WEBVTT

00:00:01.041 --> 00:00:04.175
Hello, I'm Joy, a poet of code,

00:00:04.199 --> 00:00:09.192
on a mission to stop
an unseen force that's rising,

00:00:09.216 --> 00:00:12.072
a force that I called "the coded gaze,"

00:00:12.096 --> 00:00:15.405
my term for algorithmic bias.

NOTE Paragraph

00:00:15.429 --> 00:00:19.729
Algorithmic bias, like human bias,
results in unfairness.

00:00:19.753 --> 00:00:25.775
However, algorithms, like viruses,
can spread bias on a massive scale

00:00:25.799 --> 00:00:27.381
at a rapid pace.

00:00:27.943 --> 00:00:32.330
Algorithmic bias can also lead
to exclusionary experiences

00:00:32.354 --> 00:00:34.482
and discriminatory practices.

00:00:34.506 --> 00:00:36.567
Let me show you what I mean.

NOTE Paragraph

00:00:36.980 --> 00:00:39.416
(Video) Joy Buolamwini: Hi, camera.
I've got a face.

00:00:40.162 --> 00:00:42.026
Can you see my face?

00:00:42.051 --> 00:00:43.676
No-glasses face?

00:00:43.701 --> 00:00:45.915
You can see her face.

00:00:46.237 --> 00:00:48.482
What about my face?

00:00:51.890 --> 00:00:55.640
I've got a mask. Can you see my mask?

NOTE Paragraph

00:00:56.474 --> 00:00:58.839
Joy Buolamwini: So how did this happen?

00:00:58.863 --> 00:01:02.004
Why am I sitting in front of a computer

00:01:02.028 --> 00:01:03.452
in a white mask,

00:01:03.476 --> 00:01:07.126
trying to be detected by a cheap webcam?

00:01:07.150 --> 00:01:09.441
Well, when I'm not fighting the coded gaze

00:01:09.465 --> 00:01:10.985
as a poet of code,

00:01:11.009 --> 00:01:14.281
I'm a graduate student
at the MIT Media Lab,

00:01:14.305 --> 00:01:19.222
and there I have the opportunity to work
on all sorts of whimsical projects,

00:01:19.246 --> 00:01:21.273
including the Aspire Mirror,

00:01:21.297 --> 00:01:26.431
a project I did so I could project
digital masks onto my reflection.

00:01:26.455 --> 00:01:28.805
So in the morning, if I wanted
to feel powerful,

00:01:28.829 --> 00:01:30.263
I could put on a lion.

00:01:30.287 --> 00:01:33.783
If I wanted to be uplifted,
I might have a quote.

00:01:33.807 --> 00:01:36.796
So I used generic
facial recognition software

00:01:36.820 --> 00:01:38.171
to build the system,

00:01:38.195 --> 00:01:43.298
but found it was really hard to test it
unless I wore a white mask.

NOTE Paragraph

00:01:44.282 --> 00:01:48.628
Unfortunately, I've run
into this issue before.

00:01:48.652 --> 00:01:52.955
When I was an undergraduate
at Georgia Tech studying computer science,

00:01:52.979 --> 00:01:55.034
I used to work on social robots,

00:01:55.058 --> 00:01:58.835
and one of my tasks was to get a robot
to play peek-a-boo,

00:01:58.859 --> 00:02:00.542
a simple turn-taking game

00:02:00.566 --> 00:02:04.887
where partners cover their face
and then uncover it saying, "Peek-a-boo!"

00:02:04.911 --> 00:02:09.340
The problem is, peek-a-boo
doesn't really work if I can't see you,

00:02:09.364 --> 00:02:11.863
and my robot couldn't see me.

00:02:11.887 --> 00:02:15.837
But I borrowed my roommate's face
to get the project done,

00:02:15.861 --> 00:02:17.241
submitted the assignment,

00:02:17.265 --> 00:02:21.018
and figured, you know what,
somebody else will solve this problem.

NOTE Paragraph

00:02:21.669 --> 00:02:23.672
Not too long after,

00:02:23.696 --> 00:02:27.855
I was in Hong Kong
for an entrepreneurship competition.

00:02:28.339 --> 00:02:31.033
The organizers decided
to take participants

00:02:31.057 --> 00:02:33.429
on a tour of local start-ups.

00:02:33.453 --> 00:02:36.168
One of the start-ups had a social robot,

00:02:36.192 --> 00:02:38.104
and they decided to do a demo.

00:02:38.128 --> 00:02:41.108
The demo worked on everybody
until it got to me,

00:02:41.132 --> 00:02:43.055
and you can probably guess it.

00:02:43.079 --> 00:02:46.044
It couldn't detect my face.

00:02:46.068 --> 00:02:48.579
I asked the developers what was going on,

00:02:48.603 --> 00:02:54.136
and it turned out we had used the same
generic facial recognition software.

00:02:54.160 --> 00:02:55.810
Halfway around the world,

00:02:55.834 --> 00:02:59.686
I learned that algorithmic bias
can travel as quickly

00:02:59.710 --> 00:03:02.880
as it takes to download
some files off of the internet.

NOTE Paragraph

00:03:03.745 --> 00:03:06.821
So what's going on?
Why isn't my face being detected?

00:03:06.845 --> 00:03:10.201
Well, we have to look
at how we give machines sight.

00:03:10.225 --> 00:03:13.634
Computer vision uses
machine learning techniques

00:03:13.658 --> 00:03:15.538
to do facial recognition.

00:03:15.562 --> 00:03:19.459
So how this works is, you create
a training set with examples of faces.

00:03:19.483 --> 00:03:22.301
This is a face. This is a face.
This is not a face.

00:03:22.325 --> 00:03:26.844
And over time, you can teach a computer
how to recognize other faces.

00:03:26.868 --> 00:03:30.857
However, if the training sets
aren't really that diverse,

00:03:30.881 --> 00:03:34.230
any face that deviates too much
from the established norm

00:03:34.254 --> 00:03:35.903
will be harder to detect,

00:03:35.927 --> 00:03:37.890
which is what was happening to me.

NOTE Paragraph

00:03:37.914 --> 00:03:40.296
But don't worry -- there's some good news.

00:03:40.320 --> 00:03:43.091
Training sets don't just
materialize out of nowhere.

00:03:43.115 --> 00:03:44.903
We actually can create them.

00:03:44.927 --> 00:03:49.103
So there's an opportunity to create
full-spectrum training sets

00:03:49.127 --> 00:03:52.951
that reflect a richer
portrait of humanity.

NOTE Paragraph

00:03:52.975 --> 00:03:55.196
Now you've seen in my examples

00:03:55.220 --> 00:03:56.988
how social robots

00:03:57.012 --> 00:04:01.623
was how I found out about exclusion
with algorithmic bias.

00:04:01.647 --> 00:04:06.462
But algorithmic bias can also lead
to discriminatory practices.

00:04:07.437 --> 00:04:08.890
Across the US,

00:04:08.914 --> 00:04:13.112
police departments are starting to use
facial recognition software

00:04:13.136 --> 00:04:15.595
in their crime-fighting arsenal.

00:04:15.619 --> 00:04:17.632
Georgetown Law published a report

00:04:17.656 --> 00:04:24.419
showing that one in two adults
in the US -- that's 117 million people --

00:04:24.443 --> 00:04:27.977
have their faces
in facial recognition networks.

00:04:28.001 --> 00:04:32.553
Police departments can currently look
at these networks unregulated,

00:04:32.577 --> 00:04:36.863
using algorithms that have not
been audited for accuracy.

00:04:36.887 --> 00:04:40.751
Yet we know facial recognition
is not fail proof,

00:04:40.775 --> 00:04:44.954
and labeling faces consistently
remains a challenge.

00:04:44.978 --> 00:04:46.740
You might have seen this on Facebook.

00:04:46.764 --> 00:04:49.752
My friends and I laugh all the time
when we see other people

00:04:49.776 --> 00:04:52.234
mislabeled in our photos.

00:04:52.258 --> 00:04:57.849
But misidentifying a suspected criminal
is no laughing matter,

00:04:57.873 --> 00:05:00.700
nor is breaching civil liberties.

NOTE Paragraph

00:05:00.724 --> 00:05:03.929
Machine learning is being used
for facial recognition,

00:05:03.953 --> 00:05:08.458
but it's also extending beyond the realm
of computer vision.

00:05:09.266 --> 00:05:13.282
In her book, "Weapons
of Math Destruction,"

00:05:13.306 --> 00:05:19.987
data scientist Cathy O'Neil
talks about the rising new WMDs --

00:05:20.011 --> 00:05:24.364
widespread, mysterious
and destructive algorithms

00:05:24.388 --> 00:05:27.352
that are increasingly being used
to make decisions

00:05:27.376 --> 00:05:30.553
that impact more aspects of our lives.

00:05:30.577 --> 00:05:32.447
So who gets hired or fired?

00:05:32.471 --> 00:05:34.583
Do you get that loan?
Do you get insurance?

00:05:34.607 --> 00:05:38.110
Are you admitted into the college
you wanted to get into?

00:05:38.134 --> 00:05:41.643
Do you and I pay the same price
for the same product

00:05:41.667 --> 00:05:44.109
purchased on the same platform?

NOTE Paragraph

00:05:44.133 --> 00:05:47.892
Law enforcement is also starting
to use machine learning

00:05:47.916 --> 00:05:50.205
for predictive policing.

00:05:50.229 --> 00:05:53.723
Some judges use machine-generated
risk scores to determine

00:05:53.747 --> 00:05:58.149
how long an individual
is going to spend in prison.

00:05:58.173 --> 00:06:00.627
So we really have to think
about these decisions.

00:06:00.651 --> 00:06:01.833
Are they fair?

00:06:01.857 --> 00:06:04.747
And we've seen that algorithmic bias

00:06:04.771 --> 00:06:08.145
doesn't necessarily always
lead to fair outcomes.

NOTE Paragraph

00:06:08.169 --> 00:06:10.133
So what can we do about it?

00:06:10.157 --> 00:06:13.837
Well, we can start thinking about
how we create more inclusive code

00:06:13.861 --> 00:06:16.851
and employ inclusive coding practices.

00:06:16.875 --> 00:06:19.184
It really starts with people.

00:06:19.708 --> 00:06:21.669
So who codes matters.

00:06:21.693 --> 00:06:25.812
Are we creating full-spectrum teams
with diverse individuals

00:06:25.836 --> 00:06:28.247
who can check each other's blind spots?

00:06:28.271 --> 00:06:31.816
On the technical side,
how we code matters.

00:06:31.840 --> 00:06:35.491
Are we factoring in fairness
as we're developing systems?

00:06:35.515 --> 00:06:38.428
And finally, why we code matters.

00:06:38.785 --> 00:06:43.868
We've used tools of computational creation
to unlock immense wealth.

00:06:43.892 --> 00:06:48.339
We now have the opportunity
to unlock even greater equality

00:06:48.363 --> 00:06:51.293
if we make social change a priority

00:06:51.317 --> 00:06:53.487
and not an afterthought.

00:06:54.008 --> 00:06:58.530
And so these are the three tenets
that will make up the "incoding" movement.

00:06:58.554 --> 00:07:00.206
Who codes matters,

00:07:00.230 --> 00:07:01.773
how we code matters

00:07:01.797 --> 00:07:03.820
and why we code matters.

NOTE Paragraph

00:07:03.844 --> 00:07:06.943
So to go towards incoding,
we can start thinking about

00:07:06.967 --> 00:07:10.131
building platforms that can identify bias

00:07:10.155 --> 00:07:13.233
by collecting people's experiences
like the ones I shared,

00:07:13.257 --> 00:07:16.327
but also auditing existing software.

00:07:16.351 --> 00:07:20.116
We can also start to create
more inclusive training sets.

00:07:20.140 --> 00:07:22.943
Imagine a "Selfies for Inclusion" campaign

00:07:22.967 --> 00:07:26.622
where you and I can help
developers test and create

00:07:26.646 --> 00:07:28.739
more inclusive training sets.

00:07:29.302 --> 00:07:32.130
And we can also start thinking
more conscientiously

00:07:32.154 --> 00:07:37.545
about the social impact
of the technology that we're developing.

NOTE Paragraph

00:07:37.569 --> 00:07:39.962
To get the incoding movement started,

00:07:39.986 --> 00:07:42.833
I've launched the Algorithmic
Justice League,

00:07:42.857 --> 00:07:48.729
where anyone who cares about fairness
can help fight the coded gaze.

00:07:48.753 --> 00:07:52.049
On codedgaze.com, you can report bias,

00:07:52.073 --> 00:07:54.518
request audits, become a tester

00:07:54.542 --> 00:07:57.313
and join the ongoing conversation,

00:07:57.337 --> 00:07:59.624
#codedgaze.

NOTE Paragraph

00:08:00.742 --> 00:08:03.229
So I invite you to join me

00:08:03.253 --> 00:08:06.972
in creating a world where technology
works for all of us,

00:08:06.996 --> 00:08:08.893
not just some of us,

00:08:08.917 --> 00:08:13.505
a world where we value inclusion
and center social change.

NOTE Paragraph

00:08:13.529 --> 00:08:14.704
Thank you.

NOTE Paragraph

00:08:14.728 --> 00:08:18.999
(Applause)

NOTE Paragraph

00:08:20.873 --> 00:08:23.727
But I have one question:

00:08:23.751 --> 00:08:25.810
Will you join me in the fight?

NOTE Paragraph

00:08:25.834 --> 00:08:27.119
(Laughter)

NOTE Paragraph

00:08:27.143 --> 00:08:30.830
(Applause)