WEBVTT

00:00:13.961 --> 00:00:17.095
Hello, I'm Joy, a poet of code,

00:00:17.119 --> 00:00:22.112
on a mission to stop
an unseen force that's rising,

00:00:22.136 --> 00:00:24.992
a force that I called "the coded gaze,"

00:00:25.016 --> 00:00:28.325
my term for algorithmic bias.

NOTE Paragraph

00:00:28.349 --> 00:00:32.649
Algorithmic bias, like human bias,
results in unfairness.

00:00:32.673 --> 00:00:38.695
However, algorithms, like viruses,
can spread bias on a massive scale

00:00:38.719 --> 00:00:40.301
at a rapid pace.

00:00:40.863 --> 00:00:45.250
Algorithmic bias can also lead
to exclusionary experiences

00:00:45.274 --> 00:00:47.402
and discriminatory practices.

00:00:47.426 --> 00:00:49.487
Let me show you what I mean.

NOTE Paragraph

00:00:50.340 --> 00:00:52.776
(Video) Joy Boulamwini: Hi, camera.
I've got a face.

00:00:53.242 --> 00:00:55.107
Can you see my face?

00:00:55.131 --> 00:00:56.756
No-glasses face?

00:00:58.461 --> 00:01:00.207
You can see her face.

00:01:01.084 --> 00:01:02.691
What about my face?

00:01:05.874 --> 00:01:07.054
(Laughter)

00:01:07.078 --> 00:01:09.737
I've got a mask. Can you see my mask?

NOTE Paragraph

00:01:11.914 --> 00:01:14.279
Joy Boulamwini: So how did this happen?

00:01:14.303 --> 00:01:17.444
Why am I sitting in front of a computer

00:01:17.468 --> 00:01:18.892
in a white mask,

00:01:18.916 --> 00:01:22.566
trying to be detected by a cheap webcam?

00:01:22.590 --> 00:01:24.881
Well, when I'm not fighting the coded gaze

00:01:24.905 --> 00:01:26.425
as a poet of code,

00:01:26.449 --> 00:01:29.721
I'm a graduate student
at the MIT Media Lab,

00:01:29.745 --> 00:01:34.662
and there I have the opportunity to work
on all sorts of whimsical projects,

00:01:34.686 --> 00:01:36.713
including the Aspire Mirror,

00:01:36.737 --> 00:01:41.871
a project I did so I could project
digital masks onto my reflection.

00:01:41.895 --> 00:01:44.245
So in the morning, if I wanted
to feel powerful,

00:01:44.269 --> 00:01:45.703
I could put on a lion.

00:01:45.727 --> 00:01:49.223
If I wanted to be uplifted,
I might have a quote.

00:01:49.247 --> 00:01:52.236
So I used generic
facial recognition software

00:01:52.260 --> 00:01:53.611
to build the system,

00:01:53.635 --> 00:01:58.738
but found it was really hard to test it
unless I wore a white mask.

NOTE Paragraph

00:01:59.722 --> 00:02:04.068
Unfortunately, I've run
into this issue before.

00:02:04.092 --> 00:02:08.435
When I was an undergraduate
at Georgia Tech studying computer science,

00:02:08.459 --> 00:02:10.514
I used to work on social robots,

00:02:10.538 --> 00:02:14.315
and one of my tasks was to get a robot
to play peek-a-boo,

00:02:14.339 --> 00:02:16.272
a simple turn-taking game

00:02:16.296 --> 00:02:20.617
where partners cover their face
and then uncover it saying, "Peek-a-boo!"

00:02:20.741 --> 00:02:25.170
The problem is, peek-a-boo
doesn't really work if I can't see you,

00:02:25.194 --> 00:02:27.693
and my robot couldn't see me.

00:02:27.717 --> 00:02:31.667
But I borrowed my roommate's face
to get the project done,

00:02:31.691 --> 00:02:33.071
submitted the assignment,

00:02:33.095 --> 00:02:36.848
and figured, you know what,
somebody else will solve this problem.

NOTE Paragraph

00:02:37.499 --> 00:02:39.502
Not too long after,

00:02:39.526 --> 00:02:43.685
I was in Hong Kong
for an entrepreneurship competition.

00:02:44.169 --> 00:02:46.863
The organizers decided
to take participants

00:02:46.887 --> 00:02:49.259
on a tour of local start-ups.

00:02:49.283 --> 00:02:51.998
One of the start-ups had a social robot,

00:02:52.022 --> 00:02:53.934
and they decided to do a demo.

00:02:53.958 --> 00:02:56.938
The demo worked on everybody
until it got to me,

00:02:56.962 --> 00:02:58.885
and you can probably guess it.

00:02:58.909 --> 00:03:01.874
It couldn't detect my face.

00:03:01.898 --> 00:03:04.409
I asked the developers what was going on,

00:03:04.433 --> 00:03:09.966
and it turned out we had used the same
generic facial recognition software.

00:03:09.990 --> 00:03:11.640
Halfway around the world,

00:03:11.664 --> 00:03:15.516
I learned that algorithmic bias
can travel as quickly

00:03:15.540 --> 00:03:18.710
as it takes to download
some files off of the internet.

NOTE Paragraph

00:03:19.575 --> 00:03:22.651
So what's going on?
Why isn't my face being detected?

00:03:22.675 --> 00:03:26.031
Well, we have to look
at how we give machines sight.

00:03:26.055 --> 00:03:29.464
Computer vision uses
machine learning techniques

00:03:29.488 --> 00:03:31.368
to do facial recognition.

00:03:31.392 --> 00:03:35.289
So how this works is, you create
a training set with examples of faces.

00:03:35.313 --> 00:03:38.131
This is a face. This is a face.
This is not a face.

00:03:38.155 --> 00:03:42.674
And over time, you can teach a computer
how to recognize other faces.

00:03:42.698 --> 00:03:46.687
However, if the training sets
aren't really that diverse,

00:03:46.711 --> 00:03:50.060
any face that deviates too much
from the established norm

00:03:50.084 --> 00:03:51.733
will be harder to detect,

00:03:51.757 --> 00:03:53.720
which is what was happening to me.

NOTE Paragraph

00:03:53.744 --> 00:03:56.126
But don't worry -- there's some good news.

00:03:56.150 --> 00:03:58.921
Training sets don't just
materialize out of nowhere.

00:03:58.945 --> 00:04:00.733
We actually can create them.

00:04:00.757 --> 00:04:04.933
So there's an opportunity to create
full-spectrum training sets

00:04:04.957 --> 00:04:08.781
that reflect a richer
portrait of humanity.

NOTE Paragraph

00:04:08.805 --> 00:04:11.026
Now you've seen in my examples

00:04:11.050 --> 00:04:12.818
how social robots

00:04:12.842 --> 00:04:17.453
was how I found out about exclusion
with algorithmic bias.

00:04:17.477 --> 00:04:22.292
But algorithmic bias can also lead
to discriminatory practices.

00:04:23.267 --> 00:04:24.720
Across the US,

00:04:24.744 --> 00:04:28.942
police departments are starting to use
facial recognition software

00:04:28.966 --> 00:04:31.425
in their crime-fighting arsenal.

00:04:31.449 --> 00:04:33.462
Georgetown Law published a report

00:04:33.486 --> 00:04:40.249
showing that one in two adults
in the US -- that's 117 million people --

00:04:40.273 --> 00:04:43.807
have their faces
in facial recognition networks.

00:04:43.831 --> 00:04:48.383
Police departments can currently look
at these networks unregulated,

00:04:48.407 --> 00:04:52.693
using algorithms that have not
been audited for accuracy.

00:04:52.717 --> 00:04:56.581
Yet we know facial recognition
is not fail proof,

00:04:56.605 --> 00:05:00.784
and labeling faces consistently
remains a challenge.

00:05:00.808 --> 00:05:02.570
You might have seen this on Facebook.

00:05:02.594 --> 00:05:05.582
My friends and I laugh all the time
when we see other people

00:05:05.606 --> 00:05:08.064
mislabeled in our photos.

00:05:08.088 --> 00:05:13.679
But misidentifying a suspected criminal
is no laughing matter,

00:05:13.703 --> 00:05:16.530
nor is breaching civil liberties.

NOTE Paragraph

00:05:16.554 --> 00:05:19.759
Machine learning is being used
for facial recognition,

00:05:19.783 --> 00:05:24.288
but it's also extending beyond the realm
of computer vision.

00:05:25.096 --> 00:05:29.112
In her book, "Weapons
of Math Destruction,"

00:05:29.136 --> 00:05:35.817
data scientist Cathy O'Neil
talks about the rising new WMDs --

00:05:35.841 --> 00:05:40.194
widespread, mysterious
and destructive algorithms

00:05:40.218 --> 00:05:43.182
that are increasingly being used
to make decisions

00:05:43.206 --> 00:05:46.383
that impact more aspects of our lives.

00:05:46.407 --> 00:05:48.277
So who gets hired or fired?

00:05:48.301 --> 00:05:50.413
Do you get that loan?
Do you get insurance?

00:05:50.437 --> 00:05:53.940
Are you admitted into the college
you wanted to get into?

00:05:53.964 --> 00:05:57.473
Do you and I pay the same price
for the same product

00:05:57.497 --> 00:05:59.939
purchased on the same platform?

NOTE Paragraph

00:05:59.963 --> 00:06:03.722
Law enforcement is also starting
to use machine learning

00:06:03.746 --> 00:06:06.035
for predictive policing.

00:06:06.059 --> 00:06:09.553
Some judges use machine-generated
risk scores to determine

00:06:09.577 --> 00:06:13.979
how long an individual
is going to spend in prison.

00:06:14.003 --> 00:06:16.457
So we really have to think
about these decisions.

00:06:16.481 --> 00:06:17.663
Are they fair?

00:06:17.687 --> 00:06:20.577
And we've seen that algorithmic bias

00:06:20.601 --> 00:06:23.975
doesn't necessarily always
lead to fair outcomes.

NOTE Paragraph

00:06:23.999 --> 00:06:25.963
So what can we do about it?

00:06:25.987 --> 00:06:29.667
Well, we can start thinking about
how we create more inclusive code

00:06:29.691 --> 00:06:32.681
and employ inclusive coding practices.

00:06:32.705 --> 00:06:35.014
It really starts with people.

00:06:35.538 --> 00:06:37.499
So who codes matters.

00:06:37.523 --> 00:06:41.642
Are we creating full-spectrum teams
with diverse individuals

00:06:41.666 --> 00:06:44.077
who can check each other's blind spots?

00:06:44.101 --> 00:06:47.646
On the technical side,
how we code matters.

00:06:47.670 --> 00:06:51.321
Are we factoring in fairness
as we're developing systems?

00:06:51.345 --> 00:06:54.258
And finally, why we code matters.

00:06:54.615 --> 00:06:59.698
We've used tools of computational creation
to unlock immense wealth.

00:06:59.722 --> 00:07:04.169
We now have the opportunity
to unlock even greater equality

00:07:04.193 --> 00:07:07.123
if we make social change a priority

00:07:07.147 --> 00:07:09.317
and not an afterthought.

00:07:09.838 --> 00:07:14.360
And so these are the three tenets
that will make up the "incoding" movement.

00:07:14.384 --> 00:07:16.036
Who codes matters,

00:07:16.060 --> 00:07:17.603
how we code matters

00:07:17.627 --> 00:07:19.650
and why we code matters.

NOTE Paragraph

00:07:19.674 --> 00:07:22.773
So to go towards incoding,
we can start thinking about

00:07:22.797 --> 00:07:25.961
building platforms that can identify bias

00:07:25.985 --> 00:07:29.063
by collecting people's experiences
like the ones I shared,

00:07:29.087 --> 00:07:32.157
but also auditing existing software.

00:07:32.181 --> 00:07:35.946
We can also start to create
more inclusive training sets.

00:07:35.970 --> 00:07:38.773
Imagine a "Selfies for Inclusion" campaign

00:07:38.797 --> 00:07:42.452
where you and I can help
developers test and create

00:07:42.476 --> 00:07:44.569
more inclusive training sets.

00:07:45.132 --> 00:07:47.960
And we can also start thinking
more conscientiously

00:07:47.984 --> 00:07:53.375
about the social impact
of the technology that we're developing.

NOTE Paragraph

00:07:53.399 --> 00:07:55.792
To get the incoding movement started,

00:07:55.816 --> 00:07:58.663
I've launched the Algorithmic
Justice League,

00:07:58.687 --> 00:08:04.559
where anyone who cares about fairness
can help fight the coded gaze.

00:08:04.583 --> 00:08:07.879
On codedgaze.com, you can report bias,

00:08:07.903 --> 00:08:10.348
request audits, become a tester

00:08:10.372 --> 00:08:13.143
and join the ongoing conversation,

00:08:13.167 --> 00:08:15.454
#codedgaze.

NOTE Paragraph

00:08:16.572 --> 00:08:19.059
So I invite you to join me

00:08:19.083 --> 00:08:22.802
in creating a world where technology
works for all of us,

00:08:22.826 --> 00:08:24.723
not just some of us,

00:08:24.747 --> 00:08:29.395
a world where we value inclusion
and center social change.

NOTE Paragraph

00:08:29.419 --> 00:08:30.594
Thank you.

NOTE Paragraph

00:08:30.618 --> 00:08:35.912
(Applause)

NOTE Paragraph

00:08:36.763 --> 00:08:39.617
But I have one question:

00:08:39.641 --> 00:08:41.700
Will you join me in the fight?

NOTE Paragraph

00:08:41.724 --> 00:08:43.009
(Laughter)

NOTE Paragraph

00:08:43.033 --> 00:08:46.720
(Applause)