min no_engagement_4_7 58 Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 20 - Overall behavior patterns - Facebook 11 havioral signal 18 Low attention Strength of associations Low capability Strength of associations Teens Adults Teens Page 26 of 33 Adults Overall predictor strength for linking on- platform signals with low control are 5x stronger for predicting low attention than low capability; prediction for low capability overall low with push sessions + no engagement repeat sessions on top Total time spent ~6% ~6% ~1% N/A (not top 10 predictor) Sessions originating from push notification ~5.5% ~5.75% ~1% ~1% Sessions (number) Evening time spent (4-7 pm; 7-10 pm; 10 pm-12 am) Passive sessions (overall, and evening 4-7 pm; 7-10 pm) ~5% ~6% ~.75% ~.5% ~5% 4-7, 7-10 ~4.75% 10-12 ~5.75% 4-7, 7-10 ~1% 4-7, 7-10 N/A (not top 10 predictor) ~5% overall ~4.5% evening ~5.5% ~.6% ~.5-.6% Push notifications more salient as a predictor for FB compared to IG Interesting that this is right up there with number of sessions- could sessions originating from a push be overall less intentional sessions? Evening time spent in the top 10; late night use relatively less important 59 Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 27 of 33 Global behavior correlations - FB adults Correlations surface_visits Attentiveness Capability -0.15 -0.04 engagements time_spent push_notif_session time_to_sess_from_push late_night_ts ts 47 -0.07 -0.02 -0.15 -0.03 -0.15 -0.06 0.00 0.00 -0.09 -0.02 -0.13 -0.01 ts_7_10 -0.13 -0.03 ts_10_12 -0.11 -0.03 late_night_sessions -0.13 -0.04 no_eng_sessions -0.14 -0.04 late_night_high_surface_switching_sessi ons -0.10 -0.03 notif_checks -0.10 -0.02 short_sessions -0.12 -0.04 repeat_no_eng_session -0.14 -0.06 late_night_no_eng_sessions -0.09 -0.03 no_engagement_4_7 -0.13 -0.04 no_engagement_7_10 -0.13 -0.04 no_engagement_10_12 -0.11 -0.04 61 Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 28 of 33 Global behavior correlations - IG youth Correlations Attentiveness Capability ts ts_4_7 ts_4_7_passive ts_7_10 ts_7_10_passive ts_10_12 ts_10_12_passive -0.23 -0.11 -0.10 -0.06 -0.07 -0.05 -0.19 -0.10 -0.14 -0.09 -0.18 -0.09 -0.14 -0.08 ts_late_night -0.14 -0.07 ts_late_night_passive -0.11 -0.07 no_eng_repeat_sessions -0.14 -0.10 short_sessions_count -0.11 -0.08 late_night_no_engagement_sessions -0.09 -0.07 sessions -0.18 -0.11 sessions_exceed_2hr -0.07 -0.03 avg_max_chaining_depth -0.04 0.01 sum_chaining_depth_max -0.07 0.00 post_impressions_chaining -0.08 -0.01 media_impressions_chaining -0.09 -0.02 62 Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 29 of 33 Global behavior correlations - IG adults Correlations Attentiveness Capability ts -0.22 -0.11 ts_4_7 ts_4_7_passive ts_7_10 ts_7_10_passive ts_10_12 ts_10_12_passive -0.10 -0.05 -0.08 -0.04 -0.18 -0.10 -0.15 -0.08 -0.18 -0.08 -0.14 -0.07 ts_late_night -0.13 -0.08 ts_late_night_passive -0.11 -0.07 no_eng_repeat_sessions -0.17 -0.10 short_sessions_count -0.15 -0.09 late_night_no_engagement_sessions -0.10 -0.08 sessions -0.22 -0.10 sessions_exceed_2hr -0.04 -0.02 avg_max_chaining_depth -0.05 -0.02 sum_chaining_depth_max -0.04 -0.02 post_impressions_chaining -0.06 -0.04 media_impressions_chaining -0.07 -0.05 63 Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 30 of 33 Overall behavior patterns - Instagram 21 havioral signal Low attention Strength of associations Low capability Strength of associations Teens Adults Teens Adults Total time spent ~9% ~9% ~3.5% ~3.5% Sessions ~8% ~9% ~3.5% ~3% (number) Evening time spent (7-10 pm; 10 pm- 12 am) ~7% each ~7% each ~3% each ~3% 7-10 ~2.5% 10-12 Passive evening time spent (7-10 pm; 10 pm- 12 am) Late night time spent (12 am - 4 am) (overall and passive) No engagement ~6% each ~6% each ~3% each ~2.75% 7-10 ~2% 10-12 ~6% overall ~4.5% passive ~5.5% overall ~4.75% passive ~2.5% each ~2.5% each Overall predictor strength for linking on-platform signals with low control are 2-3x stronger for predicting low attention than low capability Predictors are largely similar across adults and teens Time spent and sessions (as indices of higher overall use) consistently a top predictor Evening time spent, both overall and passive time, as a stronger predictor than late night use (particularly stronger than passive late night use) ~5.5% ~7% ~3% ~3% No engagement repeat sessions relatively stronger as a 64 23 Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 31 of 33 Next steps 66 Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 32 of 33 Measurement next steps to get to updated heuristics/metric goals 24 25 These global behaviors can be used for targeting/triggering thresholds (?already in Feed Pause) 26 Next steps involve iterating on different versions of combined signals + thresholds to optimize precision/recall of perceived low control: ?-also include team priorities - eg., late night use 1) [Here are the heuristic identification] Stacking of individual behaviors a) Alternative decision tree models that are abbreviated: simplify the metrics included (e.g., between 3-5 predictors); ensure that each are independently meaningful b) Can use those decision tree branching to determine thresholds c) This will give a smaller set of behaviors that are all together the best predictors 1) Then redo thresholding based on what specific control product would be; e.g., if we were only targeting video chaining, what would the right threshold be? (e.g., behavior thresholds conditional on other targeting factors) a) This would inform product strategy for the contexts in which features are offered 1) Opportunity size the potential impact of improved targeting for QM and TaB would be, on behavior change outcomes (even beyond only late night use); e.g., given a certain targeting threshold, what % of TAM is reached? e.g., refine what a reasonable behavior change goal would be (...leveraging efficacy of products to date; ) – a) E.g., extending Quiet mode to earlier hours b) E.g., Take a Break triggering 68 Case 4:22-md-03047-YGR Document 2885-40 Filed 03/26/26 Page 33 of 33 Focusing on control increases our measurement precision This supports our approach to focus on control: from both a conceptual and measurement standpoint, perceived control over platform use is more proximal to the actual on-platform experience, including behaviors that we can measure and potentially change. Alternative model Problematic use Focus is here Perceived lack of control over social media perceived to contribute to Negative life impacts use Control over time spent Control over experiences See 1-pager here Baseline Precision Recall AUC Former PU model on FB Outcome: low control + negative life impact ~0.11 0.17 all ages 0.73 all ages 0.66 all ages Current PU model on FB Outcome: low attentiveness ~0.3-0.35 0.67 youth 0.57 adults 0.37 youth 0.44 adults 0.74 youth 0.69 adults Current PU model on FB 0.64 youth 0.25 youth 0.76 youth Outcome: low capability ~0.15-0.2 0.46 adults 0.24 adults 0.73 adults 70