"Non-personalized ads is not a tenable user experience and undermines AFS, given adoption for AFS is predicated on the fact that we are not able to offer non-personalized ads but if we do this for Youth, there may also be a push for us to do this for adults. An age-banded solution incurs similar risks with added complexity. "
"Teens are still developing, novelty-seeking and impulsive . . . Cool, so we probably shouldn't have ranking primarily driven by p(eyeballtime) and p(engagement), which is how it currently works."
Rori ████: "[E]arlier today a man in Bosnia live-streamed the murder of a woman on IG and claimed to have murdered at least two other people, after which he reportedly took his own life… it is possible Meta may be criticised for the length of time the live stream remained on the platform (approximately three hours). There is no evidence that automation detected the video prior to removal and operational teams do not believe it should have based on the footage."
"The man live-streamed himself shooting his ex-wife in the head, whilst a child could be heard screaming in the background, before he streamed a second video approximately three hours later in which he claimed to have killed at least two other people."
"The video of the user shooting the woman in the head was reported 1,071 times by users, and had 32K VPVs at the point it was removed 2 hours and 58 minutes after being posted… There is no evidence that automation detected the video prior to removal and operational teams do not believe automation should have identified it based on the footage."
Sayed Otaru: "Rocky's just confirmed that the FPS classifier was launched on IG in June. Our teams will work together to understand why we did not detect this… Amit, we agree that there are some significant Integrity risks on IG Live."
Bullying & harassmentChild sexual exploitationDirect Messaging
"Almost all participants thought that younger users and new users should be defaulted to private"
"• Almost all participants across ages believed that a young person's account (e.g., anyone below 11-13, and in some cases up to 15) should automatically be private. Several teens said that when they were younger, they had set up their first Instagram account as public only because it was the default and they weren't aware that private was an option. Looking back, they said that private would have been a better option for them.
• Many participants also said that all accounts should default to private, for everyone regardless of age. They expressed the perspective that one should have '"practice'" on Instagram before going public, and a specific reason/goal for going public.
• Teens who began with a public account, and later switched to private, were very certain that their younger siblings and cousins should start with a private account."
"Unwanted contacts were seen as an issue related to safety, security, and privacy"
"Malicious people could stalk them and compromise their safety including showing them things they didn't want to see or tracking their location; they could invade their privacy by following them without interacting (just spying/lurking); and could use devious interactions to compromise the security of their accounts, impersonate them, and exploit their personal information."
"The biggest pain point and concern related to both public and private accounts was receiving unwanted messages* from the least intimate accounts**. These included
• Unwanted messages that were sexual in nature were particularly upsetting
•Large group messages from unknown or barely known contacts were frustrating."
"Fake robot accounts DM you and say 'watch my free sex videos' and the person sends it to 50 different kinds of people...This was kind of like 'hey get a $500 Visa gift card', I did that one time, it was the first time I ever saw it, I was really young, I thought, 'I am going to get some money!' It asked me for my address, plain old little me, I was kind of ditzy, and what it said to me after that was 'We have your address now, we will stalk you.' I was eleven, I freaked out...That's probably one of the reasons I got a private account."
"Bullying was more common for teens on public accounts Most instances of bullying that were shared came from teens with public accounts."
"Being bullied created the wish to restrict interaction, but caution about self-remediation, because blocking, unfollowing or switching to private could be interpreted as an act of fear that would draw more bullying in person. Those being bullied needed an assertive way to constrain interaction from bullies that will not be interpreted as fearfulness."
"Recommendations • Defaulting all teen accounts to private mode, and explore defaulting all accounts to private mode"
"[W]e are still missing violating content when it is reported to us more often than we would like. Of the violating child safety content that is reported to us we successfully take down 31% on Facebook and 17% on Instagram."
Michel Protti: "You around? Any chance I can call you for 5 mins re: Arkansas?"
Michel Protti: "Will try quick message thread - my take on this escalation is that, if we get a stay, we should clearly hold on our compliance plan while we assess the landscape given how invasive the age verif requirements are and how bad the user experience is with blocking new users."
Michel Protti: "Are you advocating though that we should hold on compliance even absent an injunction?"
Antigone Davis: "I agree with your position. Absent an injunction is a tougher call from the policy perspective. +s and -s."
Michel Protti: "My take is: if we don't hear anything from the court we turn on our compliance plan as planned. if/when the court rules that there's a stay on State enforcement we would pull back our compliance plan and assess the landscape. we'd then monitor the landscape and legal developments to decide if/when we turn our compliance plan back on"
Antigone Davis: "If you mean the court rules against the injunction, then I remain okay with plan of record (but generally have worries about compliance requirements outpacing our ability to push for OS solution)"
Antigone Davis: "I think we wait to hear from the court. If they miss the effective date, they will quickly follow — that is the whole point of a preliminary injunction"
Michel Protti: "Rob and the rest of the policy team on the call were of the opinion that, if we hear nothing from the court by tomorrow 6pm, we should proceed with launching with our compliance plan (and pull it back if the injunction is granted)."
"Key observation themes are summarized below:
- Inconsistent suppression on data collection of contact information for teens during the business account conversion; Lack of suppression on data collection for users with unknown age
- Inaccurate and insufficient implementation of restrictive privacy default setting on Website Embeds for teens
- Inconsistent 1:1 and group messaging restrictions between teens and unconnected adults, including High Risk Adults"
"The IG Growth Acquisition team does not always default to the more restrictive privacy setting of 'opt-out' on Website Embeds for the U.S. teen public accounts under 16 years old. As a result, teens' posts and profiles may be publicly displayed outside the IG app. For example, IA observed that out of 564K U.S. under 16 public accounts registered in May 2023, 43.7% (246K) were not pre-selected to 'opt-out'."
████: "BTW - through the protocol to identify over-age and under-age we probably have experience in identifying if a user is a teen."
████: "We do - however, we do not trigger an Age Verification in such a case today given the potential revenue impact it can cause."
"It's important to remember how crucial teenagers are to Instagram. They are early adopters, tend to have a fair amount of free time, and actively explore their identity… That's why it's so important that we stabilize Instagram teen DAP in countries where the trends are soft"
"Our biggest risk right now is that there are a handful of important countries where, while teen engagement is strong, teen DAU and MAU are soft. That's why we've taken on teen goals across so much of Instagram this half, and that's why it's so important that we reverse those trends."
"Reflecting on ████ comment about the difference between a team that finds and fixes societally important user problems and what the Integrity teams do, I think it's important to recognize what our partner teams do and don't do. Relevance: Create the value model ranking terms, and then tune them maximally empirically to optimize for vpvs, sessions, and dap. This is often a blind mathematical process with no principles. As a result, this very frequently creates exacerbates integrity problems, but beyond some very loose guardrails, that is not their problem to solve. XI: is the enforcement arm of content policy. Its toolkit is constrained to classifier-targeted Remove, Reduce, and Inform on defined content problems per the community standards. They have no capacity to makes changes to the core algorithm, they are sort of the last line of defense, and given the narrow definitions of our policies, the ceilings on demotions (a '25% demotion' is a 0.25 multiplier on the relevance score, not a 25% reduction in vpvs... often like ~8%), and the state of classifiers in 2023, the reality is that they miss a lot. FI: are sort of a 'middle' org between the two — it is responsible for bad user experiences, which often extends beyond what is strictly defined under our policies. They have a much more expansive toolkit, including deboosting (removing relevance event weights) and core algorithmic changes. Because their focus is bad experiences, they are focused on things like p(NUF) demotions, or cooling down the civic ecosystem, rather than strictly violating OR societal outcomes. IGRI: are like IG's FI, but without the toolkit. They do no deboosting or algo changes yet. This leaves very large gaps when Relevance changes create negative externalities that aren't FI priorities. FI is clear that societal outcomes are not in scope for them, and their job is to respect user preferences towards non-violating content. They are constantly firefighting and have very limited capacity for exploratory work outside of their priorities – which are mapped to growth priorities. This is how they have done nothing about Teens on FB in the last decade, and why we need to."
Problematic Use: Definition & Impact
"Problem statement: Occurs when people feel they lack control over their usage of technology and this leads to negative life impact (e.g. on sleep, productivity, parenting)"
"Importance to key stakeholders:
• Experts: Problematic use is a top cited problem by expert advisories, including the American Psychological Association and Surgeon General, as it 'can impair [adolescents'] ability to engage in daily roles and routines, and may present risk for more serious psychological harms over time'.
• Parents: 58% of parents felt concerned that social media apps are 'too easy to use too much' (Marketing Insights, n=6,146).
• Teens:
◦ 37% of teens (13-17) felt concerned that social media apps are 'too easy to use too much' (Marketing Insights, n=3,085).
◦ 38% of teens (13-17) report low attentiveness around IG use (<=2 on 5-pt scale), and these users report 20-25% higher life interference from app use (i.e., perceiving that IG use prevents the fulfillment of other responsibilities).
◦ Higher sense of control is associated with app retention and perceived IG value."
"Teens reporting low attention (aspect of low control) report 25% higher life interference (Davis & ████, 2023)."
Child sexual exploitationFriend/Account Suggestions
"Following the first Wall Street Journal article, the Child Safety Task Force has been working to tighten our existing child safety protections. This included work to ensure that Risky Adults (i.e those with specific Malicious Child Safety Actor - MaCSA scores) and minors are not recommended to each other. Implementation of the final segment of this work (recommending teens to Risky Adults on FB's PYMK) has been blocked by product based on Voice concerns. We have made extensive external statements that this filtering is in place, and based on these and user safety concerns."
"Other segments of our existing adult-minor Recommendations Policy, which preclude recommendations of adult-minor pairs without mutual connections and with large age gaps, have also never been implemented due to technical feasibility concerns."
"**Internal Policy**
Our external Recommendations Policy does not address adult-minor recommendations; however, our internal version states that we do not recommend:
- *Accounts or entities that are suspected of engaging in inappropriate interactions with children, the sexualization of children, engaging in exploitative intimate imagery and sextortion, or solicitation of CSAM, nude imagery or sexualized imagery of real or fictional children, or offline sexual encounters with children.*
- *We will not recommend the accounts of minors under the age of **16 to adults over the age of 35,** and vice-versa.*
- *Implementation: not possible today, should be in consultation with PYMK surface team and Child Safety.*
- *We will not recommend minors to non-minors, and vice-versa, in cases where the accounts share **no mutual friends**.*
- *Implementation: not possible today, should be in consultation with PYMK surface team and Child Safety."*
"A 2022 internal review noted that these components of the policy were not implemented due to technical feasibility."
Child sexual exploitationFriend/Account Suggestions
Amit ████: "We've recently learned that there are gaps implementing some previously agreed / documented limitations to adult<>minor connection recommendations in PYMK on FB and AYSF on IG. As teams have been working to implement, they've bumped into concerns about voice expressed by Growth org. Given external pressure about this topic, we'd like to escalate and resolve ASAP. The XFN (including reps from the Child Safety Task Force) is preparing an escalation document, but it's not clear how to move this forward."
Gabriel ████: "1/We'd been escalating with Growth on PYMK filters for risky adult-minors as part of the taskforce
2/ Policy were surprised that this escalation was needed at all because on the policy books, we should already be not recommending adult<>teen connections via PYMK. There are some live policy escalations that depend on having a clear answer to (i) what is our policy? and (ii) is it being enforced?
policy are now moving towards a broader escalation to align recommendations implementation with the policy lines, but this will have some significant x-app tradeoffs on friending/connections
Restricting teen <> adult AYMF also has implications for how we would subsequently do content recommendations for teen <> adult. Following a prior escalation, on IG we limit teen <> adult content recommendations to only. We are now in the process of extending that to teen <> adult AYMF recommendations (for only).
3/ Policy are now moving towards a broader escalation to align recommendations implementation with the policy lines, but this will have some significant x-app tradeoffs on friending/connections"
Robert ████: "Outside Counsel, Public Affairs, Policy, and Legal have reviewed known records of external statements and found that we have made extensive statements that we have filtering in place preventing Risky Adults and Minors from being recommended to each other on both FB's PYMK and IG's AYMF. *However, leadership within Policy can recall making broader statements; we have not been able to identify specific records but we cannot rule out the possibility that our external statements have gone beyond Risky Adult protections."
Child sexual exploitationFriend/Account Suggestions
Monika Bickert: "I wanted to get your thoughts on the options in this doc about how to restrict recommendations of minors and adults to one another. My sense right now is that Product is wedded to Option 1 (status quo: cut out "risky adults" from recommendations to and from minors) and US Public Policy is wedded to Option 3 (significantly expand restrictions).
Candidly, I am hoping we can reach a middle ground and land on Option 2, which only allows recommendations between adults/minors if they are strongly connected (e.g. mutual friends or same contact info) or if the recommended account is a public figure with over 10k followers. Andy is not confident that we could get Product to move off of Option 1, but we feel comfortable pushing Option 2 with them. Antigone and team are ok with Option 2."
Kevin Martin: "Joel - Monika and I connected and I explained I had serious reservations that Option 2 is not strong enough and may put us in a vulnerable position publicly… Unfortunately I don't think option 2 provides nearly the protection that we need or that we should be providing to youth for numerous reasons."
"When considering the best policy option I think it is important to remember that we are not talking about limitations on interactions between youth and adults, but rather protections for with whom we should RECOMEND that youth connect. When we recommend an adult to a youth or vice versa, we are much more vulnerable – as we are then responsible for anything harmful that comes from the relationship. And it is much harder to defend mistakes as we are inserting ourselves into the relationship. While I understand product is whetted to option 1, I think this issue is very likely to blow up publicly and soon. As a result, I think we are better off escalating rather than agreeing to options that will result in public criticism later. If leadership decides to go with option 1, then they will do so understanding the risks and won't be surprised later. I have several specific concerns with option 2 that Monika outlines:"
"Basically our internal recommendations policy has 3 protections: (1) limits on risky adults; (2) an age limitation (no 35 to under 16 recommendations); and (3) a requirement that there be a connection through a mutual friend. While these internal protections on recommendations have never been enforced, option 2 is a significant reduction in the protections that are 'in place' and will be perceived as weakening them to the extent that the change in guidelines leaks. For example, in the chart I have attached, you can see that there are a significant number of recommendations for teens in the over 35 category that are occurring today as that limitation has not been enforced. So eliminating this protection entirely results in a significant change and will be perceived that way to the extent that this change is uncovered. With the public attention in this area and with ongoing litigation, I worry that all of the details will come out and it will become apparent that we never enforced the earlier guidelines and that we ultimately weakened the protections in ways that will be hard to explain and even understand (why do we need to recommend a 40 year old to a 15 year old). Again I am not suggesting we restrict such interactions, but only that we be more cautious in our recommendations."
"Second, Monika describes option 2 as only allowing recommendations if the recommended account is a public figure with over 10k followers, but I don't think that is complete. What she describes is actually a prong of our option 3. Option 2 allows public figures with 10k followers 'OR is a newly created (L90D) Professional Account with >1K followers*'. As I understand it the articulated concern by product is that they want to be able to recommend celebrities to teens to follow (e.g Taylor Swift to teens). The addition of the professional account with only 1k followers seems to go well beyond existing celebrities."
"Third, it is not clear that Option 2 alone will be able to withstand the inevitable public scrutiny. The limitation on 1 mutual connection is pretty broad and it is unclear how much protection that provides. Especially since we are still going back and forth on exactly what level of behavior will be deemed risky, a mutual contact requirement alone seems to offer too little protection. To be clear it is helpful but without more I worry it is too thin a reed to bear that much weight. It is too easy to meet. For example, every child who is connected to their team's coach would then be eligible to be recommended to every adult that the coach is connected to. Every work colleague. Every high school classmate. I worry that puts too much pressure both substantively and publicly on our MSCA test. The MSCA protection is too amorphous and in flux to be able to determine whether it will offer a sufficient protection for us to defend recommending children to adults and vice versa. The experience with both Francis Haugen and the earlier WSJ story should remind us that it is not what protections we put in place that protect us but rather what protections we can explain and seem reasonable publicly that will be important. I worry that the first time our MSCA system misses risky behavior and we recommend a child to a 45 year old because they went to high school with the teen's track coach, we will be very vulnerable."
Joel Kaplan: "I share the queasiness about the 'mutual connection' being protective enough, but a two-year requirement seems overly restrictive of potentially reasonable use cases (████ and ████ follow each other on IG (I think), for instance. Don't know if that was suggested, but would certainly seem like a good use case). So maybe finding something that is more of a 2.5. I'd also like to understand what work we are doing to dial up recall on the 'risky adults' classifier. Seems to me this might be the most powerful protection of all."
Monika Bickert: "We aren't necessarily arguing for less restrictiveness. We're simply trying to find an option that is realistic for both USPP and Product (they are quite far apart) AND that the Safety team can bless. I think we may just have to kick it up to you, Nick, etc. Product can articulate their concerns with each of the additions beyond Option 1. For instance, Product will point out that we are struggling to keep up with how young people use social media these days and that is really hurting us — one big thing is that young people like discovering and following new creators. They aren't interested in following the big name celebrities so much as they want to find the ones that they think have a niche that is cool. That's why we added a lower follower threshold for 'new creator' accounts in addition to the 10k threshold. They'll have similar points on the other requirements, which I'm sure Andy and Tiggy can also articulate."
████: "I feel like i'm reading totally different litigation and legislation than our org leaders… We literally are optimizing teen feeds for engagement. We can change that, but none of the YIPs actually address this core issue."
████: "[D]o we do this differently/more for teens than adults?"
████: "[N]o, exactly the same"
████: "teens seem to just be more responsive to some of the engagement-optimization stuff we do for *all* users… which seems consistent with the external concerns"
████: "I think we need to stop optimizing teen users' algos for engagement metrics to be in compliance but nobody seems to be talking beyond building new tools and strong enforcements"
████: "yeah. prob because there's a whole other side of the company working equally frenziedly on how to achieve teen growth"
████: "Are the compliance people reading the same litigation and legislation that I am?"
████: "I am asking the same question"
████: does ████ realize how the core algo is set up? Like it's literally just mathematically optimizing for vpvs and sessions *for teens* as with everyone else, which is the thing we're being told we cannot do…"
████: "I dunno why we're doing cartwheels to avoid doing the thing we're actually being asked to do *which we can easily do at very little topline cost!*"
████: "does ████ know that we literally optimize Teen feeds for engagement?"
████: "err does he know that IG ranking is already different for teenagers… and FB ranking is different for YA… just in both cases it's to squeeze even more sessions out of them"
████: "it seems to genuinely be the case that optimizing for sessions just naturally distributes more harm to teenagers… it's been too consistent over time."
"[S]o, relevant to the lawsuits, and some foreshadowing from Mark in Q&A just now –– basically the company position is that time spent and engagement metrics proxy user value and we are 'not' just tuning the algorithm to keep people scrolling as much and as long as possible … except mechanically we absolutely do through a process called the 'guided value model' (GVM) process. GVM halves and doubles the weights on every major relevance term, which is then fed into an algorithm that is told to optimize for (a) sessions; and (b) vpvs. It spits out a bunch of changes to the weights of different terms and then we ship it. So really, the algorithm is absolutely tuned to maximize engagement in an maximally empirical, principle-less way. I am not sure that the youth legal people realize this, and that it applies to teens just as much as gen pop… so the things contained in the lawsuits are somewhat hard to refute so long as Teens are part of the GVM process."
Body image & eating disordersSuicide & self-injuryFriend/Account Suggestions
"We know that SSI and ED have a significantly disproportionate large teen audience. SSI and ED SWOP among teens is ████ higher than among adults on IG (query). When examining consumption patterns among high consumers (>= 8 pieces of violating content) of SSI and ED violating content, teens are approximately ████ more likely to be high consumers than adults."
"What are some potentially problematic SSI-ED related patterns observed on IG?"
"'Gated' Communities… Some users are involved in SSI-ED closed "communities" on IG; content production is targeted at people already in a
user's follower network. These users may feature SSI-ED related terms in profile, possibly to find and signal a non-judgmental space where SSI-ED behavior is acceptable."
"'Vent' Accounts… Accounts dedicated to discussing SSI-ED issues can serve as secretive, secondary "vent accounts" that may provide a nonjudgmental space for people to share their experiences and feelings. These accounts may also normalize problematic SSI-ED
related behavior, especially for younger users."
"Account & Content Recommendations…
After connecting with accounts that include or create SSI-ED related content (e.g., "vent" accounts), recommendations for suggested accounts related to SSI-ED may be shown to users across surfaces. Persistent recommendations for SSI-ED adjacent topics, such as body image, weight loss, fitness
and sad memes may be shown on unconnected surfaces."
Child sexual exploitationFriend/Account Suggestions
Andy ████: "Flagging this new video Horwitz just sent of the AYSF recommendations for one of the Instagram accounts he's used to do this child safety research. Can we please look into this? I'd have assumed at this point the steps we've taken would have accounted for at least most of these recommendations."
Alison ████: "Oof - thanks for flagging Andy.
1. We need to nuke these accounts ASAP. @Carolina ████ - what is the best way to have ops not only remove the accounts that are being recommended, but also can we do a fanout of the accounts that are following the accounts being recommended here.
2. MaSCA scores. Can we figure out what Jeff's account's MaSCA score is and why it is being recommended not only seemingly minor content, but what also looks like accounts that should have a MaSCA score themselves and be non-rec. @Rishi ████, is this something your team can prioritize today? It seems like our thresholds for AYSF may not be in the right place?"
Kristina ████: "i3 Child Safety identified several accounts belong to Jeff. Last we checked, they didn't have MaCSA scores."
Alison ████: "@Gabriel ████ – can we take a look here while ops is pulling the accounts from the video for why his accounts don't have a MaSCA score given how often he is searching for/viewing this type of content."
Malia ████: "There are no MaCSA signals for views"
Carolina ████: "Also, just flagging that I'm using my test user account to surface the accounts, and also started being recommended some myself."
Swithin ████: "Shouldn't these accounts also be auto enqueued for review based on account names esp pizzalink, kidseller?"
Melissa ████: "Also looks like a handful do have reactive user reports on the profile, but are sitting in the queues/pending"
Carolina ████: [B]ased on initial fan-outs, I'm seeing a lot of keyword obfuscation (i.e. "megha" or "megga" instead of "mega", and instead of "seller", using "sellar" or "zshaller") But also more obvious versions like, "pizza mega sellers" or "trusted cp seller"."
Gabriel ████: "AYMF and Content filtering are still not where we would want them to be. There are multiple levers we could pull but all would take some focus and both eng+policy"
Child sexual exploitationFriend/Account Suggestions
Andy Stone: "Hi folks - following up here with a smaller group just to make you aware of something related to Jeff's ongoing reporting of child safety issues on the platform. I don't want to come across as alarmist - and myself have pushed back vociferously on Jeff that he's created an artificial experience and this is niche content geared to a small audience - but want you to know what we're up against given the subject matter.
The following is a video from just this morning of the Accounts You Should Follow recommendations Jeff got in one of the Instagram accounts he has used to do this child safety investigation.
We have touted to Jeff our efforts to enhance and expand the use of MACSA scores across platforms and surfaces as one of the meaningful steps we're taking to address this issue. But if neither Jeff's burner IG account nor any of those that were recommended to him, as reflected in this video, has a MACSA score high enough to prevent recommendation, it's an eminently harder sell."
Sayed Otaru: "This looks pretty bad."
"Objective
- Unlock the ability for high school and college students to discover and connect with people from school (Phase 01) and provide a safe space to be authentic and build community with schoolmates (Phase 02), to give teens a new reason to join/find more value with IG."
"Risks
- PR/public perception"
"In 2024, we need to unlock these new school communities that will kickstart network connections in two phases"
"we will need to invest in virality within school communities to get teens and their friends on IG together."
Praveen ████: "shared: Instagram_Instagram_s_Algorithm_Delivers_Toxic_Video_Mix_to_Adults_Who_Follow_Children_-_WSJ.pdf"
Tomas Kabbabe: "It is true that the article seems like an intentional hit piece… But at the same time, some of the content shouldn't be in the platform"
Praveen ████: " definitely agree - my guess is that it would look much much worse if we weren't doing enforcement. But it is concerning that our systems almost classify this type of contest as an interest category, and will serve a ton of it if we think a user likes it. can lead to pretty bad rabbit holes… i suppose it just reflects the reality that it is an interest category lol"
Tomas Kabbabe: "Honestly the hardest part of my job sometimes is realizing that humans can be very gross"
"Entire Feed Ranking has a great half. Connected Ranking side, we exceeded our global cap-15 session goal and VPV goal. Blending/IFR joint-holdout also exceeds both global and teen cap-15 session goals. This is nearly mission impossible, when we set up goals in early H2. I still"
"We launched Connected Feed MB3 model this week, which improved prediction task accuracy and topline metrics, and officially completed our model can't fail milestones in H2. Compared to the beginning of H2, this model update increased our training data by 10X for all production models (100% teens and 60% for genpop), model size by 200X, and incorporated a more comprehensive set of features including cross-surface, teen-specific, and embedding features. Overall, this update improved offline prediction accuracy (~5% penter_comment_surface, ~2% pexternal_share, ~1.2% p_comment, ~0.75% preshare) and drove ecosystem value (+0.33% genpop direct reshares, +0.21% genpop likes, +0.22% teen direct reshares, +0.08% high ME imps). We expect the model to mature in the next 6 weeks and continue to see growing topline impact in the backtest."
"Increasing connected feed model training data has shown +0.06% cap15 sessions for both teen and gen pop, validating our hypothesis that increasing model training data drives value for teens and marginal users. At the beginning of H1, we hypothesized that low training data volume limits our ability to deliver model value for Teens. In H2, we took a big bet to significantly increase training data sampling rate on connected feed for both teens (5% -> 100%) and gen pop (5% -> 60%). We have now launched increased training data for 4 out of 5 production models, and are on track to complete the full migration by EOH."
"[Notifs / Stories / Feed] (10/30). This is a collaborative launch with IG Growth adds the signal p(time_spent & click), top feed / stories engagement features, and more training data to the UAR (Unified Actor Ranker) value model for push notifications, resulting in +0.067% absolute projected Global DAU, and SS wins to DM Teen MAU, cap15 sessions, and notif CTR."
Sources include internal documents, public filings and press reports. “Alleged by plaintiffs” denotes the plaintiffs’ characterization, not Meta’s own words. Source labels identify provenance, not a finding of fact.
Meta & Children's Online Safety