1
00:00:01,041 --> 00:00:02,682
Hello, I'm Joy,

2
00:00:02,682 --> 00:00:04,199
a poet of code

3
00:00:04,199 --> 00:00:06,581
on a mission to stop

4
00:00:06,581 --> 00:00:09,216
an unseen force that's rising,

5
00:00:09,216 --> 00:00:12,096
a force that I called the coded gaze,

6
00:00:12,096 --> 00:00:15,429
my term for algorithmic bias.

7
00:00:15,429 --> 00:00:17,844
Algorithmic bias, like human bias,

8
00:00:17,844 --> 00:00:20,232
results in unfairness.

9
00:00:20,232 --> 00:00:22,601
However, algorithms, like viruses,

10
00:00:22,601 --> 00:00:25,799
can spread bias on a massive scale

11
00:00:25,799 --> 00:00:28,118
at a rapid pace.

12
00:00:28,118 --> 00:00:32,528
Algorithmic bias can also lead
to exclusionary experiences,

13
00:00:32,528 --> 00:00:34,596
and discriminatory practices.

14
00:00:34,596 --> 00:00:38,377
Let me show you what I mean.

15
00:00:38,377 --> 00:00:40,506
(Video) Joy Boulamwini: Camera.
I've got a face.

16
00:00:40,506 --> 00:00:42,704
Can you see my face?

17
00:00:42,704 --> 00:00:45,627
No glasses face.

18
00:00:45,627 --> 00:00:48,343
You can see her face.

19
00:00:48,343 --> 00:00:51,016
What about my face?

20
00:00:53,819 --> 00:00:54,699
(Laughter)

21
00:00:54,699 --> 00:00:58,886
I've got a mask. Can you see my mask?

22
00:00:58,886 --> 00:01:02,195
Joy Boulamwini: So how did this happen?

23
00:01:02,195 --> 00:01:04,711
Why am I sitting in front of a computer

24
00:01:04,711 --> 00:01:06,158
in a white mask

25
00:01:06,158 --> 00:01:09,831
trying to be detected by a cheap webcam?

26
00:01:09,831 --> 00:01:12,145
Well, when I'm not fighting the coded gaze

27
00:01:12,145 --> 00:01:13,549
as a poet of code,

28
00:01:13,549 --> 00:01:16,845
I'm a graduate student
at the MIT Media Lab,

29
00:01:16,845 --> 00:01:19,058
and there I have the opportunity

30
00:01:19,058 --> 00:01:21,786
to work on all sorts
of whimsical projects,

31
00:01:21,786 --> 00:01:24,095
including the Aspire Mirror,

32
00:01:24,095 --> 00:01:28,995
a project I did so I could project
digital masks onto my reflection.

33
00:01:28,995 --> 00:01:31,369
So in the morning, if I wanted
to feel powerful,

34
00:01:31,369 --> 00:01:32,827
I could put on a lion.

35
00:01:32,827 --> 00:01:36,347
If I wanted to be uplifted,
I might have a quote.

36
00:01:36,347 --> 00:01:39,430
So I used generic facial
recognition software

37
00:01:39,430 --> 00:01:40,735
to build the system,

38
00:01:40,735 --> 00:01:43,624
but found that it was
really hard to test it

39
00:01:43,624 --> 00:01:46,659
unless I wore a white mask.

40
00:01:46,659 --> 00:01:51,091
Unfortunately, I've run
into this issue before.

41
00:01:51,091 --> 00:01:55,519
When I was an undergraduate
at Georgia Tech studying computer science,

42
00:01:55,519 --> 00:01:57,598
I used to work on social robots,

43
00:01:57,598 --> 00:01:59,974
and one of my tasks was to get a robot

44
00:01:59,974 --> 00:02:01,399
to play peek-a-boo,

45
00:02:01,399 --> 00:02:03,268
a simple turn-taking game

46
00:02:03,268 --> 00:02:07,312
where partners cover their face
and then uncover it saying "Peek-a-boo."

47
00:02:07,312 --> 00:02:12,093
The problem is, peek-a-boo
doesn't really work if I can't see you,

48
00:02:12,093 --> 00:02:14,427
and my robot couldn't see me.

49
00:02:14,427 --> 00:02:18,401
But I borrowed my roommate's face
to get the project done,

50
00:02:18,401 --> 00:02:20,372
submitted the assignment, and figured,

51
00:02:20,372 --> 00:02:24,411
you know what? Somebody else
will solve this problem.

52
00:02:24,411 --> 00:02:26,299
Not too longer after,

53
00:02:26,299 --> 00:02:30,956
I was in Hong Kong
for an entrepreneurship competition.

54
00:02:30,956 --> 00:02:33,507
The organizers decided
to take participants

55
00:02:33,507 --> 00:02:35,993
on a tour of local start-ups.

56
00:02:35,993 --> 00:02:38,732
One of the start-ups had a social robot,

57
00:02:38,732 --> 00:02:40,668
and they decided to do a demo.

58
00:02:40,668 --> 00:02:43,672
The demo worked on everybody
until it got to me,

59
00:02:43,672 --> 00:02:45,741
and you can probably guess it.

60
00:02:45,741 --> 00:02:48,608
It couldn't detect my face.

61
00:02:48,608 --> 00:02:51,143
I asked the developers
what was going on,

62
00:02:51,143 --> 00:02:52,354
and it turned out

63
00:02:52,354 --> 00:02:56,729
we had used the same generic
facial recognition software.

64
00:02:56,729 --> 00:02:58,402
Halfway around the world,

65
00:02:58,402 --> 00:03:01,997
I learned that algorithmic bias
can travel as quickly

66
00:03:01,997 --> 00:03:05,262
as it takes to download
some files off of the Internet.

67
00:03:05,262 --> 00:03:07,406
So what's going on?
Why isn't my face being detected?

68
00:03:07,406 --> 00:03:12,765
Well, we have to look at
how we give machines sight.

69
00:03:12,765 --> 00:03:16,198
Computer vision uses
machine learning techniques

70
00:03:16,198 --> 00:03:18,382
to do facial recognition.

71
00:03:18,382 --> 00:03:20,658
So how this works is
you create a training set

72
00:03:20,658 --> 00:03:22,023
with examples of faces.

73
00:03:22,023 --> 00:03:24,931
This is face. This is face.
This is not a face.

74
00:03:24,931 --> 00:03:29,455
And over time, you can teach
a computer how to recognize other faces.

75
00:03:29,455 --> 00:03:33,263
However, if the training sets
aren't really that diverse,

76
00:03:33,263 --> 00:03:36,459
any face that deviates too much
from the established norm

77
00:03:36,459 --> 00:03:38,467
will be harder to detect,

78
00:03:38,467 --> 00:03:40,454
which is what was happening to me.

79
00:03:40,454 --> 00:03:42,702
But don't worry, there's some good news.

80
00:03:42,702 --> 00:03:45,730
Training sets don't just
materialize out of nowhere.

81
00:03:45,730 --> 00:03:47,467
We actually can create them.

82
00:03:47,467 --> 00:03:51,553
So there's an opportunity to create
full spectrum training sets

83
00:03:51,553 --> 00:03:55,515
that reflect a richer
portrait of humanity.

84
00:03:55,515 --> 00:03:57,797
Now you've seen in my examples

85
00:03:57,797 --> 00:03:59,552
how social robots

86
00:03:59,552 --> 00:04:04,246
was how I found out about exclusion
with algorithmic bias,

87
00:04:04,246 --> 00:04:10,160
but algorithmic bias can also lead
to discriminatory practices.

88
00:04:10,160 --> 00:04:13,955
Across the U.S., police departments
are starting to use

89
00:04:13,955 --> 00:04:15,676
facial recognition software

90
00:04:15,676 --> 00:04:18,289
in their crime-fighting arsenal.

91
00:04:18,289 --> 00:04:18,802
Georgetown Law published a report
showing that one in two

92
00:04:18,802 --> 00:04:19,052
adults in the U.S. --
that's 117 million people --

93
00:04:19,052 --> 00:04:30,180
have their faces
in facial recognition networks.

94
00:04:30,180 --> 00:04:35,117
Police departments can currently look
at these networks unregulated

95
00:04:35,117 --> 00:04:39,484
using algorithms that have not
been audited for accuracy.

96
00:04:39,484 --> 00:04:41,691
Yet we know facial recognition

97
00:04:41,691 --> 00:04:43,232
is not failproof,

98
00:04:43,232 --> 00:04:47,518
and labeling faces consistently
remains a challenge.

99
00:04:47,518 --> 00:04:49,277
You might have seen this on Facebook.

100
00:04:49,277 --> 00:04:52,316
My friends and I laugh all the time
when we see other people

101
00:04:52,316 --> 00:04:54,918
mislabeled in our photos.

102
00:04:54,918 --> 00:04:58,406
But misidentifying a suspected criminal

103
00:04:58,406 --> 00:05:00,547
is no laughing matter,

104
00:05:00,547 --> 00:05:03,364
nor is breaching civil liberties.

105
00:05:03,364 --> 00:05:06,411
Machine learning is being used
for facial recognition,

106
00:05:06,411 --> 00:05:09,083
but it's also extending beyond the realm

107
00:05:09,083 --> 00:05:11,934
of computer vision.

108
00:05:11,934 --> 00:05:15,865
In her book
"Weapons of Math Destruction,"

109
00:05:15,865 --> 00:05:18,178
data scientist Cathy O'Neil

110
00:05:18,178 --> 00:05:22,718
talks about the rising new WMDs --

111
00:05:22,718 --> 00:05:26,806
widespread, mysterious,
and destructive algorithms

112
00:05:26,806 --> 00:05:29,673
that are increasingly being used
to make decisions

113
00:05:29,673 --> 00:05:32,883
that impact more aspects of our lives.

114
00:05:32,883 --> 00:05:35,318
So who gets hired or fired?

115
00:05:35,318 --> 00:05:37,135
Do you get that loan?
Do you get insurance?

116
00:05:37,135 --> 00:05:40,730
Are you admitted into the college
that you wanted to get into?

117
00:05:40,730 --> 00:05:44,207
Do you and I pay the same price
for the same product

118
00:05:44,207 --> 00:05:46,749
purchased on the same platform?

119
00:05:46,749 --> 00:05:50,502
Law enforcement is also starting
to use machine learning

120
00:05:50,502 --> 00:05:52,769
for predictive policing.

121
00:05:52,769 --> 00:05:56,287
Some judges use machine-generated
risk scores to determine

122
00:05:56,287 --> 00:06:00,728
how long an individual
is going to spend in prison.

123
00:06:00,728 --> 00:06:03,221
So we really have to think
about these decisions.

124
00:06:03,221 --> 00:06:04,397
Are they fair?

125
00:06:04,397 --> 00:06:07,225
And we've seen that algorithmic bias

126
00:06:07,225 --> 00:06:10,709
doesn't necessarily always
lead to fair outcomes.

127
00:06:10,709 --> 00:06:12,697
So what can we do about it?

128
00:06:12,697 --> 00:06:16,401
Well, we can start thinking about
how we create more inclusive code

129
00:06:16,401 --> 00:06:19,415
and employ inclusive coding practices.

130
00:06:19,415 --> 00:06:22,087
It really starts with people.

131
00:06:22,087 --> 00:06:24,383
So who codes matters.

132
00:06:24,383 --> 00:06:28,288
Are we creating full spectrum teams
with diverse individuals

133
00:06:28,288 --> 00:06:30,811
who can check each other's blind spots?

134
00:06:30,811 --> 00:06:32,508
On the technical side,

135
00:06:32,508 --> 00:06:34,380
how we code matters.

136
00:06:34,380 --> 00:06:38,055
Are we factoring in fairness
as we're developing systems?

137
00:06:38,055 --> 00:06:41,452
And finally, why we code matters.

138
00:06:41,452 --> 00:06:44,157
We've used tools of computational creation

139
00:06:44,157 --> 00:06:46,335
to unlock immense wealth.

140
00:06:46,335 --> 00:06:48,391
We now have the opportunity

141
00:06:48,391 --> 00:06:51,088
to unlock even greater equality

142
00:06:51,088 --> 00:06:53,565
if we make social change a priority

143
00:06:53,565 --> 00:06:56,732
and not an afterthought.

144
00:06:56,732 --> 00:07:00,989
And so these are the three tenets
that will make up the incoding movement.

145
00:07:00,989 --> 00:07:02,677
Who codes matters,

146
00:07:02,677 --> 00:07:04,337
how we code matters,

147
00:07:04,337 --> 00:07:06,505
and why we code matters.

148
00:07:06,505 --> 00:07:09,507
So to go towards incoding,
we can start thinking about

149
00:07:09,507 --> 00:07:12,589
building platforms that can identify bias

150
00:07:12,589 --> 00:07:15,913
by collecting people's experiences
like the ones I shared,

151
00:07:15,913 --> 00:07:18,891
but also auditing existing software.

152
00:07:18,891 --> 00:07:22,680
We can also start to create
more inclusive training sets.

153
00:07:22,680 --> 00:07:25,575
Imagine a selfies for inclusion campaign

154
00:07:25,575 --> 00:07:27,915
where you and I can help developers

155
00:07:27,915 --> 00:07:31,383
test and create more
inclusive training sets.

156
00:07:31,383 --> 00:07:34,738
And we can also start
thinking more conscientiously

157
00:07:34,738 --> 00:07:40,189
about the social impact
of the technology that we're developing.

158
00:07:40,189 --> 00:07:42,526
To get the incoding movement started,

159
00:07:42,526 --> 00:07:45,470
I've launched the Algorithmic
Justice League,

160
00:07:45,470 --> 00:07:48,488
where anyone who cares about fairness

161
00:07:48,488 --> 00:07:51,372
can help fight the coded gaze.

162
00:07:51,372 --> 00:07:54,165
On codedgaze.com, you can report bias,

163
00:07:54,165 --> 00:07:56,070
request audits, become a tester,

164
00:07:56,070 --> 00:07:59,877
and join the ongoing conversation,

165
00:07:59,877 --> 00:08:03,326
#codedgaze.

166
00:08:03,326 --> 00:08:05,919
So I invite you to join me

167
00:08:05,919 --> 00:08:09,615
in creating a world where technology
works for all of us,

168
00:08:09,615 --> 00:08:11,678
not just some of us,

169
00:08:11,678 --> 00:08:16,069
a world where we value inclusion
and center social change.

170
00:08:16,069 --> 00:08:18,258
Thank you.

171
00:08:18,258 --> 00:08:23,003
(Applause)

172
00:08:23,533 --> 00:08:26,928
But I have one question.

173
00:08:26,928 --> 00:08:28,780
Will you join me in the fight?

174
00:08:28,780 --> 00:08:30,988
(Laughter)

175
00:08:30,988 --> 00:08:33,348
(Applause)