WEBVTT - The Blind Spots in Facebook's Ad Machine

0:00:00.560 --> 0:00:05.439
<v Speaker 1>Earlier this year, when Facebook CEO Mark Zuckerberg testified in Washington,

0:00:05.519 --> 0:00:08.880
<v Speaker 1>d C. To defend his company, it became clear that

0:00:09.039 --> 0:00:13.680
<v Speaker 1>senators didn't really understand what his business model was. In

0:00:13.800 --> 0:00:15.520
<v Speaker 1>order to do that, we believe that we need to

0:00:15.520 --> 0:00:18.360
<v Speaker 1>offer a service that everyone can afford, and we're committed

0:00:18.400 --> 0:00:20.640
<v Speaker 1>to doing that. Well, if so, how do you sustain

0:00:20.680 --> 0:00:26.400
<v Speaker 1>a business model in which users don't pay for your service? Senator?

0:00:26.400 --> 0:00:34.680
<v Speaker 1>We run ads? Nice see in Silicon Valley. The episode

0:00:34.760 --> 0:00:38.239
<v Speaker 1>was seen as embarrassing for senators. How could they be

0:00:38.280 --> 0:00:42.120
<v Speaker 1>so clueless? A lot of people know Facebook's business model

0:00:42.240 --> 0:00:45.960
<v Speaker 1>is advertising, but they may not know exactly how it works.

0:00:46.400 --> 0:00:50.519
<v Speaker 1>In for example, when Facebook disclosed that Russia purchased a

0:00:50.600 --> 0:00:55.760
<v Speaker 1>hundred thousand dollars and ads on Facebook to manipulate US politics,

0:00:56.360 --> 0:00:58.720
<v Speaker 1>I got this question from a lot of smart people.

0:00:59.320 --> 0:01:03.080
<v Speaker 1>Who was the Facebook salesperson that sold Russia those ads?

0:01:03.160 --> 0:01:06.200
<v Speaker 1>Can we find them? That sounds like a pretty reasonable question,

0:01:06.720 --> 0:01:10.520
<v Speaker 1>But when Facebook employees here that story, they immediately laugh

0:01:11.040 --> 0:01:14.560
<v Speaker 1>because Facebook doesn't have sales people directly selling ads. In

0:01:14.640 --> 0:01:18.560
<v Speaker 1>most cases, it all happens through a website. Anyone with

0:01:18.600 --> 0:01:20.360
<v Speaker 1>a credit card can do it, and you don't even

0:01:20.440 --> 0:01:23.280
<v Speaker 1>have to be in contact with Facebook at all, just

0:01:23.400 --> 0:01:25.680
<v Speaker 1>the Facebook website. You don't even have to be who

0:01:25.760 --> 0:01:29.040
<v Speaker 1>you say you are. So much has happened since that revelation,

0:01:29.280 --> 0:01:33.440
<v Speaker 1>so many changes to Facebook's advertising system, so many apologies,

0:01:33.840 --> 0:01:37.360
<v Speaker 1>so many more questions from Congress, And even that didn't

0:01:37.360 --> 0:01:39.800
<v Speaker 1>stop Facebook from having its share of crises in the

0:01:39.840 --> 0:01:42.480
<v Speaker 1>last election cycle as I tried to reform the system.

0:01:42.520 --> 0:01:45.319
<v Speaker 1>But if there's anything this era of Facebook News has proven,

0:01:45.600 --> 0:01:48.680
<v Speaker 1>it's that sometimes it's worth it to try to understand

0:01:48.760 --> 0:02:02.400
<v Speaker 1>the basis by I'm Brad Stone and I'm Sarah Fryar,

0:02:02.560 --> 0:02:05.040
<v Speaker 1>and this week on Decrypted, we're going to try to

0:02:05.080 --> 0:02:09.200
<v Speaker 1>explain how Facebook's advertising system works so we can understand

0:02:09.200 --> 0:02:12.160
<v Speaker 1>why some people love it so much and others think

0:02:12.200 --> 0:02:15.639
<v Speaker 1>it could be so dangerous. Facebook's ad machine is easy

0:02:15.760 --> 0:02:19.120
<v Speaker 1>and quick to use, and that's made it incredibly profitable.

0:02:19.639 --> 0:02:22.160
<v Speaker 1>Wall Street is expecting it to bring in fifty five

0:02:22.480 --> 0:02:25.960
<v Speaker 1>billion in revenue this year. But the same algorithms that

0:02:26.040 --> 0:02:30.000
<v Speaker 1>make it so powerful can also create problems. For example,

0:02:30.240 --> 0:02:33.440
<v Speaker 1>if enough people say on their profile that they hate Jews,

0:02:34.040 --> 0:02:38.040
<v Speaker 1>Facebook software would make a category all by itself to

0:02:38.160 --> 0:02:42.240
<v Speaker 1>let advertisers send their contents specifically to do hitters, and

0:02:42.360 --> 0:02:45.400
<v Speaker 1>unfortunately that's not even a hypothetical. That's one of the

0:02:45.440 --> 0:02:48.040
<v Speaker 1>problems Facebook had to solve, along with a lot of

0:02:48.080 --> 0:02:51.359
<v Speaker 1>other things. But fixing oversight for the ad system isn't

0:02:51.400 --> 0:02:53.680
<v Speaker 1>as easy as you might think, and some of those

0:02:53.760 --> 0:02:57.760
<v Speaker 1>updates have had unintended consequences of their own stay with us.

0:03:04.280 --> 0:03:07.440
<v Speaker 1>So Sarah back to that congressional hearing in April, the

0:03:07.520 --> 0:03:11.280
<v Speaker 1>senators asked Mark Zuckerberg about whether Facebook sales user data

0:03:11.400 --> 0:03:15.080
<v Speaker 1>to advertisers. Um, have you ever drawn the line on

0:03:15.120 --> 0:03:19.240
<v Speaker 1>selling data to an advertiser? Yes, Senator, we don't sell

0:03:19.320 --> 0:03:22.519
<v Speaker 1>data at all. So the way the add system works

0:03:22.639 --> 0:03:25.440
<v Speaker 1>is advertisers can come to us. The company hates it

0:03:25.480 --> 0:03:29.080
<v Speaker 1>when people think that because it doesn't sell data to advertisers.

0:03:29.120 --> 0:03:31.919
<v Speaker 1>What it sells is access to certain kinds of people

0:03:32.400 --> 0:03:34.560
<v Speaker 1>based on the information they share on face and then

0:03:34.600 --> 0:03:36.600
<v Speaker 1>we help them get that message in front of people.

0:03:36.960 --> 0:03:41.720
<v Speaker 1>But this is one of the it's widely mischaracterized about

0:03:41.720 --> 0:03:45.360
<v Speaker 1>our system that we sell data, and it's actually everything

0:03:45.400 --> 0:03:48.080
<v Speaker 1>you share on Facebook, all the articles you decide to

0:03:48.080 --> 0:03:50.720
<v Speaker 1>click on or decide not to click on, everything you

0:03:50.800 --> 0:03:53.680
<v Speaker 1>write in that textbox. All the things you're friends like

0:03:54.160 --> 0:03:57.600
<v Speaker 1>help paint a picture of who you are. Facebook knows

0:03:57.640 --> 0:03:59.800
<v Speaker 1>what stories you go to, and it knows as you're

0:04:00.120 --> 0:04:02.640
<v Speaker 1>rolling what you stop to stare at, even if you

0:04:02.680 --> 0:04:05.240
<v Speaker 1>never click on it. All of these things end up

0:04:05.240 --> 0:04:07.680
<v Speaker 1>as signals that Facebook can you use to determine what

0:04:07.840 --> 0:04:10.560
<v Speaker 1>kind of ads a person might be interested in seeing.

0:04:11.120 --> 0:04:14.480
<v Speaker 1>Are they rich or poor, Asian or Latino, a mother

0:04:14.640 --> 0:04:17.680
<v Speaker 1>or a student. The advertisers can't really see who is

0:04:17.680 --> 0:04:21.680
<v Speaker 1>getting their ads, and Facebook can't at an individual level

0:04:21.720 --> 0:04:25.919
<v Speaker 1>at least either. All they know is it works. The

0:04:25.960 --> 0:04:28.520
<v Speaker 1>fact that Facebook doesn't really sell it's data, it just

0:04:28.600 --> 0:04:31.480
<v Speaker 1>sells access to it is actually one of the reasons

0:04:31.520 --> 0:04:34.680
<v Speaker 1>it's been so successful. It has all this information on

0:04:34.720 --> 0:04:38.000
<v Speaker 1>now more than two billion people basically locked up, and

0:04:38.000 --> 0:04:41.880
<v Speaker 1>it's only available to people or groups advertising directly on Facebook.

0:04:42.440 --> 0:04:44.279
<v Speaker 1>And it's actually one of the most important parts of

0:04:44.279 --> 0:04:46.680
<v Speaker 1>how Facebook works is we do not sell data. Advertisers

0:04:46.680 --> 0:04:52.120
<v Speaker 1>do not get access to people's individual data. Those advertisers

0:04:52.160 --> 0:04:55.120
<v Speaker 1>can use Facebook system to get really specific about what

0:04:55.279 --> 0:04:57.720
<v Speaker 1>kind of people they want to reach. If enough people

0:04:57.760 --> 0:05:01.320
<v Speaker 1>are into something, Facebook will auto create category for it,

0:05:01.720 --> 0:05:04.000
<v Speaker 1>like breeding insects. It might be kind of hard to

0:05:04.040 --> 0:05:06.960
<v Speaker 1>reach bug breeders if you're advertising in the New York Times,

0:05:07.360 --> 0:05:10.080
<v Speaker 1>But on Facebook you can let exactly the right people

0:05:10.160 --> 0:05:13.719
<v Speaker 1>know about your upcoming beekeeping conference. But what if you

0:05:13.760 --> 0:05:17.000
<v Speaker 1>want only white people to apply for your apartment complex,

0:05:17.200 --> 0:05:21.000
<v Speaker 1>or only young people to apply for your open job. Facebook,

0:05:21.000 --> 0:05:23.600
<v Speaker 1>for a time actually allowed people to do that too.

0:05:29.440 --> 0:05:32.479
<v Speaker 1>Of course, that's not how Facebook intended for the ad

0:05:32.520 --> 0:05:35.839
<v Speaker 1>machine to work. And while there are some awkward, unintended

0:05:35.880 --> 0:05:39.200
<v Speaker 1>consequences on the whole, the ad platform works very well

0:05:39.240 --> 0:05:41.440
<v Speaker 1>for a lot of its users. Like a couple of

0:05:41.440 --> 0:05:44.960
<v Speaker 1>months ago, I went with our producer Piaga Kari to

0:05:45.080 --> 0:05:47.960
<v Speaker 1>visit a small business in New York. It's called Two

0:05:47.960 --> 0:05:51.000
<v Speaker 1>Blind Brothers and they make T shirts. Should we give

0:05:51.040 --> 0:05:56.080
<v Speaker 1>you guys quick toward? Yeah, let's do it. Um, well,

0:05:56.080 --> 0:05:58.400
<v Speaker 1>you're pretty much looking at all of it already, but

0:05:59.560 --> 0:06:02.840
<v Speaker 1>we have a few fun things, and here we have this. Uh.

0:06:03.279 --> 0:06:06.599
<v Speaker 1>The company is run by two brothers, Brad and Brian.

0:06:07.040 --> 0:06:10.839
<v Speaker 1>They're both visually impaired, and profits from their T shirt

0:06:10.920 --> 0:06:18.400
<v Speaker 1>sales are donated to organizations that fight blindness, working on

0:06:18.680 --> 0:06:25.400
<v Speaker 1>some stuff for kids. Uh, we have t shirts Brad

0:06:25.440 --> 0:06:28.080
<v Speaker 1>and Brian say their company would never have even gotten

0:06:28.120 --> 0:06:30.960
<v Speaker 1>off the ground without Facebook's ad system, which allowed them

0:06:31.000 --> 0:06:33.680
<v Speaker 1>to start boosting posts on their business page without spending

0:06:33.680 --> 0:06:36.840
<v Speaker 1>a lot of money. It works so well that Facebook

0:06:36.880 --> 0:06:39.400
<v Speaker 1>actually published a case study about one of their Black

0:06:39.440 --> 0:06:42.800
<v Speaker 1>Friday campaigns. The actual study is there on a wall

0:06:42.839 --> 0:06:45.240
<v Speaker 1>in a frame, and it's been signed by Chief Operating

0:06:45.279 --> 0:06:49.520
<v Speaker 1>Officer Cheryl Sandberg herself. So why did you start advertising

0:06:49.520 --> 0:06:53.760
<v Speaker 1>on Facebook? You know, Facebook does a great job of

0:06:53.839 --> 0:06:57.400
<v Speaker 1>baby stepping you into experimenting with the adds. So I

0:06:57.480 --> 0:07:02.760
<v Speaker 1>remember we had put out this video and we went

0:07:02.800 --> 0:07:04.520
<v Speaker 1>to bed that night. We woke up the next day

0:07:04.520 --> 0:07:07.200
<v Speaker 1>and we're like, what the like? Normally when I put

0:07:07.200 --> 0:07:10.480
<v Speaker 1>out a post on Facebook, it's like Grandma and like

0:07:10.560 --> 0:07:13.200
<v Speaker 1>my sister that like it? And then this post just

0:07:13.800 --> 0:07:16.320
<v Speaker 1>and So they had this dynamite video and they started

0:07:16.320 --> 0:07:19.800
<v Speaker 1>targeting certain audiences with it. They could select for specific

0:07:19.840 --> 0:07:22.680
<v Speaker 1>age groups or find people with an interest in blindness.

0:07:23.040 --> 0:07:25.200
<v Speaker 1>But the brothers said it actually worked better when they

0:07:25.240 --> 0:07:27.880
<v Speaker 1>could just click a little button that would get Facebook

0:07:27.880 --> 0:07:30.800
<v Speaker 1>to do the job for them, choosing on its own

0:07:31.000 --> 0:07:33.360
<v Speaker 1>who could see their ads right. They started using a

0:07:33.400 --> 0:07:37.480
<v Speaker 1>tool called look Alike audiences. Yeah, so you know I can.

0:07:37.640 --> 0:07:40.040
<v Speaker 1>I can get on the app platform and I can

0:07:40.120 --> 0:07:45.720
<v Speaker 1>literally put into the to our audiences. I'm looking for

0:07:45.760 --> 0:07:50.920
<v Speaker 1>somebody who uh likes tom shoes and like selling Degenerous

0:07:50.920 --> 0:07:54.320
<v Speaker 1>because you know, we're featured on her show, and who

0:07:54.720 --> 0:07:58.200
<v Speaker 1>is an online shopper and all these sort of broad

0:07:58.440 --> 0:08:07.280
<v Speaker 1>categorical um characteristics. Or I can let Facebook analyze who's

0:08:07.320 --> 0:08:10.120
<v Speaker 1>interacting with our brand and then they can decide on

0:08:10.320 --> 0:08:14.720
<v Speaker 1>hundreds of factors using all their technology to figure out

0:08:14.760 --> 0:08:17.480
<v Speaker 1>who that best audience is. But that is an audience

0:08:17.480 --> 0:08:20.720
<v Speaker 1>that Facebook has determined for us looks and acts like

0:08:20.880 --> 0:08:23.800
<v Speaker 1>our best and most exciting customers, and then we can

0:08:23.840 --> 0:08:26.720
<v Speaker 1>take our ad creative and send it to that audience.

0:08:27.240 --> 0:08:30.080
<v Speaker 1>So the lookalike feature just finds people who are similar

0:08:30.120 --> 0:08:33.720
<v Speaker 1>to the users you've reached already. It's all automatic, and

0:08:33.880 --> 0:08:37.480
<v Speaker 1>as the Brothers realize pretty quickly, it's incredibly effective. It's

0:08:37.480 --> 0:08:40.680
<v Speaker 1>not just effective for businesses, though, it's effective for all

0:08:40.800 --> 0:08:44.160
<v Speaker 1>kinds of advertising. This type of ready made targeting was

0:08:44.200 --> 0:08:46.480
<v Speaker 1>one of the main tools Donald Trump used in its

0:08:46.520 --> 0:08:50.000
<v Speaker 1>presidential election campaign. It allows someone who might not be

0:08:50.040 --> 0:08:53.720
<v Speaker 1>in the mainstream conversation to find just enough people who

0:08:53.760 --> 0:08:56.480
<v Speaker 1>share an interest, and then people who look like those

0:08:56.480 --> 0:09:06.880
<v Speaker 1>people and expanded from there. So we know that Facebook's

0:09:06.880 --> 0:09:09.560
<v Speaker 1>ad targeting is super efficient, and we know that businesses

0:09:09.559 --> 0:09:12.680
<v Speaker 1>and political groups are using it. But a lot of

0:09:12.679 --> 0:09:16.319
<v Speaker 1>these tools still feel somewhat mysterious to a lot of people.

0:09:17.040 --> 0:09:19.680
<v Speaker 1>So we went to Facebook's headquarters in Menlo Park and

0:09:19.760 --> 0:09:25.880
<v Speaker 1>asked them, Yeah, Thursday traffics bad. Bad. It was like

0:09:25.960 --> 0:09:28.480
<v Speaker 1>telling me that I was going to get here. Hi,

0:09:28.600 --> 0:09:37.640
<v Speaker 1>hi Hi. This is Nikila Shwinavasan, the product manager for

0:09:37.800 --> 0:09:41.600
<v Speaker 1>Facebook Small Business Ads Growth. She knows inside and out

0:09:41.679 --> 0:09:44.080
<v Speaker 1>the portal that most people come into when they want

0:09:44.080 --> 0:09:46.800
<v Speaker 1>to buy ads, and she also knows that not everyone

0:09:46.920 --> 0:09:50.480
<v Speaker 1>understands how it works behind the scenes. To your point earlier,

0:09:50.600 --> 0:09:53.640
<v Speaker 1>I do empathize with the fact that it can feel

0:09:53.679 --> 0:09:56.240
<v Speaker 1>like a black box because after I had boost posts,

0:09:56.280 --> 0:10:00.439
<v Speaker 1>now what actually happens, right, And that's where the magic

0:10:00.480 --> 0:10:05.800
<v Speaker 1>of the auction comes in. The auction is the heart

0:10:05.880 --> 0:10:09.320
<v Speaker 1>of Facebook's advertising system. That's how it figures out how

0:10:09.400 --> 0:10:12.440
<v Speaker 1>much advertisers are willing to pay to reach someone like

0:10:12.520 --> 0:10:16.280
<v Speaker 1>those hard to reach beekeepers we mentioned earlier. But where

0:10:16.280 --> 0:10:19.680
<v Speaker 1>Facebook really excels is figuring out who those people are

0:10:19.720 --> 0:10:23.240
<v Speaker 1>in the first place. So when we think about that,

0:10:23.280 --> 0:10:26.440
<v Speaker 1>there's a few dimensions, and obviously it's a it's a

0:10:26.480 --> 0:10:30.840
<v Speaker 1>fairly complex system. But what we try to do is understand, Okay,

0:10:30.880 --> 0:10:34.200
<v Speaker 1>what do we know about this person based on previous

0:10:34.200 --> 0:10:38.440
<v Speaker 1>activity that they've had, previous ads, that they've interacted with um,

0:10:38.520 --> 0:10:42.240
<v Speaker 1>And we consider also what is their likelihood of taking

0:10:42.320 --> 0:10:45.400
<v Speaker 1>this particular action that the advertiser wants them to take,

0:10:45.480 --> 0:10:49.280
<v Speaker 1>which is in this case actually Facebook uses stuff like age, gender,

0:10:49.280 --> 0:10:53.360
<v Speaker 1>and location, of course, but that's not all. I asked

0:10:53.440 --> 0:10:57.320
<v Speaker 1>Ninkila what characteristics exactly Facebook was using to find these

0:10:57.360 --> 0:11:01.920
<v Speaker 1>target audiences. How many signals are all in the I

0:11:02.080 --> 0:11:06.000
<v Speaker 1>really couldn't tell you, because there's a whole bunch like liking,

0:11:06.440 --> 0:11:09.640
<v Speaker 1>commenting on an ad, liking a page, liking a post,

0:11:10.120 --> 0:11:15.000
<v Speaker 1>um or activities that you have done that advertisers are

0:11:15.040 --> 0:11:17.800
<v Speaker 1>able to then come back and retarget um. And then

0:11:17.840 --> 0:11:21.720
<v Speaker 1>there's also signals that we can extrapolate based on what

0:11:21.800 --> 0:11:25.720
<v Speaker 1>we know people like you tend to like. So if

0:11:25.760 --> 0:11:29.360
<v Speaker 1>we know that you visited a particular retailer before, and

0:11:29.400 --> 0:11:32.960
<v Speaker 1>there's another retailer that's selling a similar good, it's the

0:11:33.040 --> 0:11:36.520
<v Speaker 1>likelihood of view interacting with that is also higher. So

0:11:36.679 --> 0:11:40.200
<v Speaker 1>the answer to how many signals is I don't have

0:11:40.240 --> 0:11:44.080
<v Speaker 1>a number. The software is so complicated that even people

0:11:44.120 --> 0:11:46.600
<v Speaker 1>who work at Facebook can't see what's happening to the

0:11:46.600 --> 0:11:49.400
<v Speaker 1>system as a whole. Knowing gets to know exactly who

0:11:49.520 --> 0:11:53.679
<v Speaker 1>is seeing a particular ad. So what about Facebook? This Facebook? Now,

0:11:54.679 --> 0:11:56.680
<v Speaker 1>like if you go look at an advertiser, do you

0:11:56.720 --> 0:12:00.800
<v Speaker 1>know exactly whose ads that person reached? Know? We we

0:12:00.880 --> 0:12:04.120
<v Speaker 1>did not look at individual level data, nor do we

0:12:04.160 --> 0:12:06.200
<v Speaker 1>surface it to your point, We do it at an

0:12:06.240 --> 0:12:09.240
<v Speaker 1>aggregate level. We surface that to the advertiser at an

0:12:09.240 --> 0:12:12.679
<v Speaker 1>aggregate level where we're telling them, but they're now at

0:12:12.720 --> 0:12:17.000
<v Speaker 1>an we don't go look at the individual level. Now. Yeah,

0:12:17.160 --> 0:12:20.400
<v Speaker 1>So I guess, I mean, I guess it's one thing

0:12:20.440 --> 0:12:23.280
<v Speaker 1>that's interesting about this whole system is it's it is

0:12:23.400 --> 0:12:29.400
<v Speaker 1>so it's like, um, very automated and optimized. There are

0:12:29.440 --> 0:12:33.160
<v Speaker 1>all these signals that go into it, and um, you know,

0:12:33.280 --> 0:12:38.680
<v Speaker 1>out the other end comes something that's really useful and helpful. Um,

0:12:38.720 --> 0:12:42.439
<v Speaker 1>but how do you how do you look for problems?

0:12:42.520 --> 0:12:45.480
<v Speaker 1>Like how are you proactive about like the next iteration

0:12:45.559 --> 0:12:48.640
<v Speaker 1>of it. Yeah, that's a great question. We hold ourselves

0:12:48.640 --> 0:12:52.040
<v Speaker 1>even internally, to very high standards when we think about

0:12:52.160 --> 0:12:54.280
<v Speaker 1>how we developed, and it turns out with so much

0:12:54.320 --> 0:12:57.000
<v Speaker 1>content and only so many people, rooting out problems is

0:12:57.200 --> 0:13:01.040
<v Speaker 1>very very hard. Facebook doesn't really resp on two problems

0:13:01.120 --> 0:13:04.839
<v Speaker 1>until they are reported by users, and sometimes it's people

0:13:04.880 --> 0:13:08.240
<v Speaker 1>saying they're annoyed by seeing a certain ad, and other

0:13:08.320 --> 0:13:14.200
<v Speaker 1>times it's much more serious when it comes to political advertising.

0:13:14.360 --> 0:13:17.840
<v Speaker 1>Obviously the stakes are high. There's still plenty of things

0:13:17.840 --> 0:13:20.800
<v Speaker 1>that can and have gone wrong. That's despite the huge

0:13:20.840 --> 0:13:24.439
<v Speaker 1>resources that Facebook is devoted to rooting out problems. There's

0:13:24.520 --> 0:13:27.800
<v Speaker 1>just so much that goes unseen when you've tasked a

0:13:27.880 --> 0:13:31.400
<v Speaker 1>smart set of computer programs with running your business. We've

0:13:31.400 --> 0:13:35.360
<v Speaker 1>talked about the categories that get automatically created, like users

0:13:35.360 --> 0:13:38.960
<v Speaker 1>who are saying into treason in Russia. Facebook had to

0:13:38.960 --> 0:13:41.600
<v Speaker 1>go through and manually remove about five thousand of them.

0:13:41.600 --> 0:13:44.160
<v Speaker 1>That could be problematic, right, but they did so only

0:13:44.200 --> 0:13:47.880
<v Speaker 1>after a series of articles from the news organization Pro Publica.

0:13:48.040 --> 0:13:50.199
<v Speaker 1>It seems as though they're not proactively looking at the

0:13:50.200 --> 0:13:53.760
<v Speaker 1>potential problems themselves, like taking action when they think it

0:13:53.800 --> 0:13:56.120
<v Speaker 1>might be the right thing to do, like the day

0:13:56.160 --> 0:14:00.440
<v Speaker 1>before the midterms, Facebook stopped letting Trump turn and anti

0:14:00.520 --> 0:14:04.520
<v Speaker 1>immigrant video into an ad on its network. It made

0:14:04.520 --> 0:14:07.880
<v Speaker 1>the decision after major TV networks had already made it,

0:14:08.320 --> 0:14:10.839
<v Speaker 1>and also the ad flip through its network in the

0:14:10.880 --> 0:14:14.720
<v Speaker 1>first place. Also, once they decided to stop letting it

0:14:14.760 --> 0:14:18.000
<v Speaker 1>be an ad, it was still an incredibly popular organic

0:14:18.080 --> 0:14:21.160
<v Speaker 1>post on Facebook. Well, why change if you've got such

0:14:21.200 --> 0:14:23.560
<v Speaker 1>a successful thing going. What we saw with the Two

0:14:23.600 --> 0:14:26.520
<v Speaker 1>Blind Brothers is that Facebook makes a lot of money

0:14:26.520 --> 0:14:31.359
<v Speaker 1>off small, really small businesses that are targeting incredibly niche audiences.

0:14:31.840 --> 0:14:35.320
<v Speaker 1>There's no other service, I mean, maybe Google's ad platform

0:14:35.600 --> 0:14:39.000
<v Speaker 1>that can cater to such specific needs. The main area

0:14:39.080 --> 0:14:42.720
<v Speaker 1>they've tried to tame is politics. A lesson spending just

0:14:43.160 --> 0:14:46.160
<v Speaker 1>isn't a very big part of Facebook's business. The company

0:14:46.200 --> 0:14:49.640
<v Speaker 1>earlier this year even considered eliminating and entirely given the

0:14:49.640 --> 0:14:52.360
<v Speaker 1>many headaches that generates for the company. But Sarah, the

0:14:52.400 --> 0:14:55.800
<v Speaker 1>fixes generated headaches of their own, too, Right, You're totally right.

0:14:55.840 --> 0:14:58.640
<v Speaker 1>I wrote this story a few weeks ago about how

0:14:58.720 --> 0:15:03.560
<v Speaker 1>Facebook's new rules around political advertising now require anyone advertising

0:15:03.560 --> 0:15:07.640
<v Speaker 1>about a sort of electoral issue to be verified, but

0:15:07.680 --> 0:15:10.120
<v Speaker 1>the system was catching and taking down ads that hadn't

0:15:10.160 --> 0:15:14.080
<v Speaker 1>gone through the verification process, and those ads sometimes had

0:15:14.160 --> 0:15:17.400
<v Speaker 1>nothing to do with politics. Well, so there was this

0:15:17.520 --> 0:15:20.480
<v Speaker 1>vacation Bible school that got blocked because it was located

0:15:20.480 --> 0:15:24.280
<v Speaker 1>in Clinton, Indiana. In an insurance company in Clinton, Iowa

0:15:24.400 --> 0:15:26.920
<v Speaker 1>that had an ad about its family baseball night taken

0:15:26.960 --> 0:15:29.800
<v Speaker 1>down because they had the word Clinton in their ads,

0:15:30.280 --> 0:15:33.400
<v Speaker 1>as in former President Bill Clinton or Secretary of State

0:15:33.480 --> 0:15:36.400
<v Speaker 1>Hillary Clinton. But it also happens to be a common

0:15:36.480 --> 0:15:38.720
<v Speaker 1>name of towns in the US. And I think we've

0:15:38.720 --> 0:15:41.720
<v Speaker 1>heard so much about AI, but it's it sounds like

0:15:41.720 --> 0:15:44.480
<v Speaker 1>Facebook came up with a logical solution to a need

0:15:44.520 --> 0:15:48.160
<v Speaker 1>for more transparency and then signed an AI algorithm to

0:15:48.200 --> 0:15:51.320
<v Speaker 1>do the work. But AI, despite the hype, it's not

0:15:51.400 --> 0:15:57.240
<v Speaker 1>smart enough to figure out what actually might be political second,

0:15:58.400 --> 0:16:05.760
<v Speaker 1>making advertising work parents that people. Facebook is devoting huge

0:16:05.760 --> 0:16:09.520
<v Speaker 1>resources towards making sure it's more proactive about political behavior

0:16:09.560 --> 0:16:12.880
<v Speaker 1>on its site in I recently went back to Facebook

0:16:12.880 --> 0:16:15.240
<v Speaker 1>headquarters and took a tour of what it's calling the

0:16:15.280 --> 0:16:18.280
<v Speaker 1>War Room. It's basically this room of a couple dozen

0:16:18.320 --> 0:16:22.000
<v Speaker 1>people monitoring dat boards of articles that are trending and

0:16:22.200 --> 0:16:25.080
<v Speaker 1>political ad money flowing so they can tackle problems related

0:16:25.080 --> 0:16:29.480
<v Speaker 1>to election first one aunts so the biggest and most

0:16:29.520 --> 0:16:31.760
<v Speaker 1>important thing we've done here in the last two years

0:16:31.920 --> 0:16:36.080
<v Speaker 1>is huge advances and machine learning. That includes deleting fake accounts,

0:16:36.080 --> 0:16:38.840
<v Speaker 1>sending fake news to fact checkers, and a lot of

0:16:38.840 --> 0:16:41.840
<v Speaker 1>the issues that got it in trouble during the election.

0:16:42.360 --> 0:16:44.880
<v Speaker 1>Facebook has been showing this war room off to reporters

0:16:45.000 --> 0:16:47.720
<v Speaker 1>as a sign that it's taking political content on the

0:16:47.760 --> 0:16:51.040
<v Speaker 1>platform extra seriously. And they also have that at archive

0:16:51.160 --> 0:16:53.800
<v Speaker 1>right where everyone can see who's bidding on specific ads.

0:16:54.360 --> 0:16:57.120
<v Speaker 1>That's a big step, but there are still issues. For example,

0:16:57.200 --> 0:16:59.680
<v Speaker 1>the feature that allows users to see who's paying for

0:16:59.720 --> 0:17:03.120
<v Speaker 1>each ad isn't perfect. A couple of Vice reporters realized

0:17:03.160 --> 0:17:05.680
<v Speaker 1>that they could personally buy ads and claim that they

0:17:05.680 --> 0:17:09.440
<v Speaker 1>were paid for by the Senate Republican Majority leader Mr. Racconnell.

0:17:09.960 --> 0:17:12.760
<v Speaker 1>And in Virginia, a bunch of really savage attack ads

0:17:12.840 --> 0:17:15.359
<v Speaker 1>ran against a House candidate, but the buyer's managed to

0:17:15.400 --> 0:17:18.879
<v Speaker 1>completely obscure who they were or where the money came from.

0:17:18.920 --> 0:17:22.199
<v Speaker 1>In two days before election day, Facebook was alerted by

0:17:22.200 --> 0:17:25.639
<v Speaker 1>the federal government about foreign accounts that could be trying

0:17:25.640 --> 0:17:28.760
<v Speaker 1>to influence selection that it had to take down in

0:17:28.800 --> 0:17:31.159
<v Speaker 1>the day after the elections, the company announced it was

0:17:31.200 --> 0:17:35.840
<v Speaker 1>strengthening its verification process. But meanwhile it's still dealing with

0:17:36.080 --> 0:17:48.399
<v Speaker 1>a lot of influenced campaigns by foreign governments. So, Sarah,

0:17:48.400 --> 0:17:51.440
<v Speaker 1>we've devoted this whole season of Decrypted to this idea

0:17:51.480 --> 0:17:54.720
<v Speaker 1>of unintended consequences, and this is almost the biggest unintended

0:17:54.760 --> 0:17:59.159
<v Speaker 1>consequence of all. Silicon Valley designed these systems to scale infinitely,

0:17:59.760 --> 0:18:04.000
<v Speaker 1>very low human touch, um completely automated systems, and what

0:18:04.080 --> 0:18:06.800
<v Speaker 1>they failed to recognize and which in this case had

0:18:07.040 --> 0:18:11.160
<v Speaker 1>severe implications for the world, is that the problems scaled

0:18:11.520 --> 0:18:15.639
<v Speaker 1>right along with the systems. Right when you're prioritizing convenience

0:18:15.720 --> 0:18:18.800
<v Speaker 1>and speed and low touch, right, these are systems where

0:18:18.840 --> 0:18:22.480
<v Speaker 1>you really don't have to have a person selling and

0:18:22.560 --> 0:18:27.119
<v Speaker 1>placing and add um, it makes it so easy to

0:18:27.320 --> 0:18:30.840
<v Speaker 1>miss the little things. And what Facebook says over and over,

0:18:31.440 --> 0:18:33.919
<v Speaker 1>you know, they point to examples like two blind Brothers.

0:18:33.960 --> 0:18:37.639
<v Speaker 1>They point to the many small businesses that Nkuila is

0:18:37.680 --> 0:18:41.440
<v Speaker 1>helping out, and they say that the good that Facebook

0:18:41.520 --> 0:18:44.360
<v Speaker 1>is done in helping these small businesses tell their stories

0:18:44.400 --> 0:18:47.280
<v Speaker 1>and helping people have a voice in a way they

0:18:47.280 --> 0:18:52.960
<v Speaker 1>wouldn't be able to afford to otherwise that outweighs the bad. Well,

0:18:52.960 --> 0:18:55.239
<v Speaker 1>but to me, they just seem like different things. Like

0:18:55.280 --> 0:18:57.520
<v Speaker 1>it was a system that was great for helping small

0:18:57.600 --> 0:19:02.320
<v Speaker 1>businesses reach big or even very tailored audiences. But when

0:19:02.320 --> 0:19:05.640
<v Speaker 1>it comes to a tug of war over ideas and

0:19:05.720 --> 0:19:08.560
<v Speaker 1>you know, and the balance between informing people and maybe

0:19:08.600 --> 0:19:10.480
<v Speaker 1>deceiving them and to get them to do things they

0:19:10.480 --> 0:19:13.120
<v Speaker 1>wouldn't otherwise do, that's where the system kind of breaks down.

0:19:13.280 --> 0:19:16.080
<v Speaker 1>And I think you can't just go and say just

0:19:16.119 --> 0:19:18.480
<v Speaker 1>because we've done a lot of good things means that

0:19:18.600 --> 0:19:21.840
<v Speaker 1>this service is good. I think, you know, so if

0:19:21.880 --> 0:19:26.159
<v Speaker 1>you were to look at a government, for example, and say, well,

0:19:26.880 --> 0:19:29.879
<v Speaker 1>the roads are really messed up, but the schools are great, Like,

0:19:30.000 --> 0:19:32.199
<v Speaker 1>that's just not the kind of argument you would make

0:19:32.240 --> 0:19:34.360
<v Speaker 1>about your government. You want all of it to work,

0:19:34.400 --> 0:19:35.639
<v Speaker 1>and you want all of it to work in a

0:19:35.680 --> 0:19:39.080
<v Speaker 1>way that's not dangerous for society. So let's take this

0:19:39.160 --> 0:19:41.520
<v Speaker 1>to the to the logical extreme. And we touched on

0:19:41.560 --> 0:19:44.840
<v Speaker 1>this earlier, like should Facebook just not allow a certain

0:19:44.920 --> 0:19:49.760
<v Speaker 1>kinds of advertising? Should it say we haven't figured this

0:19:49.960 --> 0:19:52.000
<v Speaker 1>how to make this work for political advertising and not

0:19:52.080 --> 0:19:55.479
<v Speaker 1>accept it? Well, then you get into the question of

0:19:55.520 --> 0:19:59.720
<v Speaker 1>what political advertising actually is, right, is it something that

0:19:59.880 --> 0:20:02.360
<v Speaker 1>is is just a candidate putting out an ad to

0:20:02.400 --> 0:20:04.720
<v Speaker 1>get you to vote for them, or is it about

0:20:04.760 --> 0:20:08.000
<v Speaker 1>these issues? And that the issues are really what Russia

0:20:08.119 --> 0:20:11.840
<v Speaker 1>tried to stoke tension around in the election. They weren't

0:20:11.880 --> 0:20:14.879
<v Speaker 1>saying vote for Hillary, vote for Trump. They were saying

0:20:15.840 --> 0:20:20.240
<v Speaker 1>you should get angry about immigration and refugees and race

0:20:20.400 --> 0:20:24.240
<v Speaker 1>in America. And so I don't think that getting rid

0:20:24.280 --> 0:20:27.160
<v Speaker 1>of a category of advertising really solves it. I think

0:20:27.200 --> 0:20:30.560
<v Speaker 1>they have to think about how to proactively manage these

0:20:30.600 --> 0:20:34.439
<v Speaker 1>systems with a little bit more human involvement, even if

0:20:34.480 --> 0:20:37.439
<v Speaker 1>that doesn't scale. And verification obviously being so important, but

0:20:37.480 --> 0:20:39.840
<v Speaker 1>it's been difficult for them to get that right. Why

0:20:39.880 --> 0:20:42.479
<v Speaker 1>why is it so? Why is it so hard? Well,

0:20:42.680 --> 0:20:45.880
<v Speaker 1>when you're trying to verify, you have to go by

0:20:45.920 --> 0:20:47.679
<v Speaker 1>what the people say that and they want it to

0:20:47.720 --> 0:20:52.000
<v Speaker 1>be fast, and in politics, they've already gotten so many

0:20:52.080 --> 0:20:56.040
<v Speaker 1>complaints about how slow that system has been, where people

0:20:56.119 --> 0:20:59.520
<v Speaker 1>send in their verification and it takes a few days

0:20:59.520 --> 0:21:03.159
<v Speaker 1>to pross us. Meanwhile, you know the rival candidate in

0:21:03.200 --> 0:21:06.639
<v Speaker 1>a campaign is spending tons of money on their ads,

0:21:07.160 --> 0:21:11.159
<v Speaker 1>and that seems unfair. And so no matter what Facebook does,

0:21:11.200 --> 0:21:13.240
<v Speaker 1>there's going to be this question of are they are

0:21:13.280 --> 0:21:16.040
<v Speaker 1>they censoring or are they removing enough? And they have

0:21:16.119 --> 0:21:20.480
<v Speaker 1>to figure out what the right balances, and maybe they'll

0:21:20.480 --> 0:21:22.359
<v Speaker 1>never get there right and they can't be seen as

0:21:22.440 --> 0:21:24.879
<v Speaker 1>favoring one side over another. And we should say this

0:21:24.920 --> 0:21:27.480
<v Speaker 1>isn't just a Facebook problem, right, it's internet wide. It's

0:21:27.480 --> 0:21:30.280
<v Speaker 1>an Internet problem, and it's a problem with this kind

0:21:30.280 --> 0:21:33.720
<v Speaker 1>of this way that the assistance have been built over

0:21:33.760 --> 0:21:37.359
<v Speaker 1>the years. These scale so well, they work so efficiently,

0:21:37.400 --> 0:21:41.240
<v Speaker 1>they're so high margin. Wall Street loves it. Uh, but

0:21:41.480 --> 0:21:44.400
<v Speaker 1>you know, when it comes to the real world consequences,

0:21:45.119 --> 0:21:48.400
<v Speaker 1>it doesn't work to just wait for users to report

0:21:48.440 --> 0:21:51.720
<v Speaker 1>when things go wrong. So perhaps eventually there will be

0:21:51.760 --> 0:21:54.280
<v Speaker 1>a system wide fixed. For now, though, I'm tempted to

0:21:54.280 --> 0:21:57.720
<v Speaker 1>say that it's incumbent on users to just be more vigilant,

0:21:57.760 --> 0:22:01.600
<v Speaker 1>to be smarter. Um. There, you know our our advertisers

0:22:01.600 --> 0:22:03.320
<v Speaker 1>out there that are trying to convince you to buy

0:22:03.440 --> 0:22:06.960
<v Speaker 1>something new, maybe something you don't need. That's fairly innocuous,

0:22:07.040 --> 0:22:08.960
<v Speaker 1>but in other cases they might be trying to get

0:22:09.000 --> 0:22:11.199
<v Speaker 1>you to act in a way that you otherwise might

0:22:11.240 --> 0:22:33.160
<v Speaker 1>not and that's it for this week's Decrypted. Thanks for listening.

0:22:33.520 --> 0:22:36.000
<v Speaker 1>Do you have a story about advertising on Facebook? We

0:22:36.040 --> 0:22:38.160
<v Speaker 1>want to hear from you. You can email us at

0:22:38.200 --> 0:22:41.520
<v Speaker 1>Decrypted at bloomberg dot net or I'm on Twitter at

0:22:41.520 --> 0:22:44.280
<v Speaker 1>Sarah Fryer and I'm at brad Stone. If you're a

0:22:44.280 --> 0:22:46.400
<v Speaker 1>fan of the show, please take a moment to rate

0:22:46.440 --> 0:22:49.480
<v Speaker 1>and review us. It really helps new listeners find the show.

0:22:50.320 --> 0:22:52.960
<v Speaker 1>This episode was produced by pa good Kari and Mage

0:22:53.000 --> 0:22:56.800
<v Speaker 1>mus Hendrickson. Our story editor was Anne vander Nay. Thanks

0:22:56.840 --> 0:23:01.119
<v Speaker 1>to Akito, Emily Buso, and Liz Smith. Friend Chessca Levie

0:23:01.200 --> 0:23:03.720
<v Speaker 1>is head of Bloomberg Podcast. We'll see you next week.

0:23:16.240 --> 0:23:16.280
<v Speaker 1>M