Email: Re: Instagram Well-being Research September HPM

Raw text

Transcribed text as recorded, including headers and formatting fragments.

Exhibit 16 Message From: \_ \[/O=THEFACEBOOK/OU=EXTERNAL (FYDIBOHF255PDLT)/CN=RECIPIENTS/CN=95620A4848804CA98003431EF8051326\] Sent: To: \_ 10/1/2019 3:46:03 PM Subject: Re: Instagram Well-being Research September HPM Attachments: image001.png; image002.png; image003.png; image004.png These are so awesome at raising our awareness and cross-org collaborations, thank you for continuing to lead in elevating our work\! | love the imagery incorporation, really stands out and helps it be more digestible and memorable. I'm excited to learn more about what- finds in defining what a supportive community looks like. Max is 'very' drawn and engaged with the "supportive" element in community, we had some convos on rethinking our Mission where I got more insight on how each lead is thinking of it. Anyone you'd want to make sure I elevate and incorporate into the Community RYSK? Also appreciate the shout out© From:\_ Date: Tuesday, October T , lG\_Well-being\_Leads 1, 2019 at \_'—\_ 6:15 AM \<lG\_WeII—being\_Leads@fb.com\>, Instagram Researchers \<ig-researchers@fb.com\>,\_ Subject: lnstagram Well-being Research September HPM Things it's been methodology also pretty are been always amazing great and- busy\! to and work We inspiring even just represent wrapped more to see closely lnstagram how our the this with research community Research our FB Cl summit with and has CVI \#realtalk grown. last counterparts week, It on was bringing the great research and to together researchers see leadership our researchers own— across panel. our from lnstagram across co-lead the surface fami a w teams'gsjwe rkshfgcan of afis. D (J Of": launch into the US 2020 lockdown work. Excited to see the results of this focus at the end of this. Our team is also in the thick of 04 andrliil n kg @ey 7" co N 2w an} . a» 2: vZJ'" O 9195' -\< +7 -— "" CONFIDENTIAL - PROTECTED BY CONFIDENTIALITY AGREEMENT MT-IG-AG-00105112 planning, and we have a lot of research projects coming in hot. See below for some of our latest understanding of how the Aggression taxonomy overlaps with our lnstagram non-recommendable guidelines, lnstagrammers' experience of borderline (non-violating content), misinformation treatment testing, and more\! Highlights o Users' experience seeing non-violating problematic content (POC:—): This study aimed to explore what types of non- violating problematic content users encounter, what they do when they see this, exactly what types of control users want, and what they expect lG to do. For this research study, we presented users with a range of examples across different content areas (e.g., nudity, SSl) and discussed their experiences, reactions and wants. Users described content as falling into 3 main categories "fine/0k", "upsetting", and "disturbing". Disturbing content included content that evoked feelings of surprise, fear, and confusion. In response to seeing disturbing content, users wanted to connect with others. Many did this by reaching out to friends and family members and sometimes even sharing the content. Intent of the content was key in how users wanted us to deal with it. Content with negative intent was content that evoked anger, fear, and confusion, and in some cases the desire to retaliate. CONFIDENTIAL - PROTECTED BY CONFIDENTIALITY AGREEMENT MT-IG-AG-00105113 DISTURBING CONTENT Most users didn't want to see this at all Less familiar, More "mun" pointed (not general), gene", ('7""intEd) targeting defenseless victims Fine/OK Upsetting Disturbing Port of life Arguments/Debates Graphic news Free Speech Some humor Graphic violence Arguments/Debates Opinions contagion Some humor Rumors Fake news/misinfo Illegal things Hate o False Information Inform Treatment Foundational Insights (POC:\_): We conducted US-only interviews with teens and adults of mixed political affiliations to understand participants' expectations regarding our role with managing false information on our platform. We found that 1) most participants did not believe it was lnstagram's responsibility to validate and communicate whether a post was factual. Participants felt it was the user's responsibility to confirm whether things they viewed, posted, or shared were true. 2) Participants appreciated that Instagram was taking steps to fact-check content but most did not want us to remove content. Participants also expected friction when trying to view or share fact-checked content. 3) Participants expected lnstagram to be transparent about why content was fact-checked, when the false information treatment is applied, why actions were taken towards repeat offender accounts, and whether the account owner had the opportunity to appeal a decision. CONFIDENTIAL - PROTECTED BY CONFIDENTIALITY AGREEMENT MT-IG-AG-00105114 o False Information Inform Treatment Usability Research (POC:\_): We conducted US-onIy interviews with teens and adults of mixed political affiliations to understand whether our false information inform treatment was easy to understand and provided enough information for participants to believe the 3rd-party fact checker's determination. We found that 1) participants were receptive to Instagram obscuring fact-checked content and felt they were likely to trust that a post was false if we applied the treatment. However, most were curious to see how extreme the post was or discover whether it was something they had previously believed was true. 2) All participants favored direct and concise language and wanted additional information about the account/post to help them form their own opinion. and 3)The Repeat Offender flow was clear and easy to understand due to the prominent warning next to the account name. The options presented when the warning was tapped provided participants with a sense of control over their experience. o 5 graphs about reportinngOC:—): I looked back at some survey data from the Negative Experiences survey (NETS) I ran in May, focusing on the question of whether people reported the negative experience that they had ("Did you report this prob/em to Instagram ?"). | found that intensity predicts self-reported reporting: more intense problem types have a higher self-reported report rate. Furthermore, there are three distinct clusters of problems: 1) low-intensity 81 low-report (in blue), 2) high—intensity & high-report (in purple), and 3) very high-intensity yet only medium-report (in pink) — this outlier group contains seeing others getting bullied and seeing SSl content, which may be problems where witnesses aren't sure that it's their place to intervene despite being very upset. REPORT RATE X PROBLEM INTENSITY 6055ssi. bullying\_;eeing 50hate compromisgfixperiencing0 violen.ce buflyin-g\_experiencing ' solicitation firearms nudity 35mismformatlon.ospan1 commercml 25 SllSplCIOUS accou11l 20 Ie experiencmg. R = 0.61 . p = 0.022 20 25 30 35 40 45 50 percent reporting CONFIDENTIAL - PROTECTED BY CONFIDENTIALITY AGREEMENT MT-IG-AG-00105115 o How much aggression is also non-recommendable \[impressions edition\](POC:—): I reanalyzed the aggression & non- recommendabie overlap dataset, breaking down the aggressive content by the number of impressions it received. When | split aggressive photos by the number of impressions they got in the 72h after posting, the overlap between aggression and non-recommendable is fairly stable across different numbers of impressions, except an unexplained spike in the 1001-5000 impressions bucket. When I split aggressive photos by the number of Explore grid impressions they got, the overlap is smaller for photos that have any Explore impressions at all. As the number of 72h or Explore impressions goes up, fewer photos are non-recommendable because of bullying and hate speech, and more of them are non-recommendable because they're sexual or disturbing. o How much aggression is also non-recommendable? (POC:—): | ran a content overlap analysis to better understand the relationship between aggression guidelines & non-recommendable guidelines. This revealed: 1) About half of aggressive photos are also non-recommendable, suggesting that we can almost double the amount of aggressive content we flag by using aggression guidelines, over non-recommendable guidelines; 2) Of all the content in the overlap of aggression and non-recommendable, less than half is non- recommendable because of bullying, and almost as much content in the overlap is non-recommendable because it's sexual; 3) About 9 in 10 photos that are non-recommendable for bullying, hate speech, or threats are also aggressive, suggesting that aggression guidelines may capture almost all of the content that's non-recommendable for these reasons. CONFIDENTIAL - PROTECTED BY CONFIDENTIALITY AGREEMENT MT-IG-AG-00105116 ABOUT HALF OF AGGRESSIVE PHOTOS ARE ALSO NON-RECOMMENDABLE Non-recommendable Aggression Only E aggression: E .54 '32-» U 46% N = 4941 photos labeled as aggressive. then cross-labeled for non-recommendability Caption: Just over halfofaggressive photos in this sample were also judged to be non-recommendable — meaning thatjust under halfof the aggressive photos were not currently non-recommendable. This suggests that there's a large potential opportunity: we can almost double the amount ofaggressive content weflag by using aggression guidelines, over non-recommendable guidelines. o Feedback on Restrict: A mixed methods study on the use and usability of Restrict (POC:—): The goal of this research was to understand the use and impact of Restrict after launching to 5% English speakers worldwide. Overall, we found that users: 1) used Restrict primarily in response to unwanted contacts and insults against known actors, 2) Compared to Block, Restrict was harder to find - Users commonly discovered Restrict in Profile when looking for block, 3) Compared to Block, Restrict was significantly harder to understand and was a key barrier to use. In interviews, most users could not fully describe how Restrict worked. Success was commonly attributed to functions not included in Restrict (e.g., decreased visibility or inability to DM). o Restrict Survey Results: An update on Restrict use, usability, and understandability (POC:—): This follow-up survey on Restrict aimed to understand changes in use, usability, and understandability of Restrict among user globally (5% English speakers worldwide, 100% Australia). This survey revealed that: 1) While users continued to use Restrict and Block against a range of aggression CONFIDENTIAL - PROTECTED BY CONFIDENTIALITY AGREEMENT MT-IG-AG-00105117 types, use of Restrict against unwanted contacts in the US is decreasing, 2) Compared to Block, Restrict was less effective in completely resolving issues from US and AU respondents, but more effective in partially resolving the issue, and 3) Restrict continues to be harder to understand compared to other tools (and most notable in lD). Users made many assumptions about Restrict. Primary misconceptions were centered on visibility of content and ability to DM. a Users wonder why we only take action on inauthentic likes/followers, but don't take down IG accounts promoting these services... (POC —): Users self-com promise via third-party apps and websites (3PAs) in order to increase their engagement, which is inauthentic and violates our guidelines. We've approached this almost exclusively as an external problem, assuming that users seek out these services via off-platform resources However, users in research sessions adopted 3PA services based on recommendations from friends on IG, seeing accounts on IG offering inauthentic engagement, and via direct advertising from 3PA services most likely via data scraping. This doc assesses how our own platform can connect users to 3PA services just by in-app searches, and presents common characteristics | found across "'50 accounts. o Mental Health: A Quick Overview (POC:\_): l reviewed both internal and external research on mental health and well-being to better understand how mental health is defined, why lG should care about it and how it is impacted by social media use. (1) Mental health is defined as a state of well-being in which every individual realizes his or her potential. It therefore comprises of both positive (subjective well-being) and negative (mental illness, psychopathological symptoms and distress) mental health, which are two correlated but distinct factors. (2) On Instagram, mental health issues impact one of our core demographics (teens and younger users), as they generally demonstrate lower well-being. Mental health issues are also a leading driver of well-being complaints related to lnstagram. (3) Social media is often not the cause of problems related to mental health. However, it can and does both attenuate or exacerbate a user's experience with mental health issues. (4) Therefore solutions to mental health issues should focus on both promoting positive mental health outcomes (enhancing overall well-being) and preventing negative mental health outcomes (prevention and support for the most distressed users). In their own words... The section where we share key quotes or highlight videos from research to bring the instagrammer perspective directly to you.-nterviewed and surveyed lnstagrammers to understand the use and impact of Restrict after launching to 5% English speakers worldwide. Participants explained when and why it was really helpful to have a tool like Restrict. CONFIDENTIAL - PROTECTED BY CONFIDENTIALITY AGREEMENT MT-IG-AG-00105118 17"P';.:l'—\_|NTERV|EWS PROVIDED MORE COLOR FUR RESTRICT USE CASES AGAINST SPECIFIC ACTORS Situation and People \- - - - - Temporary Good To Bullying Unwanted you business manage wont precursor by account to contacts fights mean/inappropriate all know types to that with who block from of want people they actors for people to are/what all keep gou (n=6) types comments who know followers they of gou actors (n=14) don't wont for (n=2) (n=3) public (n=3) know when or IV .\~\_'\* A friend could have made you mad but then Female, apologized. Adult Users liked that Restrict... Sometimes goujust - - - Doesn't Doesn't aggravate show read the state situation on DMs bg - feelings classmate message Makes them (particularly could or feel didn't cause like do in they actors anything situations are to not harass particularly when Seeing alerting that the they actor read them further. hurting the the person others' is a bad) the peace. wont to Blocking keep the can make others - Mole, Adult get upset '2 users (unwanted contacts from exs) said that (he other person should know they are Restricted so they know they did something wrong 22 Upcoming Research Please reach out to the research POC to learn more about the upcoming research or to participate in and attend research sessions or readouts. o o We experiences Restrict conducted Diary using Study Restrict Restrict 9/27-10/27 research via in-degthjnterviews (POC:—): in Australia 9/23-9/27 and fgcus (POC:—). Eggs. In this diary study, we will In this continue research to monitor study, we the explored use of Restrict users' in Australia. Key areas that we will continue to explore include: how use changes, impact on the user, impact on others, evolution of understanding etc. CONFIDENTIAL - PROTECTED BY CONFIDENTIALITY AGREEMENT MT-IG-AG-00105119 o Bullying on Direct survey (POC\_) \- The top surface for experiencing bullying is Direct. What does bullying on Direct look like? This project is aimed at sizing different bullying patterns on Direct — what types of bullying occur, how much bullying is from known contacts vs. strangers, where bullying originates from (entry point), how much involves rich media, and what tools people engage when thevre bullied. o Hard Life Moments(POC—) \- This survey is aimed at understand the role that instagram plays in the hard moments of peoples' lives — for example, after a death in the family, when feeling lonely or sad, or when under stress from big life changes. How does Instagram magnify or reduce harm, pain, support, etc during these moments? Where could we be better supporting our Instagram community? What hard moments should we be prioritizing? o Bullying market landscape(POC\_ &—) — Survey to understand how Instagram fares against other social media apps on perceptions of online bullying. o Supportive Communities on Instagram (POC:\_) \- Remote qual research in the US for understanding what a supportive community looks like offline, online and specifically on instagram. Awesome Other Insights to Know o Bullying Insights from Instagram Stories(POC\_). Justin has been diving in to understand bullying in Instagram Stories. He's started by understanding reporting in stories and has been delving in deeper to understand bullying in stories. For example: 1) Are there gender/age differences in what content gets reported? 2) When is a report more likely to be about bullying? 3) Where do stories get reported from? See this group for more analyses. Please reach out if you have any feedback or questions. o Defining a Misinformation Actor on Instagram (POC:—). For misinformation enforcement on instagram, an actor should be defined as a public account with high reach or high followership (in this note, we propose 1K followers or more as criterion for high followership). Given the low volume of direct 3PFC factochecks on Instagram posts, the R0 threshold shouldn't distinguish between matched and 3PFC-fact-checked posts. We recommend setting the threshold at 2 matched hoaxes for R0 status, allowing us to treat up to 28% of public accounts sharing misinformation on Instagram in the past 30 days. o Users vulnerable to integrity harms tend to also face multiple overlapping vulnerabilities in life (e.g., homelessness, disability, substance addictionl (POC:—). Through interviews with vulnerable users in Memphis, TN we learned that these users operate in low quality information ecosystems and have high volumes of recommended content in their News Feeds. 2 additional phases of qualitative research, plus partnerships with Data Science are planned to explore harms amongst the subset of users who are targeted for a high amount of recommendations. CONFIDENTIAL - PROTECTED BY CONFIDENTIALITY AGREEMENT MT-IG-AG-00105120 geponing in Messenger is subject to low discoverabilitv, low awareness, and 'mvths' of how it works(POCs:—, - -,—). When Messenger is end-to-end encrypted, will be much more reliant on user reporting to gain insights and action on abuse. Three researchers summarized the Facebook responsibilities that emerged from UX research with 145 participants across 4 countries between H2 2018 and H2 2019. Responsibilities: address the myths that reports are not reviewed or prioritized; that reports involve the police or government; that reports notify parents or community members; that reports will reveal the reporter identity to the reportee. o Perceived Safety in Groups (POCs:\_, \_-, —). How do feelings of safety in groups change over time for users? .Through tracking and analyzing +10K users on feelings of trust and safety in groups, research scientists from the Core Data Science and Groups UX research teams are working together to examine how feelings of safety changes over time. Key findings includes: 1) Feeling high levels of safety in a group is associated with posting more frequently, writing more text, and spending more time in the group over time; 2) Admins' engagement can make members feel safer over time; 3) Increases in commenting and time spent by the respondent over time are associated with increases in feelings of overall safety. People— \- has transitioned off of lG Well-Being Account Problems and is gearing up for new work with IG Growth Acquisition and other opportunities on the horizon — - super excited about hanging with my nieces this weekend\! - has a laser cutter and is lasering all the things\! \_ — is secretly excited that my fiance is out of town for the week, so | can 'uy a quantity of plants that he will definitely find excessive \_ - is planning to catch up with friends and plan for her thanksgiving trip to Japan. \_ \- excited to show a recently relocated friend around SF. Looking for past Well-being research? Try our well-being research archive: IG WB Research Project Tracker and Research Archive CONFIDENTIAL - PROTECTED BY CONFIDENTIALITY AGREEMENT MT-IG-AG-00105121