字幕表 動画を再生する
-
Before the 1970s, people looking for jobs in the US would open up the “help wanted”
-
section of their newspapers and see this. One set of opportunities for women, and one
-
for men. We don't see job ads like this anymore,
-
largely because it's been illegal for decades. But also because advertising is now much more
-
targeted. Instead of one classified page, we have our social feeds, each crafted by
-
algorithms for an audience of one. So when this ad went out on Facebook and reached
-
a group of people that was 91% men, those outside that audience probably didn't know
-
it existed. And the same goes for this ad, which Facebook
-
displayed for an audience that was 88% women. That disparity wasn't because the advertiser
-
told Facebook to target users by gender. I know that because this is the advertiser.
-
My name is Muhammad Ali, I go by Ali. He's part of a research group at Northeastern
-
University that has spent thousands of dollars buying ads to try to figure out who Facebook
-
will show them to, and why.
-
If an ad shows up on your Facebook or Instagram
-
feed, there are two parties that decided you should see it. First, the advertiser included
-
you in their target audience, either by uploading a list of specific email addresses, phone
-
numbers, or previous visitors to their website, Or by choosing from thousands of attributes
-
that facebook offers, like Californians, under 40 who like basketball.
-
Second, Facebook decided who in that pool would actually see the ad through an automated
-
calculation based in part on what they know about you.
-
It's that second step that Ali and his colleagues wanted to study. If they uploaded a list of
-
randomly-generated American phone numbers, and then turned off all the targeting except
-
adults in the US, who would Facebook deliver the ad to?
-
So you set up a bodybuilding ad and a cosmetic ad and said we don't wanna target this any
-
further than the random phone numbers that we put in. Right? And then what were your
-
results? When Facebook started telling you who was actually seeing this ad, what did
-
they tell you? So, yeah, immediately, like we sort of expected
-
that the body building ad was more relevant to men. And that's exactly what we saw. I
-
think somewhere close to 80 to 85 percent of the audience was just men.
-
And the link that we advertise to elle.com about the makeup kits that you could buy that
-
went primarily to women. They were able to collect the results of the
-
ads over time so they knew the gender skew was there early on, suggesting that it wasn't
-
introduced by user behavior. Their experiment showed that Facebook automatically
-
analyzes the content of an ad to compare it to a user's interests.
-
How do they know what the user cares about? Well they have data from your profile and
-
everything you and your friends have done on facebook and instagram, as well as websites
-
you've visited, things you've purchased, apps you've installed, your location, your
-
devices, and more. All this information fuels automated predictions
-
about whether you are likely to engage with any given ad. And that prediction influences
-
whether the ad shows up on your feed at all. You can get a sense of what Facebook thinks
-
you're interested in on your Ad Preferences page. Or your Ad Interests on instagram.
-
Notice how some of these interests could correlate with your gender, your age, your income level,
-
or your race. And then you wanted to look at race. But it
-
sounds like Facebook does not give you data on the race of people that are seeing an ad.
-
So how do you study that? That was one of the harder things to do. We
-
thought we could use a different custom audience. Instead of random phone numbers. We could
-
take voter records from North Carolina, which are public, and they have the race of the
-
person registered as well. Then they bought ads for Rolling Stone articles
-
that were either about country albums, hip hop albums, or general top 30 albums and targeted
-
an equal number of white and Black users. And it was surprising how much the skew to
-
the Black users was for the hip bag versus the country and the top 30.
-
Facebook's algorithms are trained to not show people ads they won't be interested
-
in. But there may be cases when we're not comfortable
-
with Facebook making those predictions. One study by Ali and his colleagues investigated
-
how this plays out with political ads and found that despite targeting the same audiences,
-
using the same goal, bidding strategy, and budget,
-
an ad pointing to Bernie Sanders' site went to mostly Democrats and an ad for Trump went
-
to mostly Republicans. It cost 1.5 times more for an ad linking to
-
Sanders' site to reach the same number of conservatives as a Trump ad. Because Facebook
-
subsidizes what they consider to be “relevant” ads.
-
And then we move on to housing and employment ads, and these are considered sort of a different
-
category. Why is that? Because these are legally protected. For example, housing ads are protected by
-
the Fair Housing Act. An advertiser cannot discriminate in those
-
cases. At that point, you're excluding someone from a life opportunity which becomes much
-
more problematic. Because it's actually a legal violation that's
-
at stake? Possibly? Possibly.
-
Facebook allows advertisers to exclude certain ethnic groups from seeing an ad.
-
Dozens of employers placing job ads on Facebook that discriminate against older workers.
-
Facebook is revamping its targeted advertisements after settling lawsuits with civil rights
-
groups. In response to criticism and several lawsuits,
-
Facebook has been removing some of the targeting attributes that an advertiser could use to
-
discriminate against demographic groups, and is paying special attention to ads related
-
to employment, housing, and credit. But the role that the ad delivery system plays
-
remains unsolved. When Ali and his team tested out ads for job
-
openings in different industries, without targeting any demographic groups, facebook
-
generated some skewed audiences. The lumber industry post went to mostly men. The cleaner
-
post went to mostly women. The taxi driver ads that we ran, basically
-
seventy five percent of the audience was black users.
-
These results don't mean that Facebook is directly basing their predictions on our gender
-
or race. Instead it looks for patterns in all of our user data.
-
Maybe people who shop at a men's clothing site and like joe rogan are less likely to
-
click on an ad for a job teaching preschool. Maybe your data is similar to theirs and so
-
they predict you also wont click on that ad either.
-
Instead they show it to someone who likes skincare and feminism. And if that person
-
clicks, the system gets a new data point affirming its prediction.
-
A complaint filed by the US Department of Housing and Urban Development states that
-
this process “inevitably recreates groupings defined by their protected class.” They
-
said that Facebook's ad delivery system “prevents advertisers who want to reach
-
a broad audience of users from doing so.” According to a report by ProPublica, a construction
-
workers' union wanted to recruit diverse candidates for its apprenticeship program,
-
so they created ads featuring women, but found that Facebook still showed its them to mostly men.
-
And wouldn't any ad targeting system with
-
sort of sufficiently rich data about people have this kind of effect?
-
Well, we believe so, because a lot of these things, for example, custom audiences on all
-
of these targeting features --they're industry practice. They that also in Google's or Linkedin's or
-
Twitter's advertising platform. So the general ethos of how these systems work is the same.
-
It's a question that the industry as a whole hasn't answered:
-
When exactly is it unacceptable for an algorithm to decide that relevant audiences
-
are segregated ones?