"Bad Experiences" Measurement: Plan for a 2021 plan

Raw text

Transcribed text as recorded, including headers and formatting fragments.

HIGHLY CONFIDENTIAL (COMPETITOR) Document Provided in Native Format depobook.com EXHIBIT 19 ΜΕΤΑ3047MDL-019-00094148 PLTF EX-02801_0001 PLTF EX PLTF EX-02801 P1178.1 WELL-BEING LEADS REVIEW NOV 19, 2020 "Bad Experiences” Measurement Plan for a 2021 plan P1178.2 PLTF EX-02801_0002 Meeting Objectives P1178.3 Goals 1. Share ambition and key examples of problems to solve 2. Share high-level approach 3. Feedback on next steps Non-Goals 1. Follow up on Oct 15 review of TRIPS proxy success criteria 2. Roadmapping-level detail 3. Product strategy for bad experiences 4. "Positive Experiences" measurement (ex. promoting positive well-being) PLTF EX-02801_0003 CONTEXT WELL-BEING MISSION AND VISION Our mission is to create the safest and most supportive global community. We envision an Instagram where, everyone, especially teens, creators and underserved communities, feel safe and supported to express themselves and to push culture forward. P1178.4 PLTF EX-02801_0004 CONTEXT DEFINITIONS P1178.5 Supportive - Helping people with bad experiences, by reducing and resolving them. Not generally focused on enforcement. Policy Violating - Problems defined in the company's Community Standards content policies Borderline - Non-violating content policy intended to address part of policy narrowness Bad Experiences - Problems defined by people's perception, as implemented in survey taxonomies like TRIPS and Hard Life Moments Legitimacy - People and external stakeholders believe that our integrity work is effective at reducing harm and that our enforcement is defensible and fair. Both 1st and 3rd party perceptions Reach - Measures the size of bad experiences (number of people affected by them or proxy) Support Effectiveness - Measures how effectively we're resolving a bad experience. For people experiencing the problem (1st party), intersecting with 1st party part of legitimacy 4 PLTF EX-02801 0005 Slide 1 1 @ @fb.com - If you have a "Safe" definition to add next to "Supportive" it would help ground this dimension as well for "reach", since this framework can cover bad experiences for both. Feel free to add directly or as a comment here. _Assigned to Sam @fb.com_ 11/19/2020 3:11:13 AM P1178.6 PLTF EX-02801_0006 CONTEXT WHY BAD EXPERIENCES Our company integrity mission is 'protect the community and its voice'. Our policies have to balance both protecting the community and its voice. People feel they're having a bad experience or they don't. Their perception isn't constrained by policy balance. Understanding "bad experiences" enables us to see the fuller people-first picture of how people experience safety and support concerns, agnostic of policy. P1178.7 5 PLTF EX-02801_0007 CONTEXT WHY BAD EXPERIENCES 100% -- of Community Standards policy focuses on enforcement (delete. MAD, or resources) ~100x -- Incremental reach for B&H witness in TRIPS compared with Community Standards B&H policy violating impressions P1178.8 PLTF EX-02801_0008 Slide 1 1 2 1 P1178.9 Is this true across problems or predominantly on bullying and hate speech where context matters a lot? Miki Rothschild; 11/19/2020 5:42:46 PM I think that it's generally true if we compare content prevalence impressions to TRIPS DAP. Our policies are narrow (CS violations are very narrow, given consequences of overenforcement when we get it wrong, borderline somewhat less narrow). Perception is broad. People who take the effort to report content or otherwise self- remediate are in the middle. Sam 11/19/2020 6:23:48 AM I agree with the spirit of this fact - the perceived problem is likely much bigger than the CS-violating problem - but I think it may be a bit misleading to compare % of impressions that are bad vs. % of people who have a single bad experience (in a week's worth of impressions) Shilpa Mody; 11/19/2020 5:42:46 PM PLTF EX-02801_0009 CONTEXT DEFINING BAD EXPERIENCES P1178.10 Types of gaps between policy and bad experiences Policy can be effective, but ... 1. incident doesn't meet the policy bar though clearly bad 2. incident doesn't fit current policy definitions 3. incident is hard to detect (by human or automated review) Policy isn't the right tool when ... 4. "bad" is impossible to determine with any consistency 5. enforcement actions (hard or soft) cannot address the problem PLTF EX-02801_0010 CONTEXT EXAMPLES OF BAD EXPERIENCES P1178.11 Examples of bad experiences often in policy gaps Mass harassment Very intense, low reach. Disproportionately affects creators. Portion isn't CS policy violating but still felt as intense. Non-credible or non-violent threats Doesn't meet CS violating policy bar, but felt intensely with moderate reach. Thinspiration Easy to find content adjacent to eating disorder part of SSI and to broader social comparison and body image issue-triggers. PLTF EX-02801_0011 CONTEXT EXAMPLES: MASS HARASSMENT P1178.12 Billie Eilish Admits That Reading Instagram Comments Was "Ruining" Her Life "The cooler the things you get to do are, the more people hate you." chrissyteigen Following Aahriarpeigen Chrissy Teigen Rita Wilson • Wilfred Zaha. Dua Lipa. Milana Vayntrub CATE Θ 1810 Yuvea Site The actress who plays AT&T's Lily is facing waves of online sexual harassment, including manipulated images and objectifying memes drsavery, They ex. pige R www ♡07 by genders ryanscaggs24 ryanscaggs24 (..) 35614es Arply preston_is_ok hey i think you are very best Lily Reply robgtovicyforyou Your so gorgeous I love when you come on AT&T's commercial Seikes Rep • Only some is policy violating or borderline • ♡ The frequency and volume makes a very intense bad experience • Low reach PLTF EX-02801_0012 CONTEXT EXAMPLES: THREATS When I received rape threats in my DMs, Instagram offered me no help at all Comment FIRST PERSON Friday 24 Jul 2020 3:52 pm I've been on the receiving end of an unsolicited dick pic or two in my time. I laughed them off at first. Then the same man started sending explicit descriptions of violent sexual fantasies he had about me. This included rape, choking, putting me in crutches and forcing my mother and father to watch. P1178.13 PLTF EX-02801 0013 CONTEXT EXAMPLES: THINSPIRATION < dietcoke.calories 35 Posts 1,184 25 Followers Following thinspo current theme: pink NOT PRO DONT REPORT ME No Image JUST BLOCK ME -pro recovery- twitter: starvingiscool Follow Message ♡ < tw.thinspo_and_memes !TW! 23 Posts 683 134 Followers Following P1178.14 She/her. Not pro. All memesare mine unless stated. Love EVERY ONE OF YOU(.) Follow Message ABC diet Calories Exercise Calendars Me "We're not teaching [others] to say, hey are you okay? We're teaching them to click a report button or to call the cops and try and trace the phone or trace the location based on a Tweet. I'm kind of afraid of the resources because of that." • • One type of triggering content related to body image issues Generally not policy violating, easy to seek, often leaks into general fitness discovery experiences • Posters often suggest solution is self-support, not enforcement: "Block, don't Report", "Trigger Warning" Can trigger harassment for poster PLTF EX-02801_0014 2021 Proposal P1178.15 12 PLTF EX-02801 0015 PROPOSAL MEASURING, REDUCING, RESOLVING BAD EXPERIENCES What does success look like by end of 2021? Operationalize "bad experiences” measurement for experiment launch decisions and goaling to empower teams to work on problems in policy gaps. P1178.16 13 PLTF EX-02801_0016 PROPOSAL DEFINING BAD EXPERIENCES P1178.17 Measurement Tracks Reach Measure the number of people affected by bad experiences Support Effectiveness Measure how effectively we're resolving a bad experience No. covered here: 3rd party ("reputational") reach or intensity separate from reach Other definitions: Intensity = felt experience by users Severity = combines risk to business and risk of real world harm for people with felt intensity research PLTF EX-02801_0017 P1178.18 Slide 1 2 3 4 Why not intensity? Seems like it's as important as reach if we think working on something like mass harassment should be incentivised Miki Rothschild: 11/19/2020 3:10:39 PM There's a pretty complete answer to this common question on the comment thread on this slide. cc'd you there. tldr: intensity as a pure metric separate from reach hasn't yet shown any applications. severity-bucketed reach and intensity-weighted reach are both variants inside the Reach track Sam 11/19/2020 3:10:39 PM Common question: Shouldn't we consider intensity and/or severity? Both of these can fit inside reach to a degree. For example, how FB XI handles: 1- severity buckets for combining problems into "high harm" and "low harm" for TRIPS 2- "intensity-weighted reach" for TRIPS, where the binary "Yes, I saw in the last 7 days" is combined with the 5-point intensity scale to create metric that weights individual responses by intensity. Definitions: Intensity = felt experience by users Severity = combines risk to business and risk of real world harm for people with felt intensity research Sam 5 cc @ Sam 11/19/2020 3:14:36 PM @fb.com 11/19/2020 6:24:05 AM PLTF EX-02801_0018 PROPOSAL MEASURING REACH P1178.19 Reach measurement in 2020 Not covered here: Hard Life Moments for mental well-being concerns Violating prevalence impressions Labeled and classifier estimated metric for narrow policy violating definition. Available for a subset of problems based on CI + IGWB prioritization and operational cost. FRX Only negative user feedback with problem detail, with problems as defined by Community Standards policy. Available as aggregate for all reporting problems (dominated by spam) and by violation types and problem tags. TRIPS Survey for 13 integrity problems launched on IG in Dec'19. Also has Civic variant. 15 PLTF EX-02801 0019 Slide 1 P1178.20 1 6 It would also make sense to look into whether we can get clear signal from self-remediation efforts -- what does it tell us when someone switches their feed from public to private, or blocks someone? Andrew 11/19/2020 6:52:12 PM Blocks are in pTRIPS and USI. We (Foundation data and Bully eng) have been working to improve the attribution of account level actions so they're more predictive of problems. I don't think we've considered switching from Public to Private recently. There was some foundational analysis by CDS in H1. Probably harder to do attribution for but maybe our recent attribution work could help here. 1 2 1 7 cc @ @fb.com @ @fb.com Sam 11/19/2020 6:20:39 PM the recent switch is a good idea. We only look at whether it is currently private as a feature Carlos 11/19/2020 6:42:51 PM the logging right now is terrible, but long-term there's a lot of potential signals here even people viewing settings pages (people generally don't look at privacy settings just because they are curious) or even stuff like unfollowing/'not interested' accounts/content that lots of other people unfollow, or even people using the search box to try to find out how to get help Andrew 11/19/2020 6:52:12 PM since we're not talking about hard life moments and we're primarily looking at integrity+ (borderline experiences) - is it right to say this is primarily excluding measurement in the non-integrity well-being space for 2021? Lori 11/19/2020 7:59:52 AM No, not intended to exclude other 2021 paths. The intent of this deck is as a conversation starter. In 2020, HLM wasn't implemented as an ongoing metric like the other items on these slides. Assuming that if we can get momentum here in the integrity space, it helps us leap forward with bad experiences in other areas (ie. mental well-being). Sam 11/19/2020 7:59:52 AM PLTF EX-02801_0020 P1178.21 H2 learnings about Reach measurement Violating prevalence impressions While responsive to some changes, doesn't capture most personalized demotion impact in comments (similar on FB XI). FRX Per-product/per-surface cuts matter for seeing per-experiment impact. "FRX feedback" more inclusive than "FRX submit reports". TRIPS Responsive to IG Bullying H1 changes in comments slice (-5%). Responsive to world events, including at sub-problem detail: Since George Floyd protests in June, in the US "threats" portion of B&H and "political beliefs" for HS are elevated (ss) PLTF EX-02801_0021 P1178.22 Slide 1 3 8 One measure we could look at is sensitivity, as measured by the portion of experiments that are known negative/positive and are captured with metrics movement. Curious how well TRIPs scores on that dimension Miki Rothschild; 11/19/2020 6:42:02 AM Yes. Per the last WB review on pTRIPS, that's how we agreed to come back when we're focused on details to do a selection discussion. That level of detail was outside the scope of this review. Here just trying to share some high level updates since we last discussed. Sam 11/19/2020 6:42:02 AM PLTF EX-02801_0022 P1178.23 H2 learnings about Reach measurement (continued) US Most weight goes to FRX when predicting policy violations. Operationally challenging to add new signals in current central governance model. PTRIPS Super early results show some stat sig effects for 2 IG Bullying experiments (Remove Preview Filter Toggle, Personalized Demotions). TBD if surveys validate prediction. Product cuts for both pTRIPS proxy and TRIPS surveys probably quite important. FB XI pursuing similar approach in close collaboration with IGWB and CDS working group. Different in details. PLTF EX-02801_0023 P1178.24 Slide 1 4 9 Is there any way to look at a white box description of both of these? I think it's pretty crucial for predictive metrics to be intuitive in order to gain intuition of the limitations Miki Rothschild; 11/19/2020 6:55:13 AM Yes, in the last execution review that was narrowly focused on these we agreed to do a detailed follow up execution review for selection to land the H2 work. We'll need to do a deep dive there. This is just a light progress update on H2 progress this group may not be aware of in order to inform this 2021 topic. Sam ; 11/19/2020 6:55:13 AM PLTF EX-02801_0024 PROPOSAL MEASURING SUPPORT EFFECTIVENESS P1178.25 Support Effectiveness measurement in 2020 Adoption See how many people use our new support experiences (ex. Support Requests, Restrict, Bulk Comment Management) Supportiveness IX-developed survey: "How supportive was Instagram during this experience?" triggered after completing a specific flow LEGIT Covers much more scope than effectiveness for people experiencing problems (1st party). Responsive to actor transparency changes on IG. 18 PLTF EX-02801 0025 How to Get There P1178.26 19 PLTF EX-02801_0026 Major Next Steps P1178.27 1:10inish H2 execution TRIPS holdout read for IG bullying, pTRIPS validation, USI development. We'll have new data before Dec 17 roadmaps. 2: Coordinate with CI IMI, OC, IX At least CI IMI is actively proposing roadmap priorities that would help advance work in this area. 3: Prioritize H1 steps to unblock IG teams Which metrics would unlock the most impact? What needs Foundation and IMI help vs. being driven by teams? 4: Decide IG Bullying goal / launch metrics H1'21 is a key inflection point for IG Bullying in its 5th half of working in this direction. WB is taking Bullying PO goal. PLTF EX-02801 0027 Slide 1 5 10 P1178.28 I believe the bullying po would be oriented towards creators. Does that changes things or same measure? Miki Rothschild: 11/19/2020 3:07:39 PM I assume expanding the bullying focus to creators is more of a shift in focus than a complete change for measurement. For bullying measurement strategy, I think we'll want to consider: 1- balance of focus between "experiencing bullying" (targets) and "seeing" (witnesses) 2- how measurement execution details vary between creators, teens, and gen-pop 3- how much product strategy mirrors measurement strategy For #2, there are going to be some measures that will have CIs that are too large if we directly filter to creators. But they can still be useful when triangulating with other measures with good granularity for creators. For #3, from what we've seen with teens, we can inspire the product strategy from measurement of the problem and effectively develop products with a teen focus, then use more gen-pop measurement during launch experiment review as part of launching for everyone. So less of a full mirror between product and measurement, with more push to see the focus come through in the product strategy than the measurement. Sam 11/19/2020 3:07:39 PM PLTF EX-02801_0028 Discussion P1178.29 Feedback on ambition and examples? Is this ambitious enough? Too ambitious? Resonate as important and people-first? Feedback on approach? Is the framing intuitive? Leverage existing concepts and terminology for "extreme clarity"? Feedback on next steps? At a high-level, are we missing major steps for getting ready for H1 roadmaps? PLTF EX-02801 0029 P1178.30 Slide 1 2 6 11 not all these measures are at the same stage of evaluation and understanding. I think I'm not clear on how we're thinking teams will use all these metrics? Is the idea to use a basket of metrics for problems? Lori 11/19/2020 6:30:53 AM My main feedback is I'm worried that there's a metrics explosion in this space while it's unclear which ones are actually effective. I'd love for well-being to mostly follow others on metrics that have already proven effective and make a bet on one where we want to lead but not try to chase all of the failures with the rest of the pack because that'll be very inefficient. So the key question for me is: 1) which metrics are proven effective with great conviction that we should follow? 2) if we have to choose one metric that's so important to us that we're willing to make a bet on it, which one would that be? Miki Rothschild; 11/19/2020 6:49:04 AM Yes, this feedback seems to create inputs for step #3 on slide 20 (prioritize based on what would create impact for our teams). Sam 11/19/2020 6:49:04 AM PLTF EX-02801_0030 P1178.31 22 PLTF EX-02801_0031 CONTEXT EXISTING METRICS Track Metric Reach Violating prevalence FRX TRIPS USI Question metric answers P1178.32 How many impressions are seen for content violating CS policy? How many people report content and accounts? How many people have seen or experienced the problem in last 7d? 5-point intensity? What's the aggregate of reporting, blocking, unfollowing, and other actions? (with aggregate weighted for correlation with CS violations) PTRIPS Of the aggregate of user actions and other signals, what's our prediction for TRIPS? Support Effectiveness Adoption How many people adopt the supportive product solution? Does it cannibalize? Adoption within problem segment How many people with a specific problem adopt the supportive product solution? Supportiveness How supportive was Instagram during this experience? LEGIT Do people believe that our integrity work is effective at reducing harm and that our enforcement is defensible and fair? For the purposes of an integrity guardrail, Peter has posted options for a single metric: https://fb.workplace.com/groups/1540209922820872/permalink/167911671226 3525/ Guardrail and primary positive impact metrics will have different needs PLTF EX-02801_0032 CONTEXT ASSESSING EXISTING METRICS P1178.33 Metric Shows progress in policy gaps? Already proven in Nov'20? No 12 Track Reach Violating prevalence Yes More progress landing soon? Surface improvements landing Works for low- reach? No (except UB Prevalence) FRX Yes Yes, for some use cases Improving logging quality ? Team(s) with most progress CI OC + IG Foundation CI IX, CI OC, IGWB Foundation, IG Bullying, US2020 TRIPS Yes Yes, for some use cases Major new results landing in Dec No FB XI, IG Bullying USI No No, not operationalized H2 IG project in flight ? CI IX, CI OC PTRIPS Yes No, not operationalized Major new results landing in Dec ? IG Bullying, FB XI For the purposes of an integrity guardrail, Peter has posted options for a single metric: https://fb.workplace.com/groups/1540209922820872/permalink/167911671226 3525/ Guardrail and primary positive impact metrics will have different needs PLTF EX-02801_0033 Slide 1 12 Used as primary proxy metric for if problems were healthy during COVID surge and US2020. Sam 11/19/2020 5:20:43 PM P1178.34 PLTF EX-02801_0034 CONTEXT ASSESSING EXISTING METRICS Track Metric Shows progress in policy gaps? Already proven in Nov'20? More progress landing soon? Works for low- reach? ? Yes Yes Support Adoption P1178.35 Team(s) with most progress IG Bullying Effectiveness Adoption within Yes problem segment No, initially proposed for Limited Profile Yes IG Bullying Supportiveness ? Yes, but for different use ? CI IX LEGIT ? case 1 early proof point for different use case ? IGWB Foundation, CI Legitimacy PLTF EX-02801 0035 PLTF EX-02801_0036