IG IIC Prevalence Analysis -

Raw text

Transcribed text as recorded, including headers and formatting fragments.

Case 4:22-md-03047-YGR Document 2887-39 Filed 03/27/26 Page 1 of 5 AMENDED Exhibit 377 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 2887-39 Filed 03/27/26 Page 2 of 5 IG IIC Prevalence Analysis - a/c priv This document contains the analysis for the IG IIC Prevalence Analysis within Direct Messages. Related Docs Main research plan doc: IG IIC Prevalence Estimate Top-line Results Top-line We estimate that there are 3.5M profiles [95% Cls: 2.9M to 4.1M] conducting Inappropriate Interactions with Children (IIC) among all IG DM Sender MAP who sent a message to a minor (a population of 320M). In percent terms, this is 1.1%. (See Figure 1 for the flow chart.) IG DM Sender MAP Sent a DM in November 2018 622M Messaged a minor As determined by ANY of 4 age signals 320M 22 countries we reviewed 233M Other countries Based on available markels Random Sample Based on available markets No reviewers 10k- لـ Assume same rate Assumptions and caveats: "Potential Groomers" Review. There are a large amount of accounts marked as "potential groomers" (18M [16.6M to 19.5M]). In these cases, the reviewer was either unsure of the profile's user age, or there was ambiguity in the conversation. We will be re-submitting the "potential groomers" to a double review to be extra sure on this number, and it is possible that the IIC numbers could increase with additional review. We have already conducted double review on accounts marked as IIC. Countries we didn't review. We were only able to review profiles in 22 markets due to labelling resources. This group represents 73% of this population (233M/320M). The above 3.5M number is based on assuming that the countries we did not review have the same rate of IIC activity as the 22 countries we did review. If we do not make this assumption and restrict to the 22 countries we reviewed, our estimate is 2.55M [2.1M to 3.0M] IIC profiles (1.1% of this group). Disagreement Data. We conducted double review on accounts marked as IIC. We currently look at the last review if there was re-review present or at the expert review and treat this review as the ground truth. In general, we are trusting the expert review and placing a high weight on their decision. This is consistent with Harmful Behavior's methodology on FB (Blue). We are investigating this assumption, and it could revise our estimates. Missing Labels. There are a small number of missing labels due to technical issues with enqueuing content into Single Review Tool. We assume the missing labels are "missing at random" and not systematically biasing the results. We cannot find any systematic bias after investigating in detail. The number of missing labels is small, and we would have to assume there is a massive amount of systematic bias for this issue to have any effect on the top- line results. Next Steps: HIGHLY CONFIDENTIAL (COMPETITOR) META3047MDL-056-00002578 Case 4:22-md-03047-YGR Document 2887-39 Filed 03/27/26 Details Additional reviews Missing Iraq enqueue (~170) re-review of Potential Groomer (~500) missed review (13 not reviewed + 71 not enqueued) Disagreement labels: incorporate uncertainty when 2 labels disagree Breakdown by type of Groomer label and sub-labels Our population is IG DM Sender MAP who sent at least one DM to a minor. The numbers we report here are unique IG profiles. This is not the same as unique people. Total IG DM Sender MAP in November 2018: 622M [query] Who sent a DM to a minor: 320M [query] Page 3 of 5 • minor = ANY of our age signals identified them as a minor (FB stated age (if IG account linked to FB), FB age prediction (if linked), IG age affinity, IG Adult Classifier). • in 23 countries where we had reviewers: 236M [query]; 22 countries (excluding Iraq, where we had a bug enqueuing the content): 233M IG DM Sender MAP Sent a DM in November 2018 622M Messaged a minor As determined by ANY of 4 age signals 320M 22 countries we reviewed 233M Other countries Based on available markets No reviewers Random Sample Based on available markets 10k---- لـ Assume same rate Results Of those who sent DMs in Nov 2018 to a minor from the 22 countries we reviewed, we see: * 2.55M groomers (1.1%) [95% Cls: 2.1M to 3.0M], 2.0M missing label (0.85%) [1.6M to 2.4M), 13.2M potential groomers (5.6%) [12.0M to 14.2M], 75.7M minor age liars (32.6%) [73.5M to 78.2M], 139.8M non-groomers (59.9%) [137M to 142M] https://fburl.com/daiquery/3z4dnvvq We were only able to review 22 countries representing 72.8% of all IG DM Senders who sent to minors (233M/320M). If we assume that there is the same proportion in each category among this remaining population (e.g., 1.1% of this population is groomers), then we should multiply the above numbers by 1.37 (1/0.7281), and our total among the 320M who sent a message to a minor is: 3.5M groomers (1.1%) [95% Cls: 2.9M to 4.1M], • 2.7M missing label (0.85%) [2.2M to 3.3M], 18M potential groomers (5.6%) [16.6M to 19.5M], * 104M minor age liars (32.6%) [101M to 107M], • 192M non-groomers (59.9%) [189M to 195M] Since the remaining population did not send a DM to a minor (302M), we can assume that there were no groomers in this population. The above represents our estimate for IG DM MAP Senders in November 2018. Assumptions and Caveats HIGHLY CONFIDENTIAL (COMPETITOR) META3047MDL-056-00002579 Case 4:22-md-03047-YGR Document 2887-39 Filed 03/27/26 Page 4 of 5 Profiles, not users. Our estimate is at the profile-level. We do not have an estimate of the distinct number of people who are conducting IIC on IG. Countries we didn't review. The rate of each group is uniform between the 22 countries we reviewed (72.8% of IG DM MAP Sender) with the countries we did not review. We could test this by looking at the 22 countries we already reviewed and checking how uniform grooming activity is across country. Reporting both numbers (75% vs. 25%) Missing Labels. We assume the missing labels are missing at random and not systematically biasing the results. In that case, we can effectively add 1.1% groomers of 2.7M missing to the groomer total, which is effectively rounding error, and we can ignore this. If we discover any systematic bias, this may change our estimate. 2 groups of missing data: enqueued but not reviewed, and never enqueued. Potential Groomers. The potential groomers are cases where the reviewer was either unsure of the profile's user age or where there was ambiguity in the conversation. We will be re-submitting the "potential groomers" to a double review to be extra sure on this number, and it is possible that the groomer numbers could increase with additional review. 500 re-reviews of these profiles Disagreement Data. We currently look at the last review if there was re-review present or at the expert review and treat this review as the ground truth. In general, we are trusting the expert review and placing a high weight on their decision. We could investigate this assumption more and understand how it would change the top-line numbers. We know that the bias is typically in the direction of more false positives (reviewers have a tendency to see IIC when it is not present). We could incorporate uncertainty when 2 labels disagree by counting it as 0.5 groomer. We could also provide upper and lower bounds by only counting groomers when ANY reviewer indicated it was violating (upper bound) or when ALL reviewers agreed it was violating (lower bound). Gradations/severity of "Groomer" label. Can we differentiate levels of IIC behavior? For example, fraction of the profiles reviewed that were reported to NCMEC. I need the mapping to the below types. There are 2 groomer labels ("Groomer" and "groomer_escalate"). I'm unclear on the difference. We will break down by additional groomer types. HIGHLY CONFIDENTIAL (COMPETITOR) META3047MDL-056-00002580 Case 4:22-md-03047-YGR Document 2887-39 Filed 03/27/26 Page 5 of 5 Sexualized Conversation Child Exploitation Images Meeting/Planning to Meet for sex Position of Trust Suicidal Ideation Sibling Abuse Sextortion Prostitution of a minor Bestiality Notes 8 The population is defined in: D13440310 grp map dap 1 all dm 622,739,079 236,509,643 2 teen_receivers 320,425,593 93,909,415 https://fburl.com/daiquery/x40г714c Next Steps Disagreement labels: incorporate uncertainty when 2 labels disagree Breakdown by type of Groomer label and sub-labels Potential groomers re-review Contact Support Conversation History Quipbot#15 left the thread Nov 6 at 7:16 pm HIGHLY CONFIDENTIAL (COMPETITOR) META3047MDL-056-00002581