Twitter | Search | |
Guy Rosen
VP Integrity @ Facebook, overseeing FB's work on safety & integrity
76
Tweets
694
Following
4,663
Followers
Tweets
Guy Rosen Jul 17
Calling all Engineers and Data Scientists! FB is hosting our 4th annual child safety hackathon on Aug 1+2, together with our partners, and . Sign up here:
Reply Retweet Like
Guy Rosen May 25
Replying to @donie @ngleicher
It's standard for content marked false by a fact checker. Not new.
Reply Retweet Like
Guy Rosen May 23
Replying to @TonyRomm
The goal of this work to date has been to evaluate the *methodology* e.g. metrics, sampling, labeling, statistics. We want to validate more but also need to balance with protecting private info. We're exploring ways to make anonymized/aggregated data accessible to researchers.
Reply Retweet Like
Guy Rosen May 23
Replying to @TonyRomm
"never got to question FB workers" is incorrect. The report says the group "talked to a large number of engineers who design and oversee content enforcement and measurement systems" – including many of the engineers, product managers, data scientists etc on my team (and me too)
Reply Retweet Like
Guy Rosen May 23
Replying to @daveyalba
Correct! That's exactly why there are error bars on it (e.g. 0.23-0.25% prevalence on violence and graphic content). For the math geeks out there, these are the confidence intervals at 95% statistical significance.
Reply Retweet Like
Guy Rosen May 23
Replying to @daveyalba @mark_riedl
Proactive detection is the % of the "content actioned" number we publish (i.e. the stuff we *caught*). Prevalence of violating content (i.e. stuff we missed) is measured on all views of content on FB (thru sampling).
Reply Retweet Like
Guy Rosen May 23
Replying to @daveyalba
We posted a longer guide and a video here:
Reply Retweet Like
Guy Rosen May 23
Replying to @daveyalba
It's a great question: we sample content that people view on FB, and then manually review how much of it violates our policies. We can then statistically extrapolate the overall % of views were on violating content. That tells us what we *missed*.
Reply Retweet Like
Guy Rosen May 23
Replying to @mark_riedl @daveyalba
Amount of stuff we *missed* is measured in prevalence. More here:
Reply Retweet Like
Guy Rosen May 23
Replying to @daveyalba
The proactive rate is not an estimate, we count how much of the content we removed was detected by our systems before someone reported it to us.
Reply Retweet Like
Guy Rosen retweeted
Kurt Wagner May 23
Replying to @KurtWagner8
If you care about how FB detects and categorizes fake accts, here's a blog post they published this morning explaining some of that.
Reply Retweet Like
Guy Rosen May 23
As we've ramped up work on safety, we know it's not enough to just do the work, we also have to be open about it. Today we're publishing our 3rd community standards enforcement report, which includes metrics about how we're doing at enforcing our policies
Reply Retweet Like
Guy Rosen May 18
Excellent reporting. Reminds me of this public health inspired approach written up by - need digital equivalent to inoculation and hand-washing
Reply Retweet Like
Guy Rosen May 16
Replying to @DavidClinchNews
Yes, we've done these workshops so far in India, Singapore, Kenya, now Mexico City. US & Europe are next. There've been roundtable sessions and smaller forums in additional locations, and a public consultation process is here too:
Reply Retweet Like
Guy Rosen May 16
Just wrapped up 2 days in Mexico City with 50 experts from across 9 countries where we discussed the design and structure of Facebook’s Oversight Board. Was fascinating to hear regional perspectives on content and free expression, that will be incorporated into the design.
Reply Retweet Like
Guy Rosen May 16
We've had interest in keeping our public consultation process open longer. New deadline: Friday May 25th!
Reply Retweet Like
Guy Rosen May 15
congrats!
Reply Retweet Like
Guy Rosen May 13
Congrats!
Reply Retweet Like
Guy Rosen retweeted
SwiftOnSecurity May 9
Technically, my brain is something I have AND something I am, so a password I know is technically multi-factor authentication.
Reply Retweet Like
Guy Rosen May 7
Replying to @Pinboard
More on the sampling methodology here (around p30+)
Reply Retweet Like