Meta & Children's Online SafetyReels Mimicry integrity risks
Raw text
A/C Priv: Reels Mimicry integrity risks Kabbabe Exhibit 5 2/4/26 Jenn Dunn RMR CRR TL;DR • Roughly ~10% of reels associated to trends could come from underage users (<13) • We plan to scale mimicry by amplifying trends in H121 potentially making the size of problem bigger Problem It seems that we are not equipped integrity wise to address this challenge yet • Reels format has a particular appeal for teens but as well underage users • Trends trough mimicry seem to exacerbate the problem by making creation accessible to very young audience • We plan to scale mimicry in H121 by amplifying trend distribution, causing a potential risk to make the problem even bigger. * Based on early intermal investigation, we are not equipped to property size/ address the problem Preliminary sizing • We used the Mimicable Active Trend heuristic to identify trends in US, IN and ROW • We performed a visual observation to identify underage users (<13), this is a very rough estimate and needs further validation • A rough estimate gives 9-10% of reels associated to trends could contain underage users (based on a random sample of 400 reels in most popular trends in the US and ROW (excl IN)) • RoW (excl IN): https://fourl.com/instagram/0qre9a34 • US: https://fburl.com/instagram/c0wyy1by • Some trends seem to target particularly very young user (>90+% of reels that could feature underage users) • source: https://fburl.com/instagram/d9k0rg11 HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) PLTF EX PLTF EX-04329 ΜΕΤΑ3047MDL-294-00045120 METANMAG-159-00045120 PLTF EX-04329_0001 exhibitsticker.com Next steps Understand the current systems / guardrails to limit the distribution of reels with underage users Invest in manual curation to accurately size the problem • Depends on: PXFN approval of guidelines and process • Investigate potential solutions: Contact Support • Build ML systems to detect reels with underage user and limit distribution • Detect trends targeting underage users and filter out those trends from the trend amplification system (through curation or ML) Conversation History Quipbot#06 left the thread Aug 9 at 8:07 am Quipbot#06 locked edits on the document Aug 9 at 8:02 am Quipbot#06 Aug 9 at 8:02 am =====IMPORTANT===== This Quip file has been copied to Google Workspace (https://docs.google.com/document...) as part of the Quip2Google migration efforts. Quip is moving to Read-Only. Please work in the new Google file moving forward. Learn more htips://fb.workplace.com/groups... Questions? Visit the Quip2Google Migration Feedback Group (https://fb.workplace.com/groups...). HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) ΜΕΤΑ3047MDL-294-00045121 METANMAG-159-00045121 PLTF EX-04329_0002 Note New edits in this Quip file will NOT be automatically synced to Google Workspace. Quipbot#06 added Quipbot#06 to the thread Aug 9 at 8:07 am Quipbot#06 added you to a document Aug 9 at 8:02 am Quentin Chalvon added Tomas Kabbabe to the thread Jan 28, 2021 at 7:55 pm Quentin Chalvon added you to a document Jan 28, 2021 at 7:49 pm Shell Megan opened your conversation with Adam Tomas Kabbabe, Jon Palak Erin Quentin Chalvon Lauren Escher Anthony and Karina Newton Jan 26, 2021 at 8:33 pm Palak added you to a document Jan 22, 2021 at 11:17 pm Palak dded Karina Newton and Lauren Eschen o the thread Jan 22, 2021 at 11:22 pm Palak added you to a document Jan 22, 2021 at 11:17 pm Palak opened your conversation with Adam Erin Lauren Escher Anthony Megan Tomas Kabbabe, Sheli Jon Quentin Chalvon and Karina Newton Jan 22, 2021 at 11:13 pm Adam opened your conversation with Erin Lauren Escher Kabbabe, Sheli Jon Palak Quentin Chalvon Anthony Megan and Karina Newton Jan 22, 2021 at 8:12 pm Quentin Chalvon Jan 22, 2021 at 8:06 pm I added the details here: We used the Mimicable Active...: A/C Priv: Reels Mimicry integrity risks Jon added you to a document Jan 22, 2021 at 8:04 pm Jon added you to a document Jan 22, 2021 at 8:04 pm Jon added Megan Palak and Sheli to the thread Jan 22, 2021 at 8:10 pm HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) Tomas ΜΕΤΑ3047MDL-294-00045122 METANMAG-159-00045122 PLTF EX-04329 0003 Jon added you to a document Jan 22, 2021 at 8:04 pm Quentin Chalvon edited Jan 22, 2021 at 8:02 pm • We used the Mimicable Active Trend heuristic to identify trends in US, IN and ROW • We performed a visual observation to identify underage users (<13), this is a very rough estimate and needs further validation Erin Kabbabe, Sheli Jon ppened your conversation with Adam Palak Lauren Escher Anthony Megan Tomas Quentin Chalvon and Karina Newton Jan 22, 2021 at 7:36 pm Jon Jan 22, 2021 at 7:35 pm Quentin can you lay out the specific heuristic criteria you used for "trend" here, as well as a couple bullets of your guidelines to determine positive vs negative labels Jon added you to a document Jan 22, 2021 at 7:34 pm Jon added Adam and Erin o the thread Jan 22, 2021 at 7:40 pm Jon added you to a document Jan 22, 2021 at 7:34 pm Jon edited Jan 22, 2021 at 8:05 pm A/C Priv: Reels Mimicry integrity risks • Depends on: PXFN approval of guidelines and process Jon renamed Reels Mimicry integrity risks to A/C Priv: Reels Mimicry integrity risks Jan 22, 2021 at 7:10 pm Jon opened your conversation with Adam Tomas Kabbabe, Sheli Palak Erin Quentin Chalvon Anthony Megan Lauren Escher and Karina Newton Jan 22, 2021 at 7:02 pm Quentin Chalvon enabled link sharing View/Edit Jan 22, 2021 at 2:35 pm Quentin Chalvon added you to a document Jan 22, 2021 at 2:29 pm Quentin Chalvon added Anthony and Jon to the thread Jan 22, 2021 at 2:34 pm Quentin Chalvon added you to a document Jan 22, 2021 at 2:29 pm HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) ΜΕΤΑ3047MDL-294-00045123 METANMAG-159-00045123 PLTF EX-04329 0004 Quentin Chalvon * edited Jan 22, 2021 at 2:28 pm It seems that we are not equipped integrity wise to address this challenge yet Quentin Chalvon edited Jan 22, 2021 at 3:14 am Reels Mimicry integrity risks TL;DR • Roughly ~10% of reels associated to trends could come from underage users (<13) • We plan to scale mimicry by amplifying trends in H121 potentially making the size of problem bigger Problem It seems that we are not equipped to address this challenge yet • Reels format has a particular appeal for teens but as well underage users • Trends trough mimicry seem to exacerbate the problem by making creation accessible to very young audience • We plan to scale mimicry in H121 by amplifying trend distribution, causing a potential risk to make the problem even bigger. • Based on early intemal investigation, we are not equipped to properly size / address the problem Preliminary sizing • A rough estimate gives 9-10% of reels associated to trends could contain underage users (based on a random sample of 400 reels in most popular trends in the US and ROW (excl IN)) • ROW (excl IN): https://fburl.com/instagram/0qre9a34 • US: https://fburl.com/instagram/c0wyy1by Some trends seem to target particularly very young user (>90+% of reels that could feature underage users) • source: https://fburl.com/instagram/d9k0rg11 Next steps 8 Understand the current systems / guardrails to limit the distribution of reels with underage users Invest in manual curation to accurately size the problem • Investigate potential solutions: • Build ML systems to detect reels with underage user and limit distribution • Detect trends targeting underage users and filter out those trends from the trend amplification system (through curation or ML) Quentin Chalvon created the document Jan 22, 2021 at 1:47 am HIGHLY CONFIDENTIAL (COMPETITOR) HIGHLY CONFIDENTIAL (ATTORNEY'S EYES ONLY) ΜΕΤΑ3047MDL-294-00045124 METANMAG-159-00045124 PLTF EX-04329 0005