WEBVTT

00:00:04.545 --> 00:00:07.679
Hello, I'm Joy, a poet of code,

00:00:07.703 --> 00:00:12.696
on a mission to stop
an unseen force that's rising,

00:00:12.720 --> 00:00:15.576
a force that I called "the coded gaze,"

00:00:15.600 --> 00:00:18.909
my term for algorithmic bias.

00:00:18.933 --> 00:00:23.233
Algorithmic bias, like human bias,
results in unfairness.

00:00:23.257 --> 00:00:29.279
However, algorithms, like viruses,
can spread bias on a massive scale

00:00:29.303 --> 00:00:30.885
at a rapid pace.

00:00:31.447 --> 00:00:35.834
Algorithmic bias can also lead
to exclusionary experiences

00:00:35.858 --> 00:00:37.986
and discriminatory practices.

00:00:38.010 --> 00:00:40.071
Let me show you what I mean.

00:00:40.484 --> 00:00:42.920
(Video) Joy Buolamwini: Hi, camera.
I've got a face.

00:00:43.666 --> 00:00:45.530
Can you see my face?

00:00:45.555 --> 00:00:47.180
No-glasses face?

00:00:47.205 --> 00:00:49.419
You can see her face.

00:00:49.741 --> 00:00:51.986
What about my face?

00:00:55.394 --> 00:00:59.144
I've got a mask. Can you see my mask?

00:00:59.978 --> 00:01:02.343
Joy Buolamwini: So how did this happen?

00:01:02.367 --> 00:01:05.508
Why am I sitting in front of a computer

00:01:05.532 --> 00:01:06.956
in a white mask,

00:01:06.980 --> 00:01:10.630
trying to be detected by a cheap webcam?

00:01:10.654 --> 00:01:12.945
Well, when I'm not fighting the coded gaze

00:01:12.969 --> 00:01:14.489
as a poet of code,

00:01:14.513 --> 00:01:17.785
I'm a graduate student
at the MIT Media Lab,

00:01:17.809 --> 00:01:22.726
and there I have the opportunity to work
on all sorts of whimsical projects,

00:01:22.750 --> 00:01:24.777
including the Aspire Mirror,

00:01:24.801 --> 00:01:29.935
a project I did so I could project
digital masks onto my reflection.

00:01:29.959 --> 00:01:32.309
So in the morning, if I wanted
to feel powerful,

00:01:32.333 --> 00:01:33.767
I could put on a lion.

00:01:33.791 --> 00:01:37.287
If I wanted to be uplifted,
I might have a quote.

00:01:37.311 --> 00:01:40.300
So I used generic
facial recognition software

00:01:40.324 --> 00:01:41.675
to build the system,

00:01:41.699 --> 00:01:46.802
but found it was really hard to test it
unless I wore a white mask.

00:01:47.786 --> 00:01:52.132
Unfortunately, I've run
into this issue before.

00:01:52.156 --> 00:01:56.459
When I was an undergraduate
at Georgia Tech studying computer science,

00:01:56.483 --> 00:01:58.538
I used to work on social robots,

00:01:58.562 --> 00:02:02.339
and one of my tasks was to get a robot
to play peek-a-boo,

00:02:02.363 --> 00:02:04.046
a simple turn-taking game

00:02:04.070 --> 00:02:08.391
where partners cover their face
and then uncover it saying, "Peek-a-boo!"

00:02:08.415 --> 00:02:12.844
The problem is, peek-a-boo
doesn't really work if I can't see you,

00:02:12.868 --> 00:02:15.367
and my robot couldn't see me.

00:02:15.391 --> 00:02:19.341
But I borrowed my roommate's face
to get the project done,

00:02:19.365 --> 00:02:20.745
submitted the assignment,

00:02:20.769 --> 00:02:24.522
and figured, you know what,
somebody else will solve this problem.

00:02:25.173 --> 00:02:27.176
Not too long after,

00:02:27.200 --> 00:02:31.359
I was in Hong Kong
for an entrepreneurship competition.

00:02:31.843 --> 00:02:34.537
The organizers decided
to take participants

00:02:34.561 --> 00:02:36.933
on a tour of local start-ups.

00:02:36.957 --> 00:02:39.672
One of the start-ups had a social robot,

00:02:39.696 --> 00:02:41.608
and they decided to do a demo.

00:02:41.632 --> 00:02:44.612
The demo worked on everybody
until it got to me,

00:02:44.636 --> 00:02:46.559
and you can probably guess it.

00:02:46.583 --> 00:02:49.548
It couldn't detect my face.

00:02:49.572 --> 00:02:52.083
I asked the developers what was going on,

00:02:52.107 --> 00:02:57.640
and it turned out we had used the same
generic facial recognition software.

00:02:57.664 --> 00:02:59.314
Halfway around the world,

00:02:59.338 --> 00:03:03.190
I learned that algorithmic bias
can travel as quickly

00:03:03.214 --> 00:03:06.384
as it takes to download
some files off of the internet.

00:03:07.249 --> 00:03:10.325
So what's going on?
Why isn't my face being detected?

00:03:10.349 --> 00:03:13.705
Well, we have to look
at how we give machines sight.

00:03:13.729 --> 00:03:17.138
Computer vision uses
machine learning techniques

00:03:17.162 --> 00:03:19.042
to do facial recognition.

00:03:19.066 --> 00:03:22.963
So how this works is, you create
a training set with examples of faces.

00:03:22.987 --> 00:03:25.805
This is a face. This is a face.
This is not a face.

00:03:25.829 --> 00:03:30.348
And over time, you can teach a computer
how to recognize other faces.

00:03:30.372 --> 00:03:34.361
However, if the training sets
aren't really that diverse,

00:03:34.385 --> 00:03:37.734
any face that deviates too much
from the established norm

00:03:37.758 --> 00:03:39.407
will be harder to detect,

00:03:39.431 --> 00:03:41.394
which is what was happening to me.

00:03:41.418 --> 00:03:43.800
But don't worry -- there's some good news.

00:03:43.824 --> 00:03:46.595
Training sets don't just
materialize out of nowhere.

00:03:46.619 --> 00:03:48.407
We actually can create them.

00:03:48.431 --> 00:03:52.607
So there's an opportunity to create
full-spectrum training sets

00:03:52.631 --> 00:03:56.455
that reflect a richer
portrait of humanity.

00:03:56.479 --> 00:03:58.700
Now you've seen in my examples

00:03:58.724 --> 00:04:00.492
how social robots

00:04:00.516 --> 00:04:05.127
was how I found out about exclusion
with algorithmic bias.

00:04:05.151 --> 00:04:09.966
But algorithmic bias can also lead
to discriminatory practices.

00:04:10.941 --> 00:04:12.394
Across the US,

00:04:12.418 --> 00:04:16.616
police departments are starting to use
facial recognition software

00:04:16.640 --> 00:04:19.099
in their crime-fighting arsenal.

00:04:19.123 --> 00:04:21.136
Georgetown Law published a report

00:04:21.160 --> 00:04:27.923
showing that one in two adults
in the US -- that's 117 million people --

00:04:27.947 --> 00:04:31.481
have their faces
in facial recognition networks.

00:04:31.505 --> 00:04:36.057
Police departments can currently look
at these networks unregulated,

00:04:36.081 --> 00:04:40.367
using algorithms that have not
been audited for accuracy.

00:04:40.391 --> 00:04:44.255
Yet we know facial recognition
is not fail proof,

00:04:44.279 --> 00:04:48.458
and labeling faces consistently
remains a challenge.

00:04:48.482 --> 00:04:50.244
You might have seen this on Facebook.

00:04:50.268 --> 00:04:53.256
My friends and I laugh all the time
when we see other people

00:04:53.280 --> 00:04:55.738
mislabeled in our photos.

00:04:55.762 --> 00:05:01.353
But misidentifying a suspected criminal
is no laughing matter,

00:05:01.377 --> 00:05:04.204
nor is breaching civil liberties.

00:05:04.228 --> 00:05:07.433
Machine learning is being used
for facial recognition,

00:05:07.457 --> 00:05:11.962
but it's also extending beyond the realm
of computer vision.

00:05:12.770 --> 00:05:16.786
In her book, "Weapons
of Math Destruction,"

00:05:16.810 --> 00:05:23.491
data scientist Cathy O'Neil
talks about the rising new WMDs --

00:05:23.515 --> 00:05:27.868
widespread, mysterious
and destructive algorithms

00:05:27.892 --> 00:05:30.856
that are increasingly being used
to make decisions

00:05:30.880 --> 00:05:34.057
that impact more aspects of our lives.

00:05:34.081 --> 00:05:35.951
So who gets hired or fired?

00:05:35.975 --> 00:05:38.087
Do you get that loan?
Do you get insurance?

00:05:38.111 --> 00:05:41.614
Are you admitted into the college
you wanted to get into?

00:05:41.638 --> 00:05:45.147
Do you and I pay the same price
for the same product

00:05:45.171 --> 00:05:47.613
purchased on the same platform?

00:05:47.637 --> 00:05:51.396
Law enforcement is also starting
to use machine learning

00:05:51.420 --> 00:05:53.709
for predictive policing.

00:05:53.733 --> 00:05:57.227
Some judges use machine-generated
risk scores to determine

00:05:57.251 --> 00:06:01.653
how long an individual
is going to spend in prison.

00:06:01.677 --> 00:06:04.131
So we really have to think
about these decisions.

00:06:04.155 --> 00:06:05.337
Are they fair?

00:06:05.361 --> 00:06:08.251
And we've seen that algorithmic bias

00:06:08.275 --> 00:06:11.649
doesn't necessarily always
lead to fair outcomes.

00:06:11.673 --> 00:06:13.637
So what can we do about it?

00:06:13.661 --> 00:06:17.341
Well, we can start thinking about
how we create more inclusive code

00:06:17.365 --> 00:06:20.355
and employ inclusive coding practices.

00:06:20.379 --> 00:06:22.688
It really starts with people.

00:06:23.212 --> 00:06:25.173
So who codes matters.

00:06:25.197 --> 00:06:29.316
Are we creating full-spectrum teams
with diverse individuals

00:06:29.340 --> 00:06:31.751
who can check each other's blind spots?

00:06:31.775 --> 00:06:35.320
On the technical side,
how we code matters.

00:06:35.344 --> 00:06:38.995
Are we factoring in fairness
as we're developing systems?

00:06:39.019 --> 00:06:41.932
And finally, why we code matters.

00:06:42.289 --> 00:06:47.372
We've used tools of computational creation
to unlock immense wealth.

00:06:47.396 --> 00:06:51.843
We now have the opportunity
to unlock even greater equality

00:06:51.867 --> 00:06:54.797
if we make social change a priority

00:06:54.821 --> 00:06:56.991
and not an afterthought.

00:06:57.512 --> 00:07:02.034
And so these are the three tenets
that will make up the "incoding" movement.

00:07:02.058 --> 00:07:03.710
Who codes matters,

00:07:03.734 --> 00:07:05.277
how we code matters

00:07:05.301 --> 00:07:07.324
and why we code matters.

00:07:07.348 --> 00:07:10.447
So to go towards incoding,
we can start thinking about

00:07:10.471 --> 00:07:13.635
building platforms that can identify bias

00:07:13.659 --> 00:07:16.737
by collecting people's experiences
like the ones I shared,

00:07:16.761 --> 00:07:19.831
but also auditing existing software.

00:07:19.855 --> 00:07:23.620
We can also start to create
more inclusive training sets.

00:07:23.644 --> 00:07:26.447
Imagine a "Selfies for Inclusion" campaign

00:07:26.471 --> 00:07:30.126
where you and I can help
developers test and create

00:07:30.150 --> 00:07:32.243
more inclusive training sets.

00:07:32.806 --> 00:07:35.634
And we can also start thinking
more conscientiously

00:07:35.658 --> 00:07:41.049
about the social impact
of the technology that we're developing.

00:07:41.073 --> 00:07:43.466
To get the incoding movement started,

00:07:43.490 --> 00:07:46.337
I've launched the Algorithmic
Justice League,

00:07:46.361 --> 00:07:52.233
where anyone who cares about fairness
can help fight the coded gaze.

00:07:52.257 --> 00:07:55.553
On codedgaze.com, you can report bias,

00:07:55.577 --> 00:07:58.022
request audits, become a tester

00:07:58.046 --> 00:08:00.817
and join the ongoing conversation,

00:08:00.841 --> 00:08:03.128
#codedgaze.

00:08:04.246 --> 00:08:06.733
So I invite you to join me

00:08:06.757 --> 00:08:10.476
in creating a world where technology
works for all of us,

00:08:10.500 --> 00:08:12.397
not just some of us,

00:08:12.421 --> 00:08:17.009
a world where we value inclusion
and center social change.

00:08:17.033 --> 00:08:18.208
Thank you.

00:08:18.232 --> 00:08:22.503
(Applause)

00:08:24.377 --> 00:08:27.231
But I have one question:

00:08:27.255 --> 00:08:29.314
Will you join me in the fight?

00:08:29.338 --> 00:08:30.623
(Laughter)

00:08:30.647 --> 00:08:34.334
(Applause)