TED@BCG Milan
Margaret Mitchell: How we can build AI to help humans, not hurt us
Filmed:
Readability: 4.6
1,154,210 views
As a research scientist at Google, Margaret Mitchell helps develop computers that can communicate about what they see and understand. She tells a cautionary tale about the gaps, blind spots and biases we subconsciously encode into AI -- and asks us to consider what the technology we create today will mean for tomorrow. "All that we see now is a snapshot in the evolution of artificial intelligence," Mitchell says. "If we want AI to evolve in a way that helps humans, then we need to define the goals and strategies that enable that path now."
Margaret Mitchell - AI research scientist
Margaret Mitchell is a senior research scientist in Google's Research & Machine Intelligence group, working on artificial intelligence. Full bio
Margaret Mitchell is a senior research scientist in Google's Research & Machine Intelligence group, working on artificial intelligence. Full bio
Double-click the English transcript below to play the video.
00:13
I work on helping computers
communicate about the world around us.
communicate about the world around us.
0
1381
4015
00:17
There are a lot of ways to do this,
1
5754
1793
00:19
and I like to focus on helping computers
2
7571
2592
00:22
to talk about what they see
and understand.
and understand.
3
10187
2874
00:25
Given a scene like this,
4
13514
1571
00:27
a modern computer-vision algorithm
5
15109
1905
00:29
can tell you that there's a woman
and there's a dog.
and there's a dog.
6
17038
3095
00:32
It can tell you that the woman is smiling.
7
20157
2706
00:34
It might even be able to tell you
that the dog is incredibly cute.
that the dog is incredibly cute.
8
22887
3873
00:38
I work on this problem
9
26784
1349
00:40
thinking about how humans
understand and process the world.
understand and process the world.
10
28157
4212
00:45
The thoughts, memories and stories
11
33577
2952
00:48
that a scene like this
might evoke for humans.
might evoke for humans.
12
36553
2818
00:51
All the interconnections
of related situations.
of related situations.
13
39395
4285
00:55
Maybe you've seen
a dog like this one before,
a dog like this one before,
14
43704
3126
00:58
or you've spent time
running on a beach like this one,
running on a beach like this one,
15
46854
2969
01:01
and that further evokes thoughts
and memories of a past vacation,
and memories of a past vacation,
16
49847
4778
01:06
past times to the beach,
17
54649
1920
01:08
times spent running around
with other dogs.
with other dogs.
18
56593
2603
01:11
One of my guiding principles
is that by helping computers to understand
is that by helping computers to understand
19
59688
5207
01:16
what it's like to have these experiences,
20
64919
2896
01:19
to understand what we share
and believe and feel,
and believe and feel,
21
67839
5176
01:26
then we're in a great position
to start evolving computer technology
to start evolving computer technology
22
74094
4310
01:30
in a way that's complementary
with our own experiences.
with our own experiences.
23
78428
4587
01:35
So, digging more deeply into this,
24
83539
3387
01:38
a few years ago I began working on helping
computers to generate human-like stories
computers to generate human-like stories
25
86950
5905
01:44
from sequences of images.
26
92879
1666
01:47
So, one day,
27
95427
1904
01:49
I was working with my computer to ask it
what it thought about a trip to Australia.
what it thought about a trip to Australia.
28
97355
4622
01:54
It took a look at the pictures,
and it saw a koala.
and it saw a koala.
29
102768
2920
01:58
It didn't know what the koala was,
30
106236
1643
01:59
but it said it thought
it was an interesting-looking creature.
it was an interesting-looking creature.
31
107903
2999
02:04
Then I shared with it a sequence of images
about a house burning down.
about a house burning down.
32
112053
4004
02:09
It took a look at the images and it said,
33
117704
3285
02:13
"This is an amazing view!
This is spectacular!"
This is spectacular!"
34
121013
3500
02:17
It sent chills down my spine.
35
125450
2095
02:20
It saw a horrible, life-changing
and life-destroying event
and life-destroying event
36
128983
4572
02:25
and thought it was something positive.
37
133579
2382
02:27
I realized that it recognized
the contrast,
the contrast,
38
135985
3441
02:31
the reds, the yellows,
39
139450
2699
02:34
and thought it was something
worth remarking on positively.
worth remarking on positively.
40
142173
3078
02:37
And part of why it was doing this
41
145928
1615
02:39
was because most
of the images I had given it
of the images I had given it
42
147577
2945
02:42
were positive images.
43
150546
1840
02:44
That's because people
tend to share positive images
tend to share positive images
44
152903
3658
02:48
when they talk about their experiences.
45
156585
2190
02:51
When was the last time
you saw a selfie at a funeral?
you saw a selfie at a funeral?
46
159267
2541
02:55
I realized that,
as I worked on improving AI
as I worked on improving AI
47
163434
3095
02:58
task by task, dataset by dataset,
48
166553
3714
03:02
that I was creating massive gaps,
49
170291
2897
03:05
holes and blind spots
in what it could understand.
in what it could understand.
50
173212
3999
03:10
And while doing so,
51
178307
1334
03:11
I was encoding all kinds of biases.
52
179665
2483
03:15
Biases that reflect a limited viewpoint,
53
183029
3318
03:18
limited to a single dataset --
54
186371
2261
03:21
biases that can reflect
human biases found in the data,
human biases found in the data,
55
189283
3858
03:25
such as prejudice and stereotyping.
56
193165
3104
03:29
I thought back to the evolution
of the technology
of the technology
57
197554
3057
03:32
that brought me to where I was that day --
58
200635
2502
03:35
how the first color images
59
203966
2233
03:38
were calibrated against
a white woman's skin,
a white woman's skin,
60
206223
3048
03:41
meaning that color photography
was biased against black faces.
was biased against black faces.
61
209665
4145
03:46
And that same bias, that same blind spot
62
214514
2925
03:49
continued well into the '90s.
63
217463
1867
03:51
And the same blind spot
continues even today
continues even today
64
219701
3154
03:54
in how well we can recognize
different people's faces
different people's faces
65
222879
3698
03:58
in facial recognition technology.
66
226601
2200
04:01
I though about the state of the art
in research today,
in research today,
67
229323
3143
04:04
where we tend to limit our thinking
to one dataset and one problem.
to one dataset and one problem.
68
232490
4514
04:09
And that in doing so, we were creating
more blind spots and biases
more blind spots and biases
69
237688
4881
04:14
that the AI could further amplify.
70
242593
2277
04:17
I realized then
that we had to think deeply
that we had to think deeply
71
245712
2079
04:19
about how the technology we work on today
looks in five years, in 10 years.
looks in five years, in 10 years.
72
247815
5519
04:25
Humans evolve slowly,
with time to correct for issues
with time to correct for issues
73
253990
3142
04:29
in the interaction of humans
and their environment.
and their environment.
74
257156
3534
04:33
In contrast, artificial intelligence
is evolving at an incredibly fast rate.
is evolving at an incredibly fast rate.
75
261276
5429
04:39
And that means that it really matters
76
267013
1773
04:40
that we think about this
carefully right now --
carefully right now --
77
268810
2317
04:44
that we reflect on our own blind spots,
78
272180
3008
04:47
our own biases,
79
275212
2317
04:49
and think about how that's informing
the technology we're creating
the technology we're creating
80
277553
3857
04:53
and discuss what the technology of today
will mean for tomorrow.
will mean for tomorrow.
81
281434
3902
04:58
CEOs and scientists have weighed in
on what they think
on what they think
82
286593
3191
05:01
the artificial intelligence technology
of the future will be.
of the future will be.
83
289808
3325
05:05
Stephen Hawking warns that
84
293157
1618
05:06
"Artificial intelligence
could end mankind."
could end mankind."
85
294799
3007
05:10
Elon Musk warns
that it's an existential risk
that it's an existential risk
86
298307
2683
05:13
and one of the greatest risks
that we face as a civilization.
that we face as a civilization.
87
301014
3574
05:17
Bill Gates has made the point,
88
305665
1452
05:19
"I don't understand
why people aren't more concerned."
why people aren't more concerned."
89
307141
3185
05:23
But these views --
90
311412
1318
05:25
they're part of the story.
91
313618
1734
05:28
The math, the models,
92
316079
2420
05:30
the basic building blocks
of artificial intelligence
of artificial intelligence
93
318523
3070
05:33
are something that we call access
and all work with.
and all work with.
94
321617
3135
05:36
We have open-source tools
for machine learning and intelligence
for machine learning and intelligence
95
324776
3785
05:40
that we can contribute to.
96
328585
1734
05:42
And beyond that,
we can share our experience.
we can share our experience.
97
330919
3340
05:46
We can share our experiences
with technology and how it concerns us
with technology and how it concerns us
98
334760
3468
05:50
and how it excites us.
99
338252
1467
05:52
We can discuss what we love.
100
340251
1867
05:55
We can communicate with foresight
101
343244
2031
05:57
about the aspects of technology
that could be more beneficial
that could be more beneficial
102
345299
4857
06:02
or could be more problematic over time.
103
350180
2600
06:05
If we all focus on opening up
the discussion on AI
the discussion on AI
104
353799
4143
06:09
with foresight towards the future,
105
357966
1809
06:13
this will help create a general
conversation and awareness
conversation and awareness
106
361093
4270
06:17
about what AI is now,
107
365387
2513
06:21
what it can become
108
369212
2001
06:23
and all the things that we need to do
109
371237
1785
06:25
in order to enable that outcome
that best suits us.
that best suits us.
110
373046
3753
06:29
We already see and know this
in the technology that we use today.
in the technology that we use today.
111
377490
3674
06:33
We use smart phones
and digital assistants and Roombas.
and digital assistants and Roombas.
112
381767
3880
06:38
Are they evil?
113
386457
1150
06:40
Maybe sometimes.
114
388268
1547
06:42
Are they beneficial?
115
390664
1333
06:45
Yes, they're that, too.
116
393005
1533
06:48
And they're not all the same.
117
396236
1761
06:50
And there you already see
a light shining on what the future holds.
a light shining on what the future holds.
118
398489
3540
06:54
The future continues on
from what we build and create right now.
from what we build and create right now.
119
402942
3619
06:59
We set into motion that domino effect
120
407165
2642
07:01
that carves out AI's evolutionary path.
121
409831
2600
07:05
In our time right now,
we shape the AI of tomorrow.
we shape the AI of tomorrow.
122
413173
2871
07:08
Technology that immerses us
in augmented realities
in augmented realities
123
416566
3699
07:12
bringing to life past worlds.
124
420289
2566
07:15
Technology that helps people
to share their experiences
to share their experiences
125
423844
4312
07:20
when they have difficulty communicating.
126
428180
2262
07:23
Technology built on understanding
the streaming visual worlds
the streaming visual worlds
127
431323
4532
07:27
used as technology for self-driving cars.
128
435879
3079
07:32
Technology built on understanding images
and generating language,
and generating language,
129
440490
3413
07:35
evolving into technology that helps people
who are visually impaired
who are visually impaired
130
443927
4063
07:40
be better able to access the visual world.
131
448014
2800
07:42
And we also see how technology
can lead to problems.
can lead to problems.
132
450838
3261
07:46
We have technology today
133
454885
1428
07:48
that analyzes physical
characteristics we're born with --
characteristics we're born with --
134
456337
3835
07:52
such as the color of our skin
or the look of our face --
or the look of our face --
135
460196
3272
07:55
in order to determine whether or not
we might be criminals or terrorists.
we might be criminals or terrorists.
136
463492
3804
07:59
We have technology
that crunches through our data,
that crunches through our data,
137
467688
2905
08:02
even data relating
to our gender or our race,
to our gender or our race,
138
470617
2896
08:05
in order to determine whether or not
we might get a loan.
we might get a loan.
139
473537
2865
08:09
All that we see now
140
477494
1579
08:11
is a snapshot in the evolution
of artificial intelligence.
of artificial intelligence.
141
479097
3617
08:15
Because where we are right now,
142
483763
1778
08:17
is within a moment of that evolution.
143
485565
2238
08:20
That means that what we do now
will affect what happens down the line
will affect what happens down the line
144
488690
3802
08:24
and in the future.
145
492516
1200
08:26
If we want AI to evolve
in a way that helps humans,
in a way that helps humans,
146
494063
3951
08:30
then we need to define
the goals and strategies
the goals and strategies
147
498038
2801
08:32
that enable that path now.
148
500863
1733
08:35
What I'd like to see is something
that fits well with humans,
that fits well with humans,
149
503680
3738
08:39
with our culture and with the environment.
150
507442
2800
08:43
Technology that aids and assists
those of us with neurological conditions
those of us with neurological conditions
151
511435
4484
08:47
or other disabilities
152
515943
1721
08:49
in order to make life
equally challenging for everyone.
equally challenging for everyone.
153
517688
3216
08:54
Technology that works
154
522097
1421
08:55
regardless of your demographics
or the color of your skin.
or the color of your skin.
155
523542
3933
09:00
And so today, what I focus on
is the technology for tomorrow
is the technology for tomorrow
156
528383
4742
09:05
and for 10 years from now.
157
533149
1733
09:08
AI can turn out in many different ways.
158
536530
2634
09:11
But in this case,
159
539688
1225
09:12
it isn't a self-driving car
without any destination.
without any destination.
160
540937
3328
09:16
This is the car that we are driving.
161
544884
2400
09:19
We choose when to speed up
and when to slow down.
and when to slow down.
162
547953
3595
09:23
We choose if we need to make a turn.
163
551572
2400
09:26
We choose what the AI
of the future will be.
of the future will be.
164
554868
3000
09:31
There's a vast playing field
165
559186
1337
09:32
of all the things that artificial
intelligence can become.
intelligence can become.
166
560547
2965
09:36
It will become many things.
167
564064
1800
09:39
And it's up to us now,
168
567694
1732
09:41
in order to figure out
what we need to put in place
what we need to put in place
169
569450
3061
09:44
to make sure the outcomes
of artificial intelligence
of artificial intelligence
170
572535
3807
09:48
are the ones that will be
better for all of us.
better for all of us.
171
576366
3066
09:51
Thank you.
172
579456
1150
09:52
(Applause)
173
580630
2187
ABOUT THE SPEAKER
Margaret Mitchell - AI research scientistMargaret Mitchell is a senior research scientist in Google's Research & Machine Intelligence group, working on artificial intelligence.
Why you should listen
Margaret Mitchell's research involves vision-language and grounded language generation, focusing on how to evolve artificial intelligence towards positive goals. Her work combines computer vision, natural language processing, social media as well as many statistical methods and insights from cognitive science. Before Google, Mitchell was a founding member of Microsoft Research's "Cognition" group, focused on advancing artificial intelligence, and a researcher in Microsoft Research's Natural Language Processing group.
Margaret Mitchell | Speaker | TED.com