1
00:00:01,041 --> 00:00:04,175
Hello, I'm Joy, a poet of code,

2
00:00:04,199 --> 00:00:09,192
on a mission to stop
an unseen force that's rising,

3
00:00:09,216 --> 00:00:12,072
a force that I called "the coded gaze,"

4
00:00:12,096 --> 00:00:15,405
my term for algorithmic bias.

5
00:00:15,429 --> 00:00:19,729
Algorithmic bias, like human bias,
results in unfairness.

6
00:00:19,753 --> 00:00:25,775
However, algorithms, like viruses,
can spread bias on a massive scale

7
00:00:25,799 --> 00:00:27,381
at a rapid pace.

8
00:00:27,943 --> 00:00:32,330
Algorithmic bias can also lead
to exclusionary experiences

9
00:00:32,354 --> 00:00:34,482
and discriminatory practices.

10
00:00:34,506 --> 00:00:36,567
Let me show you what I mean.

11
00:00:38,250 --> 00:00:40,490
(Video) Joy Boulamwini: Camera.
I've got a face.

12
00:00:40,514 --> 00:00:42,387
Can you see my face?

13
00:00:42,411 --> 00:00:45,524
No-glasses face.

14
00:00:45,548 --> 00:00:48,241
You can see her face.

15
00:00:48,265 --> 00:00:50,495
What about my face?

16
00:00:53,255 --> 00:00:54,406
(Laughter)

17
00:00:54,430 --> 00:00:58,180
I've got a mask. Can you see my mask?

18
00:00:59,014 --> 00:01:01,379
Joy Boulamwini: So how did this happen?

19
00:01:01,403 --> 00:01:04,544
Why am I sitting in front of a computer

20
00:01:04,568 --> 00:01:05,992
in a white mask,

21
00:01:06,016 --> 00:01:09,666
trying to be detected by a cheap webcam?

22
00:01:09,690 --> 00:01:11,981
Well, when I'm not fighting the coded gaze

23
00:01:12,005 --> 00:01:13,525
as a poet of code,

24
00:01:13,549 --> 00:01:16,821
I'm a graduate student
at the MIT Media Lab,

25
00:01:16,845 --> 00:01:21,762
and there I have the opportunity to work
on all sorts of whimsical projects,

26
00:01:21,786 --> 00:01:23,813
including the Aspire Mirror,

27
00:01:23,837 --> 00:01:28,971
a project I did so I could project
digital masks onto my reflection.

28
00:01:28,995 --> 00:01:31,345
So in the morning, if I wanted
to feel powerful,

29
00:01:31,369 --> 00:01:32,803
I could put on a lion.

30
00:01:32,827 --> 00:01:36,323
If I wanted to be uplifted,
I might have a quote.

31
00:01:36,347 --> 00:01:39,336
So I used generic
facial recognition software

32
00:01:39,360 --> 00:01:40,711
to build the system,

33
00:01:40,735 --> 00:01:45,838
but found it was really hard to test it
unless I wore a white mask.

34
00:01:46,822 --> 00:01:51,168
Unfortunately, I've run
into this issue before.

35
00:01:51,192 --> 00:01:55,495
When I was an undergraduate
at Georgia Tech studying computer science,

36
00:01:55,519 --> 00:01:57,574
I used to work on social robots,

37
00:01:57,598 --> 00:02:01,375
and one of my tasks was to get a robot
to play peek-a-boo,

38
00:02:01,399 --> 00:02:03,082
a simple turn-taking game

39
00:02:03,106 --> 00:02:07,427
where partners cover their face
and then uncover it saying, "Peek-a-boo!"

40
00:02:07,451 --> 00:02:11,880
The problem is, peek-a-boo
doesn't really work if I can't see you,

41
00:02:11,904 --> 00:02:14,403
and my robot couldn't see me.

42
00:02:14,427 --> 00:02:18,377
But I borrowed my roommate's face
to get the project done,

43
00:02:18,401 --> 00:02:19,781
submitted the assignment,

44
00:02:19,805 --> 00:02:23,558
and figured, you know what,
somebody else will solve this problem.

45
00:02:24,209 --> 00:02:26,212
Not too longer after,

46
00:02:26,236 --> 00:02:30,395
I was in Hong Kong
for an entrepreneurship competition.

47
00:02:30,879 --> 00:02:33,573
The organizers decided
to take participants

48
00:02:33,597 --> 00:02:35,969
on a tour of local start-ups.

49
00:02:35,993 --> 00:02:38,708
One of the start-ups had a social robot,

50
00:02:38,732 --> 00:02:40,644
and they decided to do a demo.

51
00:02:40,668 --> 00:02:43,648
The demo worked on everybody
until it got to me,

52
00:02:43,672 --> 00:02:45,595
and you can probably guess it.

53
00:02:45,619 --> 00:02:48,584
It couldn't detect my face.

54
00:02:48,608 --> 00:02:51,119
I asked the developers what was going on,

55
00:02:51,143 --> 00:02:56,676
and it turned out we had used the same
generic facial recognition software.

56
00:02:56,700 --> 00:02:58,350
Halfway around the world,

57
00:02:58,374 --> 00:03:02,226
I learned that algorithmic bias
can travel as quickly

58
00:03:02,250 --> 00:03:05,420
as it takes to download
some files off of the internet.

59
00:03:06,285 --> 00:03:09,361
So what's going on?
Why isn't my face being detected?

60
00:03:09,385 --> 00:03:12,741
Well, we have to look at
how we give machines sight.

61
00:03:12,765 --> 00:03:16,174
Computer vision uses
machine learning techniques

62
00:03:16,198 --> 00:03:18,078
to do facial recognition.

63
00:03:18,102 --> 00:03:21,999
So how this works is, you create
a training set with examples of faces.

64
00:03:22,023 --> 00:03:24,841
This is a face. This is a face.
This is not a face.

65
00:03:24,865 --> 00:03:29,384
And over time, you can teach a computer
how to recognize other faces.

66
00:03:29,408 --> 00:03:33,397
However, if the training sets
aren't really that diverse,

67
00:03:33,421 --> 00:03:36,770
any face that deviates too much
from the established norm

68
00:03:36,794 --> 00:03:38,443
will be harder to detect,

69
00:03:38,467 --> 00:03:40,430
which is what was happening to me.

70
00:03:40,454 --> 00:03:42,836
But don't worry -- there's some good news.

71
00:03:42,860 --> 00:03:45,631
Training sets don't just
materialize out of nowhere.

72
00:03:45,655 --> 00:03:47,443
We actually can create them.

73
00:03:47,467 --> 00:03:51,643
So there's an opportunity to create
full-spectrum training sets

74
00:03:51,667 --> 00:03:55,491
that reflect a richer
portrait of humanity.

75
00:03:55,515 --> 00:03:57,736
Now you've seen in my examples

76
00:03:57,760 --> 00:03:59,528
how social robots

77
00:03:59,552 --> 00:04:04,163
was how I found out about exclusion
with algorithmic bias.

78
00:04:04,187 --> 00:04:09,002
But algorithmic bias can also lead
to discriminatory practices.

79
00:04:09,977 --> 00:04:11,430
Across the US,

80
00:04:11,454 --> 00:04:15,652
police departments are starting to use
facial recognition software

81
00:04:15,676 --> 00:04:18,135
in their crime-fighting arsenal.

82
00:04:18,159 --> 00:04:20,172
Georgetown Law published a report

83
00:04:20,196 --> 00:04:26,959
showing that one in two adults
in the US -- that's 117 million people --

84
00:04:26,983 --> 00:04:30,517
have their faces
in facial recognition networks.

85
00:04:30,541 --> 00:04:35,093
Police departments can currently look
at these networks unregulated,

86
00:04:35,117 --> 00:04:39,403
using algorithms that have not
been audited for accuracy.

87
00:04:39,427 --> 00:04:43,291
Yet we know facial recognition
is not fail proof,

88
00:04:43,315 --> 00:04:47,494
and labeling faces consistently
remains a challenge.

89
00:04:47,518 --> 00:04:49,280
You might have seen this on Facebook.

90
00:04:49,304 --> 00:04:52,292
My friends and I laugh all the time
when we see other people

91
00:04:52,316 --> 00:04:54,774
mislabeled in our photos.

92
00:04:54,798 --> 00:05:00,389
But misidentifying a suspected criminal
is no laughing matter,

93
00:05:00,413 --> 00:05:03,240
nor is breaching civil liberties.

94
00:05:03,264 --> 00:05:06,469
Machine learning is being used
for facial recognition,

95
00:05:06,493 --> 00:05:10,998
but it's also extending beyond the realm
of computer vision.

96
00:05:11,806 --> 00:05:15,822
In her book, "Weapons
of Math Destruction,"

97
00:05:15,846 --> 00:05:22,527
data scientist Cathy O'Neil
talks about the rising new WMDs --

98
00:05:22,551 --> 00:05:26,904
widespread, mysterious
and destructive algorithms

99
00:05:26,928 --> 00:05:29,892
that are increasingly being used
to make decisions

100
00:05:29,916 --> 00:05:33,093
that impact more aspects of our lives.

101
00:05:33,117 --> 00:05:34,987
So who gets hired or fired?

102
00:05:35,011 --> 00:05:37,123
Do you get that loan?
Do you get insurance?

103
00:05:37,147 --> 00:05:40,650
Are you admitted into the college
you wanted to get into?

104
00:05:40,674 --> 00:05:44,183
Do you and I pay the same price
for the same product

105
00:05:44,207 --> 00:05:46,649
purchased on the same platform?

106
00:05:46,673 --> 00:05:50,432
Law enforcement is also starting
to use machine learning

107
00:05:50,456 --> 00:05:52,745
for predictive policing.

108
00:05:52,769 --> 00:05:56,263
Some judges use machine-generated
risk scores to determine

109
00:05:56,287 --> 00:06:00,689
how long an individual
is going to spend in prison.

110
00:06:00,713 --> 00:06:03,167
So we really have to think
about these decisions.

111
00:06:03,191 --> 00:06:04,373
Are they fair?

112
00:06:04,397 --> 00:06:07,287
And we've seen that algorithmic bias

113
00:06:07,311 --> 00:06:10,685
doesn't necessarily always
lead to fair outcomes.

114
00:06:10,709 --> 00:06:12,673
So what can we do about it?

115
00:06:12,697 --> 00:06:16,377
Well, we can start thinking about
how we create more inclusive code

116
00:06:16,401 --> 00:06:19,391
and employ inclusive coding practices.

117
00:06:19,415 --> 00:06:21,724
It really starts with people.

118
00:06:22,248 --> 00:06:24,209
So who codes matters.

119
00:06:24,233 --> 00:06:28,352
Are we creating full-spectrum teams
with diverse individuals

120
00:06:28,376 --> 00:06:30,787
who can check each other's blind spots?

121
00:06:30,811 --> 00:06:34,356
On the technical side,
how we code matters.

122
00:06:34,380 --> 00:06:38,031
Are we factoring in fairness
as we're developing systems?

123
00:06:38,055 --> 00:06:40,968
And finally, why we code matters.

124
00:06:41,325 --> 00:06:46,408
We've used tools of computational creation
to unlock immense wealth.

125
00:06:46,432 --> 00:06:50,879
We now have the opportunity
to unlock even greater equality

126
00:06:50,903 --> 00:06:53,833
if we make social change a priority

127
00:06:53,857 --> 00:06:56,027
and not an afterthought.

128
00:06:56,548 --> 00:07:01,070
And so these are the three tenets
that will make up the "incoding" movement.

129
00:07:01,094 --> 00:07:02,746
Who codes matters,

130
00:07:02,770 --> 00:07:04,313
how we code matters

131
00:07:04,337 --> 00:07:06,360
and why we code matters.

132
00:07:06,384 --> 00:07:09,483
So to go towards encoding,
we can start thinking about

133
00:07:09,507 --> 00:07:12,671
building platforms that can identify bias

134
00:07:12,695 --> 00:07:15,773
by collecting people's experiences
like the ones I shared,

135
00:07:15,797 --> 00:07:18,867
but also auditing existing software.

136
00:07:18,891 --> 00:07:22,656
We can also start to create
more inclusive training sets.

137
00:07:22,680 --> 00:07:25,483
Imagine a "Selfies for Inclusion" campaign

138
00:07:25,507 --> 00:07:29,162
where you and I can help
developers test and create

139
00:07:29,186 --> 00:07:31,279
more inclusive training sets.

140
00:07:31,842 --> 00:07:34,670
And we can also start thinking
more conscientiously

141
00:07:34,694 --> 00:07:40,085
about the social impact
of the technology that we're developing.

142
00:07:40,109 --> 00:07:42,502
To get the incoding movement started,

143
00:07:42,526 --> 00:07:45,373
I've launched the Algorithmic
Justice League,

144
00:07:45,397 --> 00:07:51,269
where anyone who cares about fairness
can help fight the coded gaze.

145
00:07:51,293 --> 00:07:54,589
On codedgaze.com, you can report bias,

146
00:07:54,613 --> 00:07:57,058
request audits, become a tester

147
00:07:57,082 --> 00:07:59,853
and join the ongoing conversation,

148
00:07:59,877 --> 00:08:02,164
#codedgaze.

149
00:08:03,282 --> 00:08:05,769
So I invite you to join me

150
00:08:05,793 --> 00:08:09,512
in creating a world where technology
works for all of us,

151
00:08:09,536 --> 00:08:11,433
not just some of us,

152
00:08:11,457 --> 00:08:16,045
a world where we value inclusion
and center social change.

153
00:08:16,069 --> 00:08:17,244
Thank you.

154
00:08:17,268 --> 00:08:21,539
(Applause)

155
00:08:23,413 --> 00:08:26,267
But I have one question:

156
00:08:26,291 --> 00:08:28,350
Will you join me in the fight?

157
00:08:28,374 --> 00:08:29,659
(Laughter)

158
00:08:29,683 --> 00:08:33,370
(Applause)