Meta & Children's Online SafetyChat: Somewhere between zero and negligible. Child safety is an explicit non-goal this half.
Thread View
████ name redacted. Different bars may be the same person or different people.
quoted in an excerpt in the archive.
Thursday 9 July 2020
08:08████Good morning! FYI my xfinity internet has been down all night and it was flickering off and crappy yesterday. So I may be just calling in to our reviews
08:23████I've done a couple walking around outside on phone. It works great! 100
08:35████I tried a few of those early on in shelter but what's challenging is that the cell signal is also not strong around my neighborhood
08:35████So it made 1:1s challenging : (
08:37████Time for one of those 1970s era phones with super long cord? Or maybe a 1980s brick+antenna combo? ;)
08:38████Shared an image
08:59████Good morning leads! What will the day bring today? :)
08:59████what is the plan for the Monthly check-in with Comm. Leads?
09:00████@Miki shared a first draft doc via email last night.
09:00████I'm looking at it right now :)
09:00████https://fb.quip.com/H5GsAabcS8mk
09:03████Well, on the plus side, I guess our goals are aggressive enough :) Lots more missed this half.
09:18████I like the on-track/off-track grid.
09:22████Hey, folks! Just FYI that though is still in boot camp, she may be attending the check-in with Community leads today. The H1 retro and H2 lookahead is hard to pass up.
09:25████How long is her Bootcamp?
09:26████It's supposed to be two weeks but she's steamrolling through a lot of the (mostly self-driven) content.
She's likely going to start poking in early next week.
09:47████Just FYI, I have to drop off the Bullying mtg at the halfway mark to make a 10:30 pxfn review
09:51████I also will need to drop off halfway through for a wb measurement review w/ EDS KX and John hegeman
09:51████At 10:30
09:52████I'll be a few min late to this meeting running between leadership candidate debriefs.
09:52████err, the bullying session.
09:54████But I am very excited for this meeting with our CDS partners. Excited for y'all to meet them and hear some of the great work they have been doing and see and share any feedback on some of the plans for the half
10:01████getting error on bjn
10:01████████
10:01████we cannot dial in
10:01████samsies
10:01████Oh no!
10:02████can you pls update with zoom?
10:03████Just updated
10:03████To join the meeting on a computer or mobile phone: https://bluejeans.com/590196758?src=calendarLink
One-Touch: +14087407256,,,,590196758#,,#
Meeting ID: 590196758 To join via phone:
1) 2)
Want to test your video connection?
https://bluejeans.com/111
10:03████I used BJN because the last time we did Zoom people did not like you couldn't join from a portal
10:03████Happy to do either
10:04████still erroring
10:04████works for me now
10:04████https://bluejeans.com/590196758
10:05████Thanks a lot, !
10:05████it booted me out via phone when I joined.
10:05████https://bluejeans.com/590196758
10:05████working now
10:11████is it working for you?
10:11████the phone number did not so I went through app and just turned off video to reduce bandwidth for me. but it is working
10:16████fyi - Core Data Science is a very expensive resource the interview process is more akin to ML/eng interview than a product DS. These are highly capable individuals and they only work on WHAT THEY WANT TO work on (managers don't allocate). We need to keep them excited and committed to IG WB. Challenge will always be relating work to product strategies.
10:19████that latter part, making sure their work informs product, is what has been especially compelling for many of them to work with us since in the past when they've worked on WB work on fb app/central, it wasn't leading to any product changes. So thoughts on what would be helpful in shaping their research to inform product is esp great here.
10:21Michael Yehuda RothschildBTW, this session is exactly what I was envisioning for our "Understand Series"
10:21████this is actually the first of that
10:21████we had been trying to get it scheduled as part of that but eventually just put in in our review slots. I think the idea is to continue to use the review slots moving forward, correct?
10:22Michael Yehuda RothschildIn my old team we'd do these every week (rotating on product area) starting with 15 minutes on health/goals reporting update and then 45 minutes deep dive
10:28Michael Yehuda RothschildYes, I think that would be amazing. I think we said bi weekly ~30 minutes.
10:31████That seems relevant to our discussion about Ecosystem Safety's future :)
10:32████We need to be careful here with messaging to our team, if any :) Technically we don't tell them what to work on either, they choose. And we don't disclose levels, comp, etc.
10:32████(PD on bullying on this meeting) is really excited about this meeting and how it can impact her product thinking. Love this meeting.
10:33████not tied to bullying, but be great to think about how to reduce these types of comments as well...maybe through comment warnings
10:33████Shared an image
10:47████do we ever add strikes against actors using classifiers or can only co do that?
10:48████I imagine the accuracy of classifiers would need to be very high
10:56████We (Facebook Inc) add strikes against actors using classifiers all the time. Many thousands per day.
10:56████CO only does manual enforcement.
10:56████Right, for strikes. But the nice thing about comment warnings (your idea above) is that it's not a strike. It's a nudge to the user, which they can ignore. Which means we can (and do) run the same classifiers at lower precision.
10:57████Love the ideas of warnings and comment mgmt coming from the classifiers!
10:57████That's the best leverage of the in-product nudge-type interventions. They're not enforcement. We've been using the same classifiers in lower precision for comment and post warnings for many months now :)
10:58████I think we launched comment warning (using the classifiers) ~12 months ago :) Nice early bullying win.
10:58████CIFM, of course, complained about potential censorship, nanny state, etc.
10:58████Thankfully we ignored that criticism, launched it, it's worked well, and the following half (2019H2) the
10:59CI bullying and other teams started doing some of the same) This is one of the examples I use in engineer, EM onboarding, of why parity is the right goal but consistency isn't.
10:59████I used to use Daisy as well but we as a company have managed to screw that one up : (
10:59████right - but having classifiers related to user perceptions is slightly different and potentially applicable to a broader user base
11:01████How so?
11:02████The comment warning classifier is literally "some people might find this offensive" and applied to every single user...
11:03████Current classifier is trained on violating labels.. the user perception classifier will beyond violating + low quality. Warning language should likely change
11:04████but I agree @ these are incremental improvements to the existing warnings
11:17Michael Yehuda Rothschilddoes this note answer your question about precise definition of problem maturity?
https://fb.workplace.com/notes/nischal-pathania/content-integrity-problem-team-maturity- plan/383086065949597/
11:17████The bullying comment classifier is trained on a broader policy definition, not the Community Guidelines.
Lots of border-line, aggression, non-violating samples included in training (intentionally).
11:18Michael Yehuda RothschildAnd this is the complementary surface maturity definition: https://fb.workplace.com/notes/sam- england/what-is-the-role-of-a-surface-integrity-team/987352534949961/
11:18████The classifier docs (https://www.internalfb.com/intern/wiki/Instagram_Bullying_Engineering_Team/aggression/Bullying/text_filt er_summary/text_classifier/policies_labeling/) link to the specific policy, with examples.
11:32████did I see on the line?
11:32████Is this the same as s counterpart on CI? or do we have a Vlad on foundation?
11:33████nm, i have my answer
11:36████The org pairings in CI are (eng) <> (product).
reports to pairs with *and* from product, who both report to So it's not exactly 1:1 even in pairs, lol :)
11:36████is the problem team eng lead.
other reports are (CI Foundation) and (Civic).
11:37████What would I do without you diligently setting me straight every time :) ?
11:40████Very sorry to be missing Reels sync; my meeting's running over
11:41████lol not every time, only when you're asking in areas where I already made mistakes :)
11:42████Quote from (who's in the meeting now, awesome guy) last time I asked him about CI org structure:
"honestly, I wish we'd called it CI MPK, CI LON, CI SEA. Nothing else passes the sanity bar when you look at the current org structure." l01
11:42████crap... so there are many more areas where I need correction who is taking care of those?
11:42████what specifically are we doing for child grooming (something I just heard about that is happening a lot on TikTok)?
11:43████Somewhere between zero and negligible. Child safety is an explicit non-goal this half. I'd argue we're making it worse with Interop, but's that's a can of worms.
11:43████CI has the usual CEI and IIC teams working to improve classifiers. I'll share their roadmaps.
11:43████So we'll get some benefits without doing any work, hopefully.
11:45████Here you go,
https://fb.quip.com/Cj8aAejuMsD7.
11:46████Get IG to 2-measured for "tier 1" IIC. Nice win if we get it.
11:46████We've punted on our teams investing in these areas for awhile as part of prioritization and because we have central teams working on these that as Yoav said, we hope* we will benefit from. But, several of the extreme harms areas have always seemed like something we could do more to focus on and support/partner w/ central given IG challenges specifically. But, it's always a question of ruthless prioritization and partnership models.
11:47████I'm particularly thinking of SSI, CEI, IIC...
11:49████I gotta say part of the thing that makes me most anxious about Reels is that it gives me a little PTSD around when we launched TBO (Explore v2) and Trending hashtags at the top of Explore.
11:50████thankfully, our integrity processes have matured a lot since then, but it was a mess having so much bad content very quickly getting crazy distribution on the top of explore for everyone.
11:51████non-rec measurement is actually in place for certain problem areas
11:51████guardrails have been set
11:51████we can argue however whether that's good enough
12:06████I flagged to Comm. Leads that the Reels call is running late
12:06████We are waiting for all to join
12:06████It's great that you ran over because we JUST finished with Growth. If anyone asks, we were not late either.
12:54████did you want to touch on stress tests
12:56████Not more thsan these topics. Thanks for asking :)
13:09████WB Design is interested in implementing this program for making sure our craft bar remains high.
Would love to get your thoughts if this is something you think engineers would be supportive of?
Katherine is going to head this up for the design team and will have designers start the conversations with their eng partners. https://fb.workplace.com/groups/IGDSFYI/permalink/705571776684434/
13:55████[pages 7–11 fully redacted in filing]