Meta & Children's Online SafetyChat: we really need to take a very close look at what sessions (metric) is doing to vulnerable sub populations
Thread View
████ name redacted. Different bars may be the same person or different people.
quoted in an excerpt in the archive.
Sunday 27 September 2020
19:30Michael Yehuda Rothschild#quiet just got around to it. Great note! Do you know if there's any role models for this type of work across the industry? Which company's AI strategy is most aligned with human value and how. I have my own opinion but curious to hear your POV
22:50████manks! I am a part of an external working group called Partnership on AI from RAI is on the board or some such leadership position for PAI. So AI value alignment is a wett researched topic among external AI experts (the panel I am on is along with Google AI engineers, AI ethics profs from UC SF etc.
it's a good group). My 2 cents is we are a bit behind. I reached out to folks in our AI org and RAI org but there doesn't seem to be a systematic set up within the company. Several leads across the company are interested in this space though ( for example) but I am wondering how we might make this more systematic. This is not like a 6m or 1 year project but a mental shift and like a 2-5yr investment to evolve AI to better human outcomes.
22:50████understandably, it is too long term to easily generate a ton of excitement easily among engg.
22:51████Jonn is pretty supportive of this (as you can see by his comments)..
22:52████But hard, long term initiative...
22:52Michael Yehuda RothschildTotally agree. FWIW, I think we are systematically in a much worse place here compared to for eg Google Search
22:52████yes, tend to agree
22:52████BUL we are so short term focused on sessions, MSI and so on...
22:52Michael Yehuda RothschildThe reason being that most of our AI systems are using behavioral user implcit signals vs Google which uses objective human ratings
22:53████I recently talked to an ex Googler, Google Travel apparently had 2 top line metrics: sentiment (user satisfaction thru' surveys) and DAP.
22:53Michael Yehuda RothschildSo while we think we are optimizing for signals that correlate with good things we often don't actually know all the patterns that take place
22:54████yes...I think surveys are not perfect but closer to the ground truth than behavioral signals but a ton of challenges when it comes to survey based modeling and measurements
22:54Michael Yehuda RothschildI'm a pretty big fan of what Ads team on IG do
22:54Michael Yehuda RothschildI'll send you a deck
22:54████so folks probably don't want to hear this as the solution but feel like we need some sort of internal AI ethics panel...that pushes the boundary on value alignment and AI safety
22:55Michael Yehuda RothschildOOPS if you've heard about that
22:55████mm no, I don't think am familiar
22:55████prease share.
22:55Michael Yehuda Rothschild(y)
22:56████(also it's a modest plan but we finally are aligning on this feed fairness plan:
https://fb.quip.com/BkeXALXAMDYL)
22:56████II you have further ideas on AI value alignment, pls do share. I shared my note with fwiw.
22:56Michael Yehuda Rothschild(y)
22:58Michael Yehuda RothschildI just watched the social dilema yesterday and it made me sad b/c while I don't agree with a lot of what they say there's some truth to it. Have you watched it?
22:58████Yes the day it came out.
22:58████so the 3 men in the suits and the human puppet is a bit much but I agree with the core points (not the way it is sensationalized): like there seems to be a problem with our business model and definitely a problem with our metrics.
22:59████we really need to take a very close look at what sessions (metric) is doing to vulnerable sub populations for example: teens, people with mental health issues, sensitive geographies.
22:59Michael Yehuda RothschildYes
22:59████Within feed, we are attempting to operationalize a sentiment based metric...
23:00████put survey based metrics don't have a level playing ground with behavioral metrics like sessions and MSI.
23:00Michael Yehuda RothschildAnd into "rabbit holes" caused by our collaborative filters loops
23:00████our metrics are partly to blame and this ties to value alignment. You give the AI system a metric. It assumes that IS the ground truth and intent and will keep on optimizing until you build in strong guardrails.
23:01Michael Yehuda RothschildI once got enticed by some snake video in watch, now all my recommendations occasionally have snake videos show up. It's terrible
23:01████So if a person vulnerable to conspiracy theories will watch more of those videos and comment more and share more, we will keep showing them more
23:01████ugn
23:01████pecause it gives us more MSI and sessions.
23:01Michael Yehuda Rothschild50, BTW, I think it might be simpler to rely more on product design and controls, and reduce complexity of our AI systems
23:01████and Humane tech have a podcast, don't know if you are familiar...
23:02Michael Yehuda RothschildYeah, I am
23:02████I once obsessively heard all the podcasts and compiled notes here: https://fb.quip.com/aSTeAQkYam8f
23:02████It's bad.
23:02████(we hosted for a workshop, wasn't great)
23:03████imis is my tldr; one slide compilation
23:03████Shared an image