Chat: Our age models are totally blind to this extraordinarily obvious exploit for the very abusive type we're trying to solve for

Raw text

Transcribed text as recorded, including headers and formatting fragments.

Case 4:22-md-03047-YGR Document 2887-16 Filed 03/27/26 Page 1 of 4 AMENDED Exhibit 340 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 2887-16 Filed 03/27/26 DOCUMENT SLIPSHEET Bates Number: META3047MDL-047-00531116-META3047MDL-047-00531117 Custodian: Author: Filename: MESSAGE SUMMARY [{"OTHERUSERFBID":NULL,"THREADFBID":4517677971579541}] Create Date: 6-25-2020 1:25 PM Last Modified Date: 6-25-2020 1:25 PM TNAG BEGBATES: TNAG ENDBATES: MAAG BEGBATES: MAAG ENDBATES: depobook.com EXHIBIT Meta- Castaneda-zo Page 2 of 4 Message Case 4:22-md-03047-YGR Document 2887-16 Filed 03/27/26 From: [/O=THEFACEBOOK/OU=EXTERNAL Sent: To: (FYDIBOHF25SPDLT)/CN=RECIPIENTS/ 6/25/2020 8:25:59 PM Subject: Message summary [{"otherUserFbId":null,"threadFbId":4517677971579541}] A] Page 3 of 4 >Hey (6/25/2020 09:52:18 PDT): - any thoughts/progress on this? Conscious that this data point will be helpful in establishing the right window of time and particularly if we're considering leaning on reporting more after carve- out convo this am (6/25/2020 09:52:38 PDT): >Thanks, (6/25/2020 09:53:11 PDT): >Thanks, and FYI I'm running down the CEI/IIC TAT with Child Safety co EOD is when I expect numbers back (6/25/2020 10:24:01 PDT): >hey, we were able to track down the select messages data in android which will give us at least some directional guidance here (best we can do at the moment). we're dealing with some data format nonsense (life of a DS really...) but i'll have something for you before tomorrow (6/25/2020 11:29:59 PDT): >thanks so much! (6/25/2020 12:47:50 PDT): >99 did Core Dims confirm that their age models are optimized to catch age-lying minors and not age-lying adults? I thought it would be operational for both sets of age-liars. (6/25/2020 12:48:28 PDT): >yep, definitely confirmed that their focus has been on misrepresenting minors and there has never been much focus on misrepresenting adults (6/25/2020 12:50:40 PDT): >i think this is largely an artifact of the history of product use cases for age models - facebook generally has gated minors out of experiences, so an adult lying about being a minor just gets gated out, no harm there. it is the integrity use case which is unique in that both kinds of misrepresentation are very important, and the lying adult may be much worse (6/25/2020 12:51:26 PDT): >And they are, in fact, bad at predicting age lying adults? B/c if the most relevant signal is the average age of your friend graph, age-lying adults could also be fairly predictable. Although, I know from experience that age-lying adults who do so to groom children create fake minor profiles for this explicit purpose and would usually have majority minor friends in their graph. (6/25/2020 12:51:45 PDT): >(Just asking in case their model, even though trained one way, happened to work the other way too) (6/25/2020 12:56:13 PDT): >not only are they bad at predicting them, they don't really know how bad they are because of a lack of high quality labelled data (6/25/2020 12:56:26 PDT): >but they know it's very bad (6/25/2020 12:57:32 PDT): >> age-lying adults who do so to groom children create fake minor profiles for this explicit purpose and would usually have majority minor friends in their graph > >this is a big part of the problem. i can't confidently talk about "most relevant signals" but i know that the age of the people in your network is definitely a very important signal our (6/25/2020 12:59:05 PDT): >all the other nuance and stuff aside, this exact thing reason i think age carve outs won't work age models are totally blind to this extraordinarily obvious exploit for the very abusive type we're trying to solve for CONFIDENTIAL META3047MDL-047-00531116 Case 4:22-md-03047-YGR Document 2887-16 Filed 03/27/26 (6/25/2020 12:59:21 PDT): >this exact thing is the reason* Page 4 of 4 (6/25/2020 13:02:28 PDT): >yeah, totally hear you on that b/c it's true: that fact pattern is a significant part of the highest severity harms, in addition to the IRL adult who is using messaging to abuse a minor he/she has access to or can persuade to meet up for IRL sex abuse. BUT the other part of the picture is the large numbers of adults who are quite brazen and honest about how old they are when reaching out to kids as demonstrated by this undercover investigation where a woman pretended to be an 11 year old and within days had 94 adult men reach out to her soliciting sexual contact: https://fightthenewdrug.org/went- undercover-on-instagram-online-sexual-exploitation-happens/ (6/25/2020 13:03:12 PDT): >but this kind of case can and should be solved for by breaking _ unconnected adult<>minor contact (6/25/2020 13:03:14 PDT): >I think this is especially a problem on IG. I'd encourage you to watch the short documentary video (6/25/2020 13:03:36 PDT): >we are actively working now on breaking ig message requests from unconnected adults to minors g (6/25/2020 13:03:50 PDT): >it's a mitigation we have on msgr and want to get igd to parity (6/25/2020 13:04:28 PDT): >yes, absolutely. Do we know ETA on that in the Interop launch timeline? I'd think there are some age model dependencies there as well (6/25/2020 13:05:34 PDT): >there are age model dependencies but they are less stringent. because we can tolerate many more false positives in that flowW (6/25/2020 13:05:58 PDT): >krish is the best person to ask about timelines there, but i believe it is considered launch blocking for 1m2 (6/25/2020 13:06:24 PDT): >the reason we wanted to try out best to age gate minors out of the early smoke tests is because this mitigation would not be ready before the smoke tests (6/25/2020 13:08:02 PDT): >the key driver behind the priority is actually to prevent a regression on msgr by suddenly allowing adults on IGD to send requests to teens on msgr (6/25/2020 13:08:55 PDT): >(there may be a decoupling between bringing the mitigation to x-network vs bringing it to igd native in- network) (6/25/2020 13:15:32 PDT): >there seem to be 4 populations of abusive adults we need to think through > >1) unconnected adults who we know are adults >2) connected adults we we know are adults >3) unconnected adults who we do not know are adults >4) connected adults who do not know are adults > >i'd be very comfortable advocating for using our age models, even with very low thresholds, even as imperfect as they are, to help with (1) which is probably "highest volume". (3) and (4) feel fundamentally unsolved today with any of our current ideas if we just don't know the abuser is an adult to begin with, how do we mitigate the abuse? (2) would be "solved" by hard age carve outs, but it would come at absolutely tremendous relative cost to the product, and i feel like we could do better than that > >does that summary make sense? (6/25/2020 13:22:26 PDT): >Yes, perfectly. Thank you!! (6/25/2020 13:25:59 PDT): >+1 on everything you guys covered :) CONFIDENTIAL META3047MDL-047-00531117