WEBVTT - Facebook Papers: How Their Algorithm Spread Anger and Misinformation

0:00:00.080 --> 0:00:04.280
<v Speaker 1>It's Thursday, October. I'm Oscar Ramiras from the Daily Dive

0:00:04.360 --> 0:00:08.479
<v Speaker 1>podcast in Los Angeles, and this is reopening America. The

0:00:08.520 --> 0:00:11.080
<v Speaker 1>Facebook papers are out and it's shedding more light on

0:00:11.119 --> 0:00:14.040
<v Speaker 1>how executives their weight, various trade offs between their bottom

0:00:14.080 --> 0:00:17.520
<v Speaker 1>line and impacts on public safety. In one instance, it

0:00:17.560 --> 0:00:19.960
<v Speaker 1>took them years to implement a fix for an algorithm

0:00:20.079 --> 0:00:23.760
<v Speaker 1>that was feeding people angry emotional content filled with misinformation.

0:00:24.239 --> 0:00:26.680
<v Speaker 1>If you use the angry emoji on a post instead

0:00:26.680 --> 0:00:29.480
<v Speaker 1>of the like button, it carried more weight, and then

0:00:29.640 --> 0:00:31.720
<v Speaker 1>they would feed you more of the same despite the

0:00:31.760 --> 0:00:35.240
<v Speaker 1>signal that you did not like it. Jeremy Merrill, Data

0:00:35.280 --> 0:00:38.160
<v Speaker 1>report at the Washington Post joins us for more. Thanks

0:00:38.159 --> 0:00:40.600
<v Speaker 1>for joining us, Jeremy, thanks for having me on it.

0:00:41.080 --> 0:00:43.440
<v Speaker 1>Let's talk a little bit about the Facebook papers. We're

0:00:43.520 --> 0:00:47.239
<v Speaker 1>learning a lot more about kind of how things were

0:00:47.240 --> 0:00:49.800
<v Speaker 1>working in the background at Facebook. You know a lot

0:00:49.800 --> 0:00:54.480
<v Speaker 1>of different things, internal dialogues between between workers, and how

0:00:54.520 --> 0:00:56.800
<v Speaker 1>things were being portrayed out there. One of the articles

0:00:56.800 --> 0:01:01.040
<v Speaker 1>that you wrote about was about this, uh, these anger emojis,

0:01:01.080 --> 0:01:03.880
<v Speaker 1>these new set of emojis that popped up in seen

0:01:04.319 --> 0:01:08.840
<v Speaker 1>that lets people react to comments beyond the the just

0:01:08.880 --> 0:01:11.600
<v Speaker 1>the like button. Um, you know, you can laugh at something,

0:01:11.680 --> 0:01:15.200
<v Speaker 1>heart something. You know, there's like a sad emoji or

0:01:15.200 --> 0:01:17.360
<v Speaker 1>a sorry emoji I think they called it, uh, and

0:01:17.400 --> 0:01:20.400
<v Speaker 1>then the the angry emoji also. But what happened with

0:01:20.440 --> 0:01:24.040
<v Speaker 1>all of this is uh, things were waited differently, and

0:01:24.080 --> 0:01:26.400
<v Speaker 1>these emojis were waited in such a way that it

0:01:26.480 --> 0:01:29.880
<v Speaker 1>started promoting a lot of misinformation, a lot of things

0:01:29.920 --> 0:01:33.000
<v Speaker 1>that would give you more of an emotional response. And

0:01:33.319 --> 0:01:35.120
<v Speaker 1>at the time they thought that this was the way

0:01:35.160 --> 0:01:37.160
<v Speaker 1>to go, but it just kind of led to all

0:01:37.240 --> 0:01:39.120
<v Speaker 1>sorts of mishaps and it took them a long time

0:01:39.160 --> 0:01:42.400
<v Speaker 1>to really remedy that as well. So Jeremy, help us

0:01:42.400 --> 0:01:43.920
<v Speaker 1>walk through some of this. What are we seeing with it?

0:01:44.720 --> 0:01:47.040
<v Speaker 1>That's exactly right. So what we're talking about here is

0:01:47.040 --> 0:01:49.680
<v Speaker 1>whenever you go to Facebook dot com or open the

0:01:49.720 --> 0:01:53.120
<v Speaker 1>Facebook app on your phone, you get your news feeds, right,

0:01:53.600 --> 0:01:57.120
<v Speaker 1>and just just imagine all of the possible things that

0:01:57.280 --> 0:02:00.400
<v Speaker 1>Facebook could show you. At the very top could be

0:02:00.480 --> 0:02:02.720
<v Speaker 1>posts from your friends a couple of days ago. It

0:02:02.800 --> 0:02:05.520
<v Speaker 1>could be a baby picture from someone you wrote the

0:02:05.520 --> 0:02:09.000
<v Speaker 1>school bus with a high school fifteen years ago. It

0:02:09.080 --> 0:02:11.480
<v Speaker 1>could be posts from a news organization or a buy

0:02:11.520 --> 0:02:14.400
<v Speaker 1>nothing group in your town. There's all sorts of things

0:02:14.400 --> 0:02:17.480
<v Speaker 1>and Facebook has to make a decision of what you

0:02:17.639 --> 0:02:19.400
<v Speaker 1>go first and what you goes back in what should

0:02:19.440 --> 0:02:22.880
<v Speaker 1>go you know three thousands which that you'll never see it,

0:02:23.440 --> 0:02:27.400
<v Speaker 1>And that algorithm is sort of one of the secret

0:02:27.440 --> 0:02:30.960
<v Speaker 1>sauces that that keeps people, that keeps Facebook, um sort

0:02:31.000 --> 0:02:36.760
<v Speaker 1>of a life. And broadly speaking, um, there are two

0:02:36.760 --> 0:02:39.079
<v Speaker 1>sets of things that make up that algorithm. One is

0:02:39.080 --> 0:02:43.040
<v Speaker 1>a really complicated, really masthei for what you think of

0:02:43.080 --> 0:02:45.000
<v Speaker 1>when you think of an algorithm that it's super complicated.

0:02:45.360 --> 0:02:48.880
<v Speaker 1>That's meant to predict what's the likelihood that if we

0:02:48.960 --> 0:02:51.919
<v Speaker 1>show this post to Jeremy, what's the likelihood that he's

0:02:51.919 --> 0:02:54.920
<v Speaker 1>going to click like, what's the likelihood that he's going

0:02:54.960 --> 0:02:58.160
<v Speaker 1>to click angry, what's the likelihood that he's going to

0:02:58.200 --> 0:03:00.639
<v Speaker 1>write a comment? Um? So that's the first part. Are

0:03:00.680 --> 0:03:04.359
<v Speaker 1>all those predictions, and the second part is really simple.

0:03:04.760 --> 0:03:08.200
<v Speaker 1>It's just numbers that are picks by human beings who

0:03:08.240 --> 0:03:13.560
<v Speaker 1>worked in Facebook, and those numbers are how important relatively

0:03:14.280 --> 0:03:19.919
<v Speaker 1>those different predictions are. And when Facebook launched a new

0:03:19.960 --> 0:03:22.240
<v Speaker 1>way of ranking your news speed, a new way of

0:03:22.240 --> 0:03:25.680
<v Speaker 1>picking what stuff shows up first. Those weights were that

0:03:25.760 --> 0:03:30.000
<v Speaker 1>alike is worth one point. These reaction emojis, including angry,

0:03:30.040 --> 0:03:32.760
<v Speaker 1>but also all the other ones were worth five points,

0:03:32.919 --> 0:03:35.720
<v Speaker 1>and comments were worth between fifteen and thirty points, and

0:03:35.760 --> 0:03:38.480
<v Speaker 1>there's some other a couple other ingredients as well. But

0:03:38.640 --> 0:03:41.640
<v Speaker 1>sort of the minute that this launched, someone from Facebook

0:03:41.680 --> 0:03:44.360
<v Speaker 1>wrote on their internal message board and said, you know

0:03:44.440 --> 0:03:48.760
<v Speaker 1>that they were worried that posts that make people react

0:03:48.880 --> 0:03:54.960
<v Speaker 1>with the angry emoji might inadvertently open the door to spam, abuse, clickbait,

0:03:55.760 --> 0:03:59.560
<v Speaker 1>and somebody else said, you know it's possible. Yeah, I

0:03:59.560 --> 0:04:02.280
<v Speaker 1>mean it makes natural sense, right, something that angers you,

0:04:02.440 --> 0:04:04.800
<v Speaker 1>something that you're like, well, you know there that's full

0:04:04.800 --> 0:04:07.800
<v Speaker 1>of crap. You know, I'm gonna anger emoji that hopefully

0:04:07.800 --> 0:04:09.840
<v Speaker 1>you don't get any more of it. But it was

0:04:09.840 --> 0:04:13.280
<v Speaker 1>actually the opposite, because they waited it so much more

0:04:13.360 --> 0:04:16.839
<v Speaker 1>than just the like or whatever. Uh, the the algorithm

0:04:16.839 --> 0:04:19.480
<v Speaker 1>would actually just serve you up more. And and in

0:04:19.520 --> 0:04:23.919
<v Speaker 1>that sense, you know, the worst of the platform started, uh,

0:04:24.120 --> 0:04:27.960
<v Speaker 1>started shining through even more for people's news feeds. That's

0:04:28.080 --> 0:04:30.599
<v Speaker 1>that's exactly right. If the algorithm learns that you're going

0:04:30.640 --> 0:04:33.479
<v Speaker 1>to react annoying, it's worth five points, then it's going

0:04:33.520 --> 0:04:35.159
<v Speaker 1>to show you more stuff like that, because that's what

0:04:35.200 --> 0:04:39.160
<v Speaker 1>it considers. The Facebook's term was meaningful social interaction. They

0:04:39.160 --> 0:04:41.440
<v Speaker 1>consider that to be a meaningful social interaction, so they're

0:04:41.440 --> 0:04:44.559
<v Speaker 1>gonna try to do more of it. But gradually, um

0:04:44.800 --> 0:04:48.400
<v Speaker 1>evidence started building up. People found that things that sparked

0:04:48.400 --> 0:04:54.520
<v Speaker 1>a lot of angry angry reactions were disproportionately misinformation what

0:04:54.600 --> 0:04:58.679
<v Speaker 1>places Book called toxicity, as well as low quality news,

0:04:59.160 --> 0:05:03.640
<v Speaker 1>and people started proposing, hey, maybe this isn't a great

0:05:03.680 --> 0:05:07.920
<v Speaker 1>idea to count the anger emojis so high, and they

0:05:08.200 --> 0:05:11.600
<v Speaker 1>counted as exactly the same as these other these other

0:05:11.600 --> 0:05:15.640
<v Speaker 1>emojis like love or care. Yeah, and that's part of

0:05:15.680 --> 0:05:18.760
<v Speaker 1>what that's part of what's interesting about the Facebook papers

0:05:18.760 --> 0:05:21.000
<v Speaker 1>and what we're kind of learning about how Facebook was

0:05:21.040 --> 0:05:24.400
<v Speaker 1>operating is that while these experiments were going on, right

0:05:24.480 --> 0:05:26.320
<v Speaker 1>how the different weights and how to serve up the

0:05:26.320 --> 0:05:31.000
<v Speaker 1>algorithm best to people, the internal conversations from employees calling

0:05:31.040 --> 0:05:33.040
<v Speaker 1>some of it out, saying, hey, this is a problem,

0:05:33.120 --> 0:05:35.720
<v Speaker 1>this is becoming a problem. This problem is too big,

0:05:36.080 --> 0:05:39.160
<v Speaker 1>and uh, you know, really not being able to remedy

0:05:39.160 --> 0:05:41.760
<v Speaker 1>it right away. Uh. That's one of the other interesting

0:05:41.760 --> 0:05:43.680
<v Speaker 1>aspects about it, because that's part of what is We're

0:05:43.720 --> 0:05:47.560
<v Speaker 1>we're seeing these internal dialogues happening behind the scenes, now right,

0:05:47.600 --> 0:05:49.960
<v Speaker 1>That's that's exactly right, And you know, to be fair

0:05:50.040 --> 0:05:53.120
<v Speaker 1>to Facebook, Um, in some cases people didn't have evidence

0:05:53.160 --> 0:05:56.000
<v Speaker 1>yet it was just a suspicious sort of an instinct. Hey,

0:05:56.360 --> 0:05:58.960
<v Speaker 1>ranking anger so highly doesn't seem like a great idea,

0:05:59.560 --> 0:06:02.840
<v Speaker 1>or it would be kind of weird, um to make

0:06:02.920 --> 0:06:05.599
<v Speaker 1>distinctions between different kinds of emotions. They felt kind of

0:06:05.640 --> 0:06:09.320
<v Speaker 1>uncomfortable with that, as well as raising the possibility that

0:06:09.920 --> 0:06:13.400
<v Speaker 1>stuff that makes people angry angry might be important for

0:06:13.480 --> 0:06:17.680
<v Speaker 1>social change. And um, but you know, eventually, this book

0:06:17.720 --> 0:06:20.280
<v Speaker 1>did end up cutting the weight of this emoji all

0:06:20.279 --> 0:06:22.920
<v Speaker 1>the way down to zero. They ended up making distinctions

0:06:22.960 --> 0:06:25.840
<v Speaker 1>between emotions. Um, So for a while, love and care

0:06:25.920 --> 0:06:29.719
<v Speaker 1>were were double alike, and stuff that they maybe go

0:06:29.839 --> 0:06:33.120
<v Speaker 1>ha ha or wow was only worth the same as alike.

0:06:33.560 --> 0:06:36.560
<v Speaker 1>So to their credit, they did end up making this

0:06:36.760 --> 0:06:41.760
<v Speaker 1>fix eventually, But it was the case that they set

0:06:41.800 --> 0:06:44.680
<v Speaker 1>these without They set the weights alto five times alike

0:06:45.080 --> 0:06:48.440
<v Speaker 1>without looking into what would happen, And then once they've

0:06:48.480 --> 0:06:50.880
<v Speaker 1>made that choice, they stuck with it for a couple

0:06:50.920 --> 0:06:53.400
<v Speaker 1>of years until the weight of the evidence was so

0:06:53.560 --> 0:06:56.560
<v Speaker 1>much that they finally had to act. Yeah, and sticking

0:06:56.560 --> 0:06:58.120
<v Speaker 1>with it too, right, And you know, it's a it's

0:06:58.120 --> 0:07:01.120
<v Speaker 1>a new feature on the website, so it makes sense

0:07:01.120 --> 0:07:02.800
<v Speaker 1>that they want to point people to it as much

0:07:02.839 --> 0:07:05.080
<v Speaker 1>as possible. So that was part of the problem too.

0:07:05.120 --> 0:07:07.880
<v Speaker 1>Is that why they didn't act so quickly, is that

0:07:07.920 --> 0:07:10.360
<v Speaker 1>you know, they wanted people to start using that new feature,

0:07:10.800 --> 0:07:12.800
<v Speaker 1>um thinking you know, it was going to be that

0:07:13.080 --> 0:07:15.800
<v Speaker 1>next great thing for them. Uh, And so that posed

0:07:15.840 --> 0:07:18.240
<v Speaker 1>the problem too. And and as you mentioned once, they

0:07:18.360 --> 0:07:21.880
<v Speaker 1>um did kind of start reducing the weights of these things.

0:07:22.320 --> 0:07:24.440
<v Speaker 1>They found that, you know, in other parts of the

0:07:24.440 --> 0:07:27.240
<v Speaker 1>algorithm also, you know, when a post would get too

0:07:27.240 --> 0:07:31.560
<v Speaker 1>popular or something, and they would demote certain things. Even then,

0:07:31.600 --> 0:07:33.760
<v Speaker 1>it was it was tough because uh, you know, there

0:07:33.800 --> 0:07:36.560
<v Speaker 1>was no cap on how many points let's say a

0:07:36.600 --> 0:07:39.000
<v Speaker 1>post could get, so um, you know, even if you

0:07:39.080 --> 0:07:42.120
<v Speaker 1>demote something, I mean, it still could have a billion points, right,

0:07:42.360 --> 0:07:44.880
<v Speaker 1>and it could be at the top of everybody news feed.

0:07:45.880 --> 0:07:50.520
<v Speaker 1>That's exactly right. It's really difficult for us to understand

0:07:51.200 --> 0:07:54.600
<v Speaker 1>how all the sort of weird mass that goes into

0:07:54.640 --> 0:07:56.560
<v Speaker 1>stay going out, what goes to the top of your

0:07:56.600 --> 0:07:59.400
<v Speaker 1>news feed. And some of the documents showed that even

0:07:59.520 --> 0:08:01.800
<v Speaker 1>these people with access to a lot of resources that

0:08:01.840 --> 0:08:05.120
<v Speaker 1>Facebook and a lot of people who can understand it offensibly,

0:08:05.440 --> 0:08:08.520
<v Speaker 1>even they didn't really understand how the algorithm worked, and

0:08:08.560 --> 0:08:11.000
<v Speaker 1>they were finding that, you know, the scores for the

0:08:11.080 --> 0:08:13.560
<v Speaker 1>things that would appear at the top of your news

0:08:13.560 --> 0:08:16.840
<v Speaker 1>speed usually are about a hundred, sorry, in the hundreds,

0:08:17.120 --> 0:08:21.480
<v Speaker 1>a couple hundred, but some some of the post could

0:08:21.480 --> 0:08:25.000
<v Speaker 1>get scores in the millions up to a billion. And

0:08:25.400 --> 0:08:28.960
<v Speaker 1>what what you've got is a sort of battle between uh,

0:08:29.000 --> 0:08:32.439
<v Speaker 1>the integrity teams, who are the ones tasked with finding

0:08:32.480 --> 0:08:36.000
<v Speaker 1>stuff that's probably bad and trying to push it down

0:08:36.480 --> 0:08:39.520
<v Speaker 1>in people's feeds so it doesn't go viral, and the

0:08:39.600 --> 0:08:43.360
<v Speaker 1>growth teams, who are tasked with getting people to use Facebook.

0:08:43.800 --> 0:08:45.240
<v Speaker 1>And so you should have had this battle where the

0:08:45.240 --> 0:08:48.840
<v Speaker 1>growth team was giving these things these humongous scores, but

0:08:48.920 --> 0:08:51.440
<v Speaker 1>the integrity team was limited and like the best that

0:08:51.480 --> 0:08:53.880
<v Speaker 1>they could do is like cut it the score it half.

0:08:54.200 --> 0:08:56.640
<v Speaker 1>So if you you know, the post number one has

0:08:56.720 --> 0:08:59.840
<v Speaker 1>you know, five million as its score and post number

0:08:59.840 --> 0:09:03.120
<v Speaker 1>two has eight hundred, you're trying to push post number

0:09:03.120 --> 0:09:05.439
<v Speaker 1>one down because because it's probably garbage and all you

0:09:05.480 --> 0:09:07.880
<v Speaker 1>can do is divide the five million and a half, well,

0:09:08.000 --> 0:09:10.640
<v Speaker 1>two and a half million is still huge. So they

0:09:10.640 --> 0:09:15.040
<v Speaker 1>found that they had this trouble of actually making their

0:09:15.120 --> 0:09:17.960
<v Speaker 1>principled ways of trying to make the platform safer and

0:09:18.600 --> 0:09:21.960
<v Speaker 1>more enjoyable for people. They weren't able to work because

0:09:22.000 --> 0:09:24.959
<v Speaker 1>the math was so complicated. They were always fighting that

0:09:25.000 --> 0:09:28.280
<v Speaker 1>losing battle. And you know, it's it's interesting, you know,

0:09:28.360 --> 0:09:31.760
<v Speaker 1>to the point, right, it's why is Facebook serving up

0:09:31.800 --> 0:09:33.840
<v Speaker 1>so much crap? Why is it serving up so much

0:09:33.880 --> 0:09:36.679
<v Speaker 1>anger and divisiveness? And this is exactly why it is.

0:09:36.760 --> 0:09:39.200
<v Speaker 1>The different weights that were put on things, and the

0:09:39.200 --> 0:09:42.120
<v Speaker 1>math was so complicated it was hard to go back

0:09:42.320 --> 0:09:45.000
<v Speaker 1>and fix it retroactively. So you know, we're learning a

0:09:45.000 --> 0:09:48.560
<v Speaker 1>lot from these Facebook papers. Um, you know, Facebook choosing

0:09:48.559 --> 0:09:52.800
<v Speaker 1>maximum engagement over user safety, dropping the ball in different areas.

0:09:52.840 --> 0:09:55.720
<v Speaker 1>So we'll keep finding out more and uh and bring

0:09:55.720 --> 0:09:58.720
<v Speaker 1>it all to you. Jeremy Merrill data reported at the

0:09:58.720 --> 0:10:02.079
<v Speaker 1>Washington Post. Thank very much for joining us, Thanks for

0:10:02.160 --> 0:10:06.480
<v Speaker 1>having me on. I'm Oscar Roomiers and this has been

0:10:06.520 --> 0:10:10.240
<v Speaker 1>reopening America. Don't forget effort. Today's big news stories you

0:10:10.240 --> 0:10:12.400
<v Speaker 1>can check me out on the Daily Dive podcast every

0:10:12.440 --> 0:10:14.880
<v Speaker 1>Monday through Friday, so follow us on I Heart Radio

0:10:15.160 --> 0:10:16.560
<v Speaker 1>or wherever you get your podcast