[] IG Search Integrity: Keyword Enforcement Proposal for SCC Less

Raw text

Transcribed text as recorded, including headers and formatting fragments.

IG TRUST AND SAFETY OFFICE HOURS - 5/9 - A/C PRIV DISCOVERY SAFETY - SEARCH [A/C PRIV] IG Search Integrity: Keyword Enforcement Proposal for SCC Less PLTF EX-04323_0001 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050609 METANMAG-130-00050609 Goals for today Context - A key goal in H1 2024 is to improve Safety for Teens. The top opportunity to do this is to improve enforcements on non-recommendable queries such as "porn", "meth", "self harm", and "skinny body", a gap that many competitors have already addressed. - Currently, IG shows account results when someone searches for a non-rec query. Some of these accounts have non-rec content, violating content, or sometimes neither, but these accounts can still cause perception impact for the users (teens, parents, adults) - especially when the user has opted to be in a Safe setting (SCC Less). We are seeking a decision from leads to improve search safety for non-rec queries for SCC Less (Designs, Traffic light): - Option 1 (Increased Friction - Policy Preference) that adds SERP friction for non-rec queries to mitigate some screenshot risk without significant over-enforcement impact, but is unlikely to improve topline teen safety. - Option 2 (Exact Match/Connected Only - Product/Legal Preference, Policy Supported) that only shows the exact match account and connected account results for a non-rec query. This option prioritizes safety over voice and has the potential to reduce overall teen SWVP (Severity Weighted Views Prevalence) by 2.1% and overall teen sexual non-rec prevalence by 21%. - Option 3 (Fully Blocking) completely empties SERP for non-rec queries, which is Tiktok's approach Non-goals for today - Discuss overarching Search principles: Search is an incredibly complex surface for enforcement principles and 30 min is insufficient to cover all enforcement principles, hence we would like to address this one-piece at a time. - Discussing other improvements: a) protections for inherently violating queries b) Typeahead safety (the team is actively working on closing gaps) c) improvements to the overall SCC product PLTF EX-04323_0002 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050610 METANMAG-130-00050610 Context Proposal Discussion PLTF EX-04323_0003 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050611 METANMAG-130-00050611 Terminology - SCC - Sensitive Content Control - SCC is a setting that allows users to control how much sensitive content they see on IG - Currently, ~54% of teens are in SCC Less, but this will increase to ~95% of teens through Project Nido. U16 teens will be hard defaulted into SCC Less and 16-17 y/o teens will be soft defaulted into SCC Less. - Inherently non-rec query - Based on IG's Keyword Recommendations Guidelines - Examples - "porn", "meth", "self harm", "skinny body", "sex", "horny" - What is currently not covered - "pause challenge", "new drake song" and other queries that may currently show a high amount of non-rec content on SERP, but are not inherently non-rec per our Keyword Recommendations Guidelines. Policy is being updated to use SERP Media rollup to inform whether a query is considered non-rec. - Note: Policy is updating the Search Principles and Enforcement Framework. Some non-rec categories (e.g. Life Events) will be excluded from the proposed enforcements. - Inherently violating query - Based on IG's Community Guidelines. Inherently violating queries are already completely blocked (we do not return any SERP) on IG search through the blocking inform module keyword lists. - Examples - "child porn links", "how to kill myself", "pro anorexia" - Search surfaces - Typeahead - Recommended query matches that appear as the user types a search query - SERP - Search Engine Results Page - Media SERP - Content results on search that appear on the For You tab - Account SERP - Account results on search that appear on the Accounts tab PLTF EX-04323_0004 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050612 METANMAG-130-00050612 Context - Problem Problem Teens use Search to find problematic content, as measured by the fact that Search is a top pathway contributing to teen SWVP and non-rec prevalence. However, parents expect IG to limit the discoverability of age-inappropriate content for teens on search. Opportunity Prioritizing safety over voice for Option 2 (Exact Match/Connected Only) can reduce overall teen SWVP by an estimated 2.1% (-4% UWVP) and reduce overall teen sexual non-rec prevalence by ~21% (Source) while also building into IG's broader efforts to address parental concerns around age-appropriate content. Overall: - Profile contribution to SSPx (67%) * search contribution to Profile (48%) * SSPx contribution to SWVP (6%) * Enforcable Rate (77%) = 1.5% - Profile contribution to ANSA (54%) * Search contribution to Profile (48%) * ANSA contribution to SWVP (3%) * Enforcable Rate (77%) = 0.60% - Profile contribution to sex non-rec (57%) * search contribution to profile (48%) * Enforcable Rate (77%) = 21.1% PLTF EX-04323_0005 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050613 METANMAG-130-00050613 What we know - Search is a key pathway to violating VPVs: at least 48% of SSPX VPVs and 26% of ANSA VPVs. - Out of the top 200 searched users that led to violating ANSA/SSPX VPVs, 77% are results from inherently non-rec queries (others came from "adult" or "lesbian") - Classifier upgrades and other enforcement improvements have limited opportunity to improve content safety. - Inherently non-rec queries (e.g. "porn") mostly produce bad or worse results. In many cases, users are intending to discover violating content and entities. Current IG inventory cannot support safe search since there are very few good results, and ranking/filtering are challenging and not effective. - We currently filter non-rec accounts from account results, but since non-rec account recall is currently ~38%, we are not enforcing >60% of non-rec accounts. There are also many accounts that look non-rec but barely scrape through policy lines. - Since recall for content/entity classifiers will never be perfect, users can continue relying on search to aggregate all False Negatives in one place. See violating and non-rec samples in Appendix. - Non-rec prevalence can be driven down to MDE (Minimum Detectable Effect) on recs, but this will not be possible on Search without updating search principles - IG needs to cater to the user's explicit non-rec intent on Search and return relevant results. Since the starting query is inherently non-rec, a concentration of unsafe exposure cannot be avoided. PLTF EX-04323_0006 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050614 METANMAG-130-00050614 Context Proposal Discussion PLTF EX-04323_0007 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050615 METANMAG-130-00050615 Today's Search Experience Problem: Despite SCC Less being our safest setting for teens and IG completely blocking inherently violating queries on search, teens still discover violating and non-rec content/entities through inherently non-rec queries. Results for SCC Less are not significantly safer than SCC Standard results [Screenshots: side-by-side comparison of search results for query "horny" under "Less" (SCC Less) and "Standard" settings, both showing similar lists of sexually-suggestive account results] Note: There is stricter downranking for SCC Less, but it's not helpful for risky queries since we're ranking between poor and worse results PLTF EX-04323_0008 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050616 METANMAG-130-00050616 Proposal TL;DR - Enforcement Options [Screenshots: three iPhone mockups showing search results for "horny" under three proposed enforcement options] Option #1: Increased Friction Option #2: Exact Match/Connected Only Option #3: Fully Blocking PLTF EX-04323_0009 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050617 METANMAG-130-00050617 A lightweight option that still allows discovery, but adds additional friction We will hide all search results behind one layer of friction. At a minimum, we propose shipping this experience to bring IG to parity with FB. We will be ready to public test this experience in the next 2 weeks. [Screenshots: before/after search results for "horny", showing an interstitial "We've hidden most results" screen with a "See results anyway" option] PLTF EX-04323_0010 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050618 METANMAG-130-00050618 Only allow exact account matches (for unconnected accounts) and connected accounts only Making these 2 allowances will help us preserve voice and over enforcement against benign accounts that happen to have an inherently non-rec keyword in their username and prior connections. [Screenshots: before/after search results for "horny", showing results hidden except for exact-match/connected accounts] PLTF EX-04323_0011 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050619 METANMAG-130-00050619 Fully blocking search, similar to TikTok's approach Terms that Tiktok has completely blocked from search (for all users, not just teens) include "porn", "sex", "horny", "adderall", "fentanyl", "anorexia", "self harm", "suicide". IG currently allows results for all of these terms for teens. [Screenshots: before/after search results for "horny", "after" showing a TikTok-style "No results found" screen] PLTF EX-04323_0012 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050620 METANMAG-130-00050620 Options for keyword enforcement (Traffic Light) Option #2 (Exact Match/Connected Only - Product/Legal Recommended) prioritizes safety over voice. We believe that building trust with parents and the estimated 2.1% teen SWVP and 21% teen sexual non-rec prevalence reductions are worth the overenforcement trade-off to protect teens. [Table comparing Status Quo, Option #1 Increased Friction, Option #2 Exact Match/Connected Only and Option #3 Fully Blocking, across "Safety" (discoverability of non-rec/violating content and entities) and "Voice" (creator reach, discoverability of benign and popular accounts) columns. Status Quo: non-rec accounts filtered but no query-level enforcement, recall gaps make it extremely easy for teens to find non-rec and violating content; all recommendable accounts eligible to show up in search. Option 1: teens will easily click through and find non-rec content and entities, no anticipated SWVP or non-rec prevalence reductions; moderately reduced discoverability for some recommendable accounts. Option 2: estimated 2.1% teen SWVP reduction and 21% teen sexual non-rec prevalence reduction; difficult for teens to find recommendable accounts associated with inherently non-rec keywords. Option 3: estimated 2.1% teen SWVP reduction and 21% teen sexual non-rec prevalence reduction; fully impossible for a teen to find any accounts associated with inherently non-rec keywords.] PLTF EX-04323_0013 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050621 METANMAG-130-00050621 Risks We are likely to make voice tradeoffs (see below) with Option 2 that can be managed via a) Comms b) Mitigations (next slide) Risk 1: Over-enforcement of benign accounts for inherently non-rec keywords Examples: @fuckjerry (popular meme account), @foodporn, @thesuicidesquad. We anticipated moderate to severe creator backlash. Risk 2: Hard for users to find benign accounts with long names due to misspellings etc Example: Misspellings or typos for benign and popular accounts such as @thehornyrammyc or @thesexlifeofcollegegirls may cause users to have difficulty finding the intended account. Risk 3: False positive queries flagged by our classifiers Example: "eating disorder awareness week", "food porn", "toxic relationship", and other related non-rec keywords Overall Risk: Although this experience does not directly link to SCC settings, Teens can still attempt to change their SCC settings if they find that Search isn't comprehensive for them. PLTF EX-04323_0014 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050622 METANMAG-130-00050622 Mitigations Although there will be significant tradeoffs for Option 2 (Exact Match/Connected Only), we have mitigations that will help. Mitigation for Risk 1 and 2 (Benign accounts and typos) a) Content instructing users to type the full username b) Showing verified accounts, managed partners, and accounts with higher followers (currently not policy approved due to bias concerns) [Note: this mitigation won't capture all benign/popular accounts. E.g. @food_porn (300k followers) may not be shown] Mitigation for Risk 3 (False Positive queries) a) Phase 1: Manually curated keyword bank b) Phase 2: Slowly open to keyword classifiers with proactive Ops review and allowlisting to mitigate most False Positives [Allowlist "safe" accounts] PLTF EX-04323_0015 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050623 METANMAG-130-00050623 Context Proposal Discussion PLTF EX-04323_0016 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050624 METANMAG-130-00050624 Discussion 1. Are you aligned with moving forward with Option 2 (Exact Match/Connected Only)? - If we align on Option 2 (Exact Match/Connected Only), the team will start testing at the end of H1 with a set of manually curated keywords to enable a safe launch, and then scale to use a classifier in H2 (Test Details). We will also test Option 1 (Increased Friction) as an enforcement for SCC standard, in parity with FB. - If we align on Option 1 (Increased Friction), the team will start testing in late May (Test Details). Since we do not expect safety impact, the team will brainstorm other enforcements to reduce teen search pathways to problematic content. If we're aligned on Option 2, 1. Are you comfortable with the mitigations we have proposed for the over enforcement and voice risks? 2. Are you aligned with the test setup and going forward, what precision tolerance should the team aim towards for non-rec query classifier enforcement? - We were aiming for north star p75, but starting at p50. Note: Our current keyword classifier has ~30% precision at 40% recall (text classifiers have lower performance in general), hence we will start with a manually curated list of keywords with high precision. 4. Are you aligned on the Option 2 experience being limited to SCC Less only? - Unsafe results for non-rec queries will remain a gap for gen-pop users (including parents), but we do not believe the over-enforcement impact is worth it to extend either enforcement to gen-pop. 5. Do we need to plan for a Youth Integrity review for this work? PLTF EX-04323_0017 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050625 METANMAG-130-00050625 Appendix PLTF EX-04323_0018 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050626 METANMAG-130-00050626 Today's Search Experience The benign search experience consists of typeahead suggestions and media SERP This includes queries like "home", "Taylor swift", "latte art", etc. [Screenshots: typeahead suggestions and media SERP grid for the query "home"] PLTF EX-04323_0019 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050627 METANMAG-130-00050627 Today's Search Experience For violating queries, we fully block typeahead and SERP This includes terms such as "how to kill myself", "child porn", "pro anorexia", etc. [Screenshots: search for "how to kill myself" showing no typeahead suggestions, and a "Help is available" screen with options to contact a helpline, see suggestions from professionals outside of Meta, and reach out to a friend] PLTF EX-04323_0020 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050628 METANMAG-130-00050628 Today's Search Experience For inherently non-rec queries, we block typeahead suggestions but still show account SERP This includes terms like "porn", "sex", "cocaine", "skinny body". For non-rec search queries, we currently: 1. Filter typeahead suggestions against our Search recommendation guidelines. 2. Show account SERP instead of media SERP to help avoid directly showing potentially problematic images and videos. However, the accounts we show still offer a gateway into accounts that post violating content. [Screenshots: search for "horny" showing no typeahead suggestions, but account SERP results still displayed] PLTF EX-04323_0021 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050629 METANMAG-130-00050629 Today's Search Experience As a reminder, our search enforcements for inherently non-rec queries are impacted by a user's SCC setting These enforcements include downranking and filtering entities on SERP. 54% of teens are in SCC Less today. With our teen accounts work, we will be defaulting all teens into SCC Less (either hard or soft default depending on their age). We offer transparency that SCC also impacts search results. [Screenshot: Sensitive Content Control settings screen showing "Less" selected] PLTF EX-04323_0022 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050630 METANMAG-130-00050630 It is quite easy to find violating and non-rec content by searching non-rec queries or violating adjacent search queries [Screenshots/redacted content boxes labeled SSI/ED and ANSA, showing sample search results marked SCA/PRIV] Table: User Search by query length query length 3-6: search_query_text / queries / sessions / user -- sex 1116002 / 345289 / 3161; xxx 614000 / 217320 / 1993; porn 537140 / 207926 / 1951; sexy 501750 / 138300 / 1278; porno 415431 / 116860 / 1061; xnxx 248002 / 76792 / 7123 query length above 6: search_query_text -- hentai; pornhub; breastfeeding; frases; sex video Many top queries on search are inherently non-rec (sexual) PLTF EX-04323_0023 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050631 METANMAG-130-00050631 TW: Non-rec and Violating Search Samples (ED) Account-level recall is low, and it is extremely easy for any teen to find pro-eating disorder accounts with very triggering content. [Screenshots of sample eating-disorder-related account profiles and search results found via non-rec search queries; account content redacted/marked SCA in source] PLTF EX-04323_0024 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050632 METANMAG-130-00050632 TW: Violating Search Samples (SSI) [Screenshots of sample self-harm-related account profiles and search results found via non-rec search queries; account content redacted/marked SCA in source] PLTF EX-04323_0025 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050633 METANMAG-130-00050633 Violating ED Query Experience The ideal experience is to block all triggering eating disorder content and entities. Not all eating disorder related queries are inherently violating and most cannot be blocked from search. However, account and content-level recall gaps pose a huge risk to teen safety. [Screenshots: "Help is available" screens shown for queries "pro anorexia" and "thinspo", each offering to contact a helpline, reach out to a friend, or see suggestions from professionals outside of Meta] PLTF EX-04323_0026 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050634 METANMAG-130-00050634 Users have high intent to find sexual content The top queries on search are all inherently sexual and non-rec. Table: User Search query length 3-6: search_query_text / queries / sessions / user -- sex 1116002 / 345289 / 3161; xxx 614000 / 217320 / 1993; porn 537140 / 207926 / 1951; sexy 501750 / 138300 / 1278; porno 415431 / 116860 / 1061; xnxx 248002 / 76792 / 7127 query length above 6: search_query_text -- hentai; pornhub; breastfeeding; frases; sex video PLTF EX-04323_0027 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050635 METANMAG-130-00050635 TW: Non-rec and Violating Search Samples (Sexual) [Screenshots of sample sexually-explicit account profiles found via non-rec search queries; account content redacted/marked SCA in source] PLTF EX-04323_0028 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050636 METANMAG-130-00050636 TW: Non-rec and Violating Hashtag Samples (Sexual) [Screenshots of sample sexually-explicit hashtag search results (e.g. searches for obfuscated/adjacent hashtags); content redacted/marked SCA in source] PLTF EX-04323_0029 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050637 METANMAG-130-00050637 Public Test Details Option 1 (Increased Friction) - ETA: mid to end of May - Keyword list: Matching the FB implementation to start (high precision manually curated regex), then will move to v2 of the IG Offensive Keyword Classifier when it is upgraded at the end of H1 24 with proactive manual review and FB offensive keyword classifiers. - Expected metric movement: No anticipated topline metric movement; we will track click-through rate for the interstitial Option 2 (Exact Match/Connected Only) - ETA: End of H1/in line with Project Nido dates (working with Youth teams to finalize) - Keyword list: Matching the FB implementation to start (high precision manually curated regex), then will move to v2 of the IG Offensive Keyword Classifier when it is upgraded at the end of H1 24, with proactive manual review. - Expected metric movement: Topline teen sexual non-rec prevalence: -21%; Topline teen SWVP: -2.1%. Tier 0/1 Growth metrics: Possible teen friending regressions PLTF EX-04323_0030 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050638 METANMAG-130-00050638 Proposal FB Parity FB shows an interstitial that requires users to explicitly select "Continue" to continue to search results for risky and inherently non-rec queries such as "sex", "porn", and "horny". [Screenshot: Facebook search interstitial reading "Are you sure you want to continue? The term you searched for is sometimes associated with nudity or sexual activity, which isn't allowed on Facebook." with "Continue" and "Go to News Feed" options] PLTF EX-04323_0031 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050639 METANMAG-130-00050639 Mocks TikTok fully blocks search (no results at all) for risky and non-rec queries Terms that Tiktok has completely blocked from search (for all users, not just teens) include "porn", "sex", "horny", "adderall", "fentanyl", "self harm", "suicide". [Screenshots: TikTok search results for "Ecstasy" and "Porn" both showing "No results found. This phrase may be associated with behavior or content that violates our guidelines. Promoting a safe and positive experience is TikTok's top priority. For more information, we invite you to review our Community Guidelines."] PLTF EX-04323_0032 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050640 METANMAG-130-00050640 Search x SCC: High Level Overview Table with columns: Search query | SCC state | Typeahead/Popular filtering | SERP landing experience | SERP filtering | SERP downranking Risky terms ("porn"): - Less: Same as Standard | Account SERP | Same as Standard but applies when query <= 10 char, e.g. horny, sex xxx | Same as Standard - Standard: Filter (unless exact match or followed) MAXA enforcements, impersonation accounts | Account SERP | Filter in either of following cases, except if exact match, followed, query > 4 chars. Does NOT include horny, drugs, etc | Downrank for AQVM and MAXA enforcements - More: Same as Standard | | Same as Standard | Same as Standard Violating terms ("how to kill myself"): - Less / Standard / More: Fully block all results (content, accounts) -- exact match typeahead and search results Caveat this is extremely high level PLTF EX-04323_0033 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050641 METANMAG-130-00050641 Discovery Safety Parents recognize Instagram as one of the safest and most age-appropriate apps for their teens to discover, create, and connect. WS1 - Parent Attitudes Measurement Measure how Instagram is leading with parents with standardized parent and teen sentiment tracking. WS2 - Winning Over Parents Bundle advocacy, parental controls, and default safety protections into a simple GTM package, centered around age bands. WS3 - Unwanted Contact with Strangers Reduce teen exposure to unwanted contact with adult strangers. WS4 - Age-Appropriate Content Reduce teen exposure to inappropriate content with a simple, explainable content framework. WS5 - Bullying & Harassment Demonstrate that bullying is not tolerated on Instagram. WS6 - SSI/ED Reduce potential repeat SSI-ED teen exposures and empower teens in recovery. WS7 - Parental Controls Address key parental concerns with Supervision features. WS8 - CSE & Child Predators Mitigate threats with investigation-led enforcement and surface-level measures. PLTF EX-04323_0034 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) META3047MDL-278-00050642 METANMAG-130-00050642