The era of blind faith in big data must end | Cathy O'Neil

242,918 views ใƒป 2017-09-07

TED


ืื ื ืœื—ืฅ ืคืขืžื™ื™ื ืขืœ ื”ื›ืชื•ื‘ื™ื•ืช ื‘ืื ื’ืœื™ืช ืœืžื˜ื” ื›ื“ื™ ืœื”ืคืขื™ืœ ืืช ื”ืกืจื˜ื•ืŸ.

ืชืจื’ื•ื: Nurit Noy ืขืจื™ื›ื”: Shlomo Adam
00:12
Algorithms are everywhere.
0
12795
1596
ื”ืืœื’ื•ืจื™ืชืžื™ื ื ืžืฆืื™ื ื‘ื›ืœ ืžืงื•ื.
00:15
They sort and separate the winners from the losers.
1
15931
3125
ื”ื ืžืžื™ื™ื ื™ื ืื ืฉื™ื ื•ืžืคืจื™ื“ื™ื ื‘ื™ืŸ ืžื ืฆื—ื™ื ืœืžืคืกื™ื“ื™ื.
00:19
The winners get the job
2
19839
2264
ื”ืžื ืฆื—ื™ื ื–ื•ื›ื™ื ื‘ืžืฉืจื” ื”ื ื—ืฉืงืช
00:22
or a good credit card offer.
3
22127
1743
ืื• ื‘ื”ืฆืขื” ืœื›ืจื˜ื™ืก ืืฉืจืื™ ื˜ื•ื‘.
00:23
The losers don't even get an interview
4
23894
2651
ื”ืžืคืกื™ื“ื™ื ืœื ื–ื•ื›ื™ื ืืคื™ืœื• ื‘ืจืื™ื•ืŸ
00:27
or they pay more for insurance.
5
27410
1777
ืื• ืžืฉืœืžื™ื ื™ื•ืชืจ ืขืœ ื”ื‘ื™ื˜ื•ื—.
00:30
We're being scored with secret formulas that we don't understand
6
30017
3549
ื ื•ืกื—ืื•ืช ืกื•ื“ื™ื•ืช ืฉืื™ื ื ื• ืžื‘ื™ื ื™ื ืžื“ืจื’ื•ืช ืื•ืชื ื•,
00:34
that often don't have systems of appeal.
7
34495
3217
ื•ื‘ื“ืจืš ื›ืœืœ ืื™ืŸ ืืคืฉืจื•ืช ืœืขืจืขืจ ืขืœ ื”ื—ืœื˜ื•ืชื™ื”ืŸ.
00:39
That begs the question:
8
39060
1296
ืžืชื‘ืงืฉืช ื”ืฉืืœื”:
00:40
What if the algorithms are wrong?
9
40380
2913
ืžื” ืื ื”ืืœื’ื•ืจื™ืชืžื™ื ื˜ื•ืขื™ื?
00:44
To build an algorithm you need two things:
10
44920
2040
ื›ื“ื™ ืœื‘ื ื•ืช ืืœื’ื•ืจื™ืชื ื ื—ื•ืฆื™ื ืฉื ื™ ื“ื‘ืจื™ื:
00:46
you need data, what happened in the past,
11
46984
1981
ื ืชื•ื ื™ื: ืžื” ืงืจื” ื‘ืขื‘ืจ,
00:48
and a definition of success,
12
48989
1561
ื•ื”ื’ื“ืจื” ืฉืœ ื”ืฆืœื—ื”,
00:50
the thing you're looking for and often hoping for.
13
50574
2457
ืžื” ืฉืืชื ืจื•ืฆื™ื ืื• ืžืงื•ื•ื™ื ืœื•.
00:53
You train an algorithm by looking, figuring out.
14
53055
5037
ื”ืืœื’ื•ืจื™ืชื ืœื•ืžื“ ืข"ื™...
00:58
The algorithm figures out what is associated with success.
15
58116
3419
ื”ืืœื’ื•ืจื™ืชื ืžื–ื”ื” ืžื” ืžืชืงืฉืจ ืœื”ืฆืœื—ื”.
01:01
What situation leads to success?
16
61559
2463
ืื™ืœื• ืžืฆื‘ื™ื ืžื•ื‘ื™ืœื™ื ืœื”ืฆืœื—ื”?
01:04
Actually, everyone uses algorithms.
17
64701
1762
ื‘ืขืฆื, ื›ื•ืœื ื• ืžืฉืชืžืฉื™ื ื‘ืืœื’ื•ืจื™ืชืžื™ื,
01:06
They just don't formalize them in written code.
18
66487
2718
ืื‘ืœ ืœื ืžื ืกื—ื™ื ืื•ืชื ื‘ืฆื•ืจืช ืงื•ื“ ื›ืชื•ื‘.
01:09
Let me give you an example.
19
69229
1348
ืืชืŸ ืœื›ื ื“ื•ื’ืžื”.
01:10
I use an algorithm every day to make a meal for my family.
20
70601
3316
ืื ื™ ืžืฉืชืžืฉืช ื‘ื›ืœ ื™ื•ื ื‘ืืœื’ื•ืจื™ืชื ื›ื“ื™ ืœื”ื›ื™ืŸ ืœืžืฉืคื—ืชื™ ืืจื•ื—ื”.
01:13
The data I use
21
73941
1476
ื”ื ืชื•ื ื™ื ื‘ื”ื ืื ื™ ืžืฉืชืžืฉืช
01:16
is the ingredients in my kitchen,
22
76214
1659
ื”ื ื”ืžื•ืฆืจื™ื ื‘ืžื˜ื‘ื— ืฉืœื™,
01:17
the time I have,
23
77897
1527
ื”ื–ืžืŸ ืฉืขื•ืžื“ ืœืจืฉื•ืชื™,
01:19
the ambition I have,
24
79448
1233
ื”ืฉืื™ืคื•ืช ืฉืœื™,
01:20
and I curate that data.
25
80705
1709
ื•ืื ื™ ืžืืจื’ื ืช ืืช ื”ื ืชื•ื ื™ื.
01:22
I don't count those little packages of ramen noodles as food.
26
82438
4251
ืื ื™ ืœื ืžื—ืฉื™ื‘ื” "ืžื ื” ื—ืžื”" ื›ืžื–ื•ืŸ.
01:26
(Laughter)
27
86713
1869
(ืฆื—ื•ืง)
01:28
My definition of success is:
28
88606
1845
ื”ื”ื’ื“ืจื” ืฉืœื™ ืœื”ืฆืœื—ื”:
01:30
a meal is successful if my kids eat vegetables.
29
90475
2659
ืืจื•ื—ื” ื ื—ืฉื‘ืช ืœืžื•ืฆืœื—ืช ืื ื”ื™ืœื“ื™ื ืฉืœื™ ืื•ื›ืœื™ื ื™ืจืงื•ืช.
01:34
It's very different from if my youngest son were in charge.
30
94001
2854
ืื ื‘ื ื™ ื”ืฆืขื™ืจ ื™ื”ื™ื” ืื—ืจืื™ ืœื›ืš ื–ื” ื™ื”ื™ื” ืื—ืจืช ืœื’ืžืจื™.
01:36
He'd say success is if he gets to eat lots of Nutella.
31
96879
2788
ื”ื•ื ื™ื’ื™ื“ ืฉื”ืฆืœื—ื” ืคื™ืจื•ืฉื” ืฉื”ื•ื ืื›ืœ ื”ืจื‘ื” ื—ืžืืช-ื‘ื•ื˜ื ื™ื.
01:40
But I get to choose success.
32
100999
2226
ืื‘ืœ ืื ื™ ื”ื™ื ื–ื• ืฉื‘ื•ื—ืจืช ืžื”ื™ ื”ืฆืœื—ื”.
01:43
I am in charge. My opinion matters.
33
103249
2707
ืื ื™ ื”ืื—ืจืื™ืช. ื”ื“ืขื” ืฉืœื™ ืงื•ื‘ืขืช.
01:45
That's the first rule of algorithms.
34
105980
2675
ื–ื”ื• ื”ื—ื•ืง ื”ืจืืฉื•ืŸ ืฉืœ ื”ืืœื’ื•ืจื™ืชืžื™ื.
01:48
Algorithms are opinions embedded in code.
35
108679
3180
ืืœื’ื•ืจื™ืชืžื™ื ื”ื ื“ืขื•ืช ืฉืžื•ื˜ืžืขื•ืช ื‘ืงื•ื“.
01:53
It's really different from what you think most people think of algorithms.
36
113382
3663
ื–ื” ืฉื•ื ื” ืžืื“ ืžืžื” ืฉืจื•ื‘ ื”ืื ืฉื™ื ื—ื•ืฉื‘ื™ื ืขืœ ืืœื’ื•ืจื™ืชืžื™ื.
01:57
They think algorithms are objective and true and scientific.
37
117069
4504
ื”ื ื—ื•ืฉื‘ื™ื ืฉื”ืืœื’ื•ืจื™ืชืžื™ื ื”ื ืื•ื‘ื™ื™ืงื˜ื™ื‘ื™ื™ื, ื ื›ื•ื ื™ื ื•ืžื“ืขื™ื™ื.
02:02
That's a marketing trick.
38
122207
1699
ื–ื• ืชื—ื‘ื•ืœื” ืฉื™ื•ื•ืงื™ืช.
02:05
It's also a marketing trick
39
125089
2125
ืชื—ื‘ื•ืœื” ืฉื™ื•ื•ืงื™ืช ื ื•ืกืคืช
02:07
to intimidate you with algorithms,
40
127238
3154
ื”ื™ื ืœื”ืคื—ื™ื“ ืืชื›ื ื‘ืืœื’ื•ืจื™ืชืžื™ื,
02:10
to make you trust and fear algorithms
41
130416
3661
ื›ื“ื™ ืฉืชื‘ื˜ื—ื• ื‘ื”ื ื•ืชื—ืฉืฉื• ืžื”ื
02:14
because you trust and fear mathematics.
42
134101
2018
ื›ื™ ืืชื ื‘ื•ื˜ื—ื™ื ื‘ืžืชืžื˜ื™ืงื” ื•ื—ื•ืฉืฉื™ื ืžืžื ื”.
02:17
A lot can go wrong when we put blind faith in big data.
43
137387
4830
ื”ืจื‘ื” ื™ื›ื•ืœ ืœื”ืฉืชื‘ืฉ ื›ืฉืื ื—ื ื• ื ื•ืชื ื™ื ืืžื•ืŸ ืขื™ื•ื•ืจ ื‘ื ืชื•ื ื™-ืขืชืง.
02:23
This is Kiri Soares. She's a high school principal in Brooklyn.
44
143504
3373
ื–ื• ืงื™ืจื™ ืกื•ืืจื–, ืžื ื”ืœืช ื‘ื™"ืก ืชื™ื›ื•ืŸ ื‘ื‘ืจื•ืงืœื™ืŸ.
02:26
In 2011, she told me her teachers were being scored
45
146901
2586
ื‘-2011 ื”ื™ื ืืžืจื” ืœื™ ืฉื”ืžื•ืจื™ื ืฉืœื” ืžื“ื•ืจื’ื™ื
02:29
with a complex, secret algorithm
46
149511
2727
ื‘ืขื–ืจืช ืืœื’ื•ืจื™ืชื ืกื•ื“ื™ ื•ืžื•ืจื›ื‘,
02:32
called the "value-added model."
47
152262
1489
ืฉื ืงืจื "ืžื•ื“ืœ ื”ืขืจืš ื”ืžื•ืกืฃ".
02:34
I told her, "Well, figure out what the formula is, show it to me.
48
154325
3092
ืืžืจืชื™ ืœื”, "ืชื‘ืจืจื™ ืžื”ื™ ื”ื ื•ืกื—ื” ื•ืชืจืื™ ืœื™ ืื•ืชื”.
"ืื ื™ ืืกื‘ื™ืจ ืœืš ืื•ืชื”"
02:37
I'm going to explain it to you."
49
157441
1541
ื”ื™ื ืืžืจื”, "ื ื™ืกื™ืชื™ ืœืงื‘ืœ ืืช ื”ื ื•ืกื—ื”.
02:39
She said, "Well, I tried to get the formula,
50
159006
2141
"ืืš ื‘ืžืฉืจื“ ื”ื—ื™ื ื•ืš ืืžืจื• ืœื™ ืฉื–ืืช ืžืชืžื˜ื™ืงื”,
02:41
but my Department of Education contact told me it was math
51
161171
2772
02:43
and I wouldn't understand it."
52
163967
1546
"ื•ืฉืื ื™ ืœื ืื‘ื™ืŸ ืื•ืชื”."
02:47
It gets worse.
53
167086
1338
ื–ื” ื ื”ื™ื” ื™ื•ืชืจ ื’ืจื•ืข.
02:48
The New York Post filed a Freedom of Information Act request,
54
168448
3530
ื”"ื ื™ื•-ื™ื•ืจืง ืคื•ืกื˜" ื”ื’ื™ืฉ ื‘ืงืฉื” ืœืคื™ ื—ื•ืง ื—ื•ืคืฉ ื”ืžื™ื“ืข.
02:52
got all the teachers' names and all their scores
55
172002
2959
ืงื™ื‘ืœ ืืช ื›ืœ ืฉืžื•ืช ื”ืžื•ืจื™ื ื•ื”ื“ื™ืจื•ื’ ืฉืœื”ื,
02:54
and they published them as an act of teacher-shaming.
56
174985
2782
ื•ืคื™ืจืกื ืื•ืชืŸ ื›ืฆืขื“ ืฉืœ ื‘ื™ื•ืฉ ืžื•ืจื™ื.
02:58
When I tried to get the formulas, the source code, through the same means,
57
178904
3860
ื›ืฉื ื™ืกื™ืชื™ ืœื”ืฉื™ื’ ืืช ื”ื ื•ืกื—ืื•ืช, ืืช ื”ืงื•ื“ ื”ืžืงื•ืจื™, ื‘ืื•ืชื ื”ืืžืฆืขื™ื,
03:02
I was told I couldn't.
58
182788
2149
ืืžืจื• ืœื™, "ืื™-ืืคืฉืจ".
03:04
I was denied.
59
184961
1236
ื“ื—ื• ืื•ืชื™.
03:06
I later found out
60
186221
1174
ืžืื•ื—ืจ ื™ื•ืชืจ ื’ื™ืœื™ืชื™
03:07
that nobody in New York City had access to that formula.
61
187419
2866
ืฉืœืืฃ ืื—ื“ ื‘ืขื™ืจ ื ื™ื•-ื™ื•ืจืง ืื™ืŸ ื’ื™ืฉื” ืœื ื•ืกื—ื” ื”ื”ื™ื.
03:10
No one understood it.
62
190309
1305
ืฉืื™ืฉ ืœื ืžื‘ื™ืŸ ืื•ืชื”.
03:13
Then someone really smart got involved, Gary Rubinstein.
63
193749
3224
ื•ืื– ื ื›ื ืก ืœืชืžื•ื ื” ืžื™ืฉื”ื• ืžืžืฉ ื—ื›ื. ื’ืจื™ ืจื•ื‘ื™ื ืฉื˜ื™ื™ืŸ.
03:16
He found 665 teachers from that New York Post data
64
196997
3621
ื”ื•ื ื–ื™ื”ื” ื‘ื ืชื•ื ื™ ื”"ื ื™ื•-ื™ื•ืจืง ืคื•ืกื˜" 665 ืžื•ืจื™ื ืขื ืฉื ื™ ื“ื™ืจื•ื’ื™ื.
03:20
that actually had two scores.
65
200642
1866
03:22
That could happen if they were teaching
66
202532
1881
ื–ื” ื™ื›ื•ืœ ื”ื™ื” ืœืงืจื•ืช ืื ื”ื ืœื™ืžื“ื• ืžืชืžื˜ื™ืงื” ื‘ื›ื™ืชื•ืช ื–' ื•ื’ื ื‘ื›ื™ืชื•ืช ื—'.
03:24
seventh grade math and eighth grade math.
67
204437
2439
03:26
He decided to plot them.
68
206900
1538
ื”ื•ื ื”ื—ืœื™ื˜ ืœื”ืฆื™ื’ ื–ืืช ื‘ื’ืจืฃ.
03:28
Each dot represents a teacher.
69
208462
1993
ื›ืœ ื ืงื•ื“ื” ืžืกืžืœืช ืžื•ืจื”.
03:30
(Laughter)
70
210924
2379
(ืฆื—ื•ืง)
03:33
What is that?
71
213327
1521
ืžื” ื–ื”?
03:34
(Laughter)
72
214872
1277
(ืฆื—ื•ืง)
03:36
That should never have been used for individual assessment.
73
216173
3446
ื–ื” ืœื ืžืฉื”ื• ืฉืืžื•ืจ ืœืฉืžืฉ ืœืฆื•ืจืš ื”ืขืจื›ื•ืช ืื™ืฉื™ื•ืช.
03:39
It's almost a random number generator.
74
219643
1926
ื–ื”ื• ื›ืžืขื˜ ืžื—ื•ืœืœ ืžืกืคืจื™ื ืืงืจืื™.
03:41
(Applause)
75
221593
2946
(ืžื—ื™ืื•ืช ื›ืคื™ื™ื)
03:44
But it was.
76
224563
1162
ืืš ื–ื” ืฉื™ืžืฉ ืœื›ืš.
03:45
This is Sarah Wysocki.
77
225749
1176
ื–ื•ื”ื™ ืฉืจื” ื•ื•ื™ืกื•ืงื™. ื”ื™ื ืคื•ื˜ืจื” ื™ื—ื“ ืขื ืขื•ื“ 205 ืžื•ืจื™ื
03:46
She got fired, along with 205 other teachers,
78
226949
2175
03:49
from the Washington, DC school district,
79
229148
2662
ืžื”ืžื—ื•ื– ื”ื‘ื™ืช-ืกื™ืคืจื™ ืฉืœ ื•ื•ืฉื™ื ื’ื˜ื•ืŸ ื”ื‘ื™ืจื”,
03:51
even though she had great recommendations from her principal
80
231834
2909
ืœืžืจื•ืช ืฉื”ื™ื• ืœื” ื”ืžืœืฆื•ืช ืžืขื•ืœื•ืช ืžื”ื ื”ืœืช ื‘ื™ื”"ืก
03:54
and the parents of her kids.
81
234767
1428
ื•ื’ื ืžื”ื”ื•ืจื™ื ืฉืœ ื”ื™ืœื“ื™ื ืฉืœื™ืžื“ื”.
03:57
I know what a lot of you guys are thinking,
82
237210
2032
ืื ื™ ื™ื•ื“ืขืช ืฉืจื‘ื™ื ืžื›ื ื—ื•ืฉื‘ื™ื,
ื‘ืžื™ื•ื—ื“ ื—ื•ืงืจื™ ื”ื ืชื•ื ื™ื ื•ืžื•ืžื—ื™ ื”ื‘ื™ื ื” ื”ืžืœืื›ื•ืชื™ืช ืฉื›ืืŸ,
03:59
especially the data scientists, the AI experts here.
83
239266
2487
04:01
You're thinking, "Well, I would never make an algorithm that inconsistent."
84
241777
4226
"ืื ื™ ืืฃ ืคืขื ืœื ืื›ืชื•ื‘ ืืœื’ื•ืจื™ืชื ื›ืœ-ื›ืš ืœื ืขื™ืงื‘ื™."
04:06
But algorithms can go wrong,
85
246673
1683
ืื‘ืœ ืืœื’ื•ืจื™ืชืžื™ื ื™ื›ื•ืœื™ื ืœื˜ืขื•ืช,
04:08
even have deeply destructive effects with good intentions.
86
248380
4598
ื•ืืคื™ืœื• ืœื’ืจื•ื ืœืชื•ืฆืื•ืช ื”ืจืกื ื™ื•ืช ื‘ื™ื•ืชืจ ืžืชื•ืš ื›ื•ื•ื ื•ืช ื˜ื•ื‘ื•ืช.
04:14
And whereas an airplane that's designed badly
87
254351
2379
ื•ื‘ืขื•ื“ ืฉืžื˜ื•ืก ืฉืชื•ื›ื ืŸ ื’ืจื•ืข ืžืชืจืกืง ื•ื›ื•ืœื ืจื•ืื™ื ื–ืืช,
04:16
crashes to the earth and everyone sees it,
88
256754
2001
04:18
an algorithm designed badly
89
258779
1850
ื”ืจื™ ื›ืฉืืœื’ื•ืจื™ืชื ืžืขื•ืฆื‘ ื’ืจื•ืข,
04:22
can go on for a long time, silently wreaking havoc.
90
262065
3865
ื”ื•ื ื™ื›ื•ืœ ืœืขื‘ื•ื“ ื”ืจื‘ื” ื–ืžืŸ ื•ืœื–ืจื•ืข ื‘ืฉืงื˜ ืชื•ื”ื•.
04:27
This is Roger Ailes.
91
267568
1570
ื–ื”ื• ืจื•ื’'ืจ ืื™ื™ืœืก.
04:29
(Laughter)
92
269162
2000
(ืฆื—ื•ืง)
04:32
He founded Fox News in 1996.
93
272344
2388
ื”ื•ื ื™ื™ืกื“ ืืช "ื—ื“ืฉื•ืช ืคื•ืงืก" ื‘-1996.
04:35
More than 20 women complained about sexual harassment.
94
275256
2581
ื™ื•ืชืจ ืž-20 ื ืฉื™ื ื”ืชืœื•ื ื ื• ืขืœ ื”ื˜ืจื“ื” ืžื™ื ื™ืช
04:37
They said they weren't allowed to succeed at Fox News.
95
277861
3235
ื•ืืžืจื• ืฉื”ืŸ ืœื ื”ื™ืจืฉื• ืœื”ืŸ ืœื”ืฆืœื™ื— ื‘"ื—ื“ืฉื•ืช ืคื•ืงืก".
04:41
He was ousted last year, but we've seen recently
96
281120
2520
ื”ื•ื ื”ื•ื“ื— ื‘ืฉื ื” ืฉืขื‘ืจื”, ืืš ืœืื—ืจื•ื ื” ื ื•ื“ืข ืœื ื•
04:43
that the problems have persisted.
97
283664
2670
ืฉื”ื‘ืขื™ื” ื ืžืฉื›ืช.
04:47
That begs the question:
98
287474
1400
ื ืฉืืœืช ื”ืฉืืœื”:
04:48
What should Fox News do to turn over another leaf?
99
288898
2884
ืžื” ืฆืจื™ื›ื” ืจืฉืช "ื—ื“ืฉื•ืช ืคื•ืงืก" ืœืขืฉื•ืช ื›ื“ื™ ืœืคืชื•ื— ื“ืฃ ื—ื“ืฉ?
04:53
Well, what if they replaced their hiring process
100
293065
3041
ืžื” ืื ื”ื ื™ื—ืœื™ืคื• ืืช ืชื”ืœื™ืš ื”ื”ืขืกืงื” ืฉืœื”ื
ื‘ืืœื’ื•ืจื™ืชื ืฉืœ ืœืžื™ื“ืช-ืžื›ื•ื ื”?
04:56
with a machine-learning algorithm?
101
296130
1654
04:57
That sounds good, right?
102
297808
1595
ื ืฉืžืข ื˜ื•ื‘, ื ื›ื•ืŸ?
04:59
Think about it.
103
299427
1300
ื—ื™ืฉื‘ื• ืขืœ ื–ื”.
05:00
The data, what would the data be?
104
300751
2105
ื”ื ืชื•ื ื™ื, ืžื” ื”ื ื™ื”ื™ื•?
05:02
A reasonable choice would be the last 21 years of applications to Fox News.
105
302880
4947
ื”ื’ื™ื•ื ื™ ืฉืืœื” ื™ื”ื™ื• ื ืชื•ื ื™ 21 ื”ืฉื ื™ื ื”ืื—ืจื•ื ื•ืช ืฉืœ ื‘ืงืฉื•ืช ืขื‘ื•ื“ื” ื‘"ื—ื“ืฉื•ืช ืคื•ืงืก".
05:07
Reasonable.
106
307851
1502
ื”ื’ื™ื•ื ื™.
05:09
What about the definition of success?
107
309377
1938
ืžื” ืœื’ื‘ื™ ื”ื”ื’ื“ืจื” ืœื”ืฆืœื—ื”?
05:11
Reasonable choice would be,
108
311741
1324
ื‘ื—ื™ืจื” ื”ื’ื™ื•ื ื™ืช ืชื”ื™ื”,
05:13
well, who is successful at Fox News?
109
313089
1778
ืžื™ ืžืฆืœื™ื— ื‘"ื—ื“ืฉื•ืช ืคื•ืงืก"?
05:14
I guess someone who, say, stayed there for four years
110
314891
3580
ืื•ืœื™ ืžื™ืฉื”ื• ืฉืขื•ื‘ื“ ืฉื ื›ื‘ืจ 4 ืฉื ื™ื,
05:18
and was promoted at least once.
111
318495
1654
ื•ืงื™ื‘ืœ ืงื™ื“ื•ื ืœืคื—ื•ืช ืคืขื ืื—ืช.
05:20
Sounds reasonable.
112
320636
1561
ื ืฉืžืข ื”ื’ื™ื•ื ื™.
05:22
And then the algorithm would be trained.
113
322221
2354
ื•ืื– ื”ืืœื’ื•ืจื™ืชื ื™ืขื‘ื•ืจ ืœื™ืžื•ื“.
05:24
It would be trained to look for people to learn what led to success,
114
324599
3877
ื”ื•ื ื™ืœืžื“ ืœื—ืคืฉ ืื ืฉื™ื ื›ื“ื™ ืœืœืžื•ื“ ืžื” ื”ื•ื‘ื™ืœ ืœื”ืฆืœื—ื”,
05:29
what kind of applications historically led to success
115
329039
4318
ืื™ืœื• ืžื•ืขืžื“ื™ื ื”ืคื›ื• ืœืขื•ื‘ื“ื™ื ืžื•ืฆืœื—ื™ื,
05:33
by that definition.
116
333381
1294
ืœืคื™ ื”ื”ื’ื“ืจื” ื”ื–ื•.
ืขื›ืฉื™ื• ื—ื™ืฉื‘ื• ืžื” ื™ืงืจื” ืื ื ื™ื™ืฉื ื–ืืช ืœืžืื’ืจ ืžื•ืขืžื“ื™ื ื‘ื”ื•ื•ื”:
05:36
Now think about what would happen
117
336020
1775
05:37
if we applied that to a current pool of applicants.
118
337819
2555
05:40
It would filter out women
119
340939
1629
ื”ืืœื’ื•ืจื™ืชื ื™ืกื ืŸ ื”ื—ื•ืฆื” ื ืฉื™ื,
05:43
because they do not look like people who were successful in the past.
120
343483
3930
ื›ื™ ื”ืŸ ืื™ื ืŸ ื“ื•ืžื•ืช ืœืื ืฉื™ื ืฉื”ืฆืœื™ื—ื• ื‘ืขื‘ืจ.
05:51
Algorithms don't make things fair
121
351572
2537
ื”ืืœื’ื•ืจื™ืชืžื™ื ืื™ื ื ืžืชืงื ื™ื ืืช ื”ืขื•ืœื
05:54
if you just blithely, blindly apply algorithms.
122
354133
2694
ืื ืžื™ื™ืฉืžื™ื ืื•ืชื ื‘ืฉืžื—ื” ื•ื‘ืขื™ื•ื•ืจื•ืŸ
05:56
They don't make things fair.
123
356851
1482
ื”ื ืœื ืžืชืงื ื™ื ืืช ื”ืขื•ืœื
05:58
They repeat our past practices,
124
358357
2128
ืืœื ืจืง ื—ื•ื–ืจื™ื ืขืœ ืžื” ืฉืขืฉื™ื ื• ื‘ืขื‘ืจ,
06:00
our patterns.
125
360509
1183
ืขืœ ื”ื“ืคื•ืกื™ื ืฉืœื ื•.
06:01
They automate the status quo.
126
361716
1939
ื”ื ื”ื•ืคื›ื™ื ืืช ื”ืžืฆื‘ ื”ืงื™ื™ื ืœืื•ื˜ื•ืžื˜ื™.
06:04
That would be great if we had a perfect world,
127
364538
2389
ื”ื™ื” ื ื”ื“ืจ ืื ื”ื™ื” ืœื ื• ืขื•ืœื ืžื•ืฉืœื,
06:07
but we don't.
128
367725
1312
ืื‘ืœ ืื™ืŸ ืœื ื•.
06:09
And I'll add that most companies don't have embarrassing lawsuits,
129
369061
4102
ื•ืื ื™ ืื•ืกื™ืฃ ืฉืจื•ื‘ ื”ื—ื‘ืจื•ืช ืœื ืžืชืžื•ื“ื“ื•ืช ืขื ืชื‘ื™ืขื•ืช ืžื‘ื™ื›ื•ืช,
06:14
but the data scientists in those companies
130
374266
2588
ืื‘ืœ ืžื•ืžื—ื™ ื”ื ืชื•ื ื™ื ื‘ื—ื‘ืจื•ืช ืืœื” ืžื—ื•ื™ื™ื‘ื™ื ืœืฆื™ื™ืช ืœื ืชื•ื ื™ื,
06:16
are told to follow the data,
131
376878
2189
06:19
to focus on accuracy.
132
379091
2143
ืœื”ืชืžืงื“ ื‘ื“ื™ื•ืง.
06:22
Think about what that means.
133
382093
1381
ื—ื™ืฉื‘ื• ืžื” ื–ื” ืื•ืžืจ,
06:23
Because we all have bias, it means they could be codifying sexism
134
383498
4027
ื”ืจื™ ืœื›ื•ืœื ื• ื™ืฉ ื”ื˜ื™ื•ืช.
ืื•ืœื™ ื”ื ืžืชื›ื ืชื™ื ืœืืคืœื™ื” ืขืœ ืจืงืข ืžื™ืŸ,
06:27
or any other kind of bigotry.
135
387549
1836
ืื• ื›ืœ ืกื•ื’ ืื—ืจ ืฉืœ ื’ื–ืขื ื•ืช.
06:31
Thought experiment,
136
391308
1421
ื ื™ืกื•ื™ ืžื—ืฉื‘ืชื™,
06:32
because I like them:
137
392753
1509
ื›ื™ ืื ื™ ืื•ื”ื‘ืช ื›ืืœื”:
06:35
an entirely segregated society --
138
395394
2975
ื“ืžื™ื™ื ื• ื—ื‘ืจื” ืฉืœืžื” ืฉืžื•ืคืจื“ืช ืœืคื™ ื’ื–ืขื™ื:
06:40
racially segregated, all towns, all neighborhoods
139
400067
3328
ื›ืœ ื”ืขื™ื™ืจื•ืช, ื›ืœ ื”ืฉื›ื•ื ื•ืช,
06:43
and where we send the police only to the minority neighborhoods
140
403419
3037
ื•ืืช ื”ืžืฉื˜ืจื” ืฉื•ืœื—ื™ื ืœื—ืคืฉ ืคืฉื™ืขื” ืจืง ื‘ืฉื›ื•ื ื•ืช ืฉืœ ืžื™ืขื•ื˜ื™ื.
06:46
to look for crime.
141
406480
1193
06:48
The arrest data would be very biased.
142
408271
2219
ื ืชื•ื ื™ ื”ืžืขืฆืจื™ื ื™ื”ื™ื• ืžื•ื˜ื™ื ื‘ืฆื•ืจื” ืžื•ื‘ื”ืงืช.
06:51
What if, on top of that, we found the data scientists
143
411671
2575
ืžื” ืื ื‘ื ื•ืกืฃ, ืžืฆืื ื• ืžื•ืžื—ื” ืœื ืชื•ื ื™ื
06:54
and paid the data scientists to predict where the next crime would occur?
144
414270
4161
ื•ืฉื™ืœืžื ื• ืœื• ื›ื“ื™ ืฉื™ื ื‘ื ืื™ืคื” ื™ืงืจื” ื”ืคืฉืข ื”ื‘ื?
06:59
Minority neighborhood.
145
419095
1487
ืฉื›ื•ื ืช ืžื™ืขื•ื˜ื™ื.
07:01
Or to predict who the next criminal would be?
146
421105
3125
ืื• ื›ื“ื™ ืฉื™ื ื‘ื ืžื™ ื™ื”ื™ื” ื”ืคื•ืฉืข ื”ื‘ื?
07:04
A minority.
147
424708
1395
ื‘ืŸ ืžื™ืขื•ื˜ื™ื.
07:07
The data scientists would brag about how great and how accurate
148
427769
3541
ืžื•ืžื—ื™ ื”ื ืชื•ื ื™ื ื™ืชืคืืจื• ื›ืžื” ื ื”ื“ืจ ื•ืžื“ื•ื™ื™ืง ื”ืžื•ื“ืœ ืฉืœื”ื.
07:11
their model would be,
149
431334
1297
07:12
and they'd be right.
150
432655
1299
ื•ื”ื ื™ืฆื“ืงื•.
07:15
Now, reality isn't that drastic, but we do have severe segregations
151
435771
4615
ื”ืžืฆื™ืื•ืช ืœื ื›ืœ ื›ืš ื“ืจืกื˜ื™ืช, ืื‘ืœ ื™ืฉ ืœื ื• ื‘ืืžืช ื”ืคืจื“ื” ื—ืžื•ืจื”
07:20
in many cities and towns,
152
440410
1287
ื‘ืขืจื™ื ื•ืขื™ื™ืจื•ืช ืจื‘ื•ืช,
07:21
and we have plenty of evidence
153
441721
1893
ื•ื™ืฉ ืœื ื• ืฉืคืข ืจืื™ื•ืช
07:23
of biased policing and justice system data.
154
443638
2688
ืœื”ื˜ื™ื•ืช ื‘ื ืชื•ื ื™ื ื”ืžืฉื˜ืจืชื™ื™ื ืžื•ื˜ื™ื ื•ื‘ืžืขืจื›ืช ื”ืžืฉืคื˜.
07:27
And we actually do predict hotspots,
155
447452
2815
ื•ืื ื—ื ื• ืื›ืŸ ื—ื•ื–ื™ื ื ืงื•ื“ื•ืช ืกื™ื›ื•ืŸ,
07:30
places where crimes will occur.
156
450291
1530
ืžืงื•ืžื•ืช ื‘ื”ื ื™ืงืจื• ืคืฉืขื™ื.
07:32
And we do predict, in fact, the individual criminality,
157
452221
3866
ื•ืื ื—ื ื• ื’ื ืžื ื‘ืื™ื ืืช ืžื™ื“ืช ื”ื ื˜ื™ื” ื”ืื™ืฉื™ืช ืœืคืฉื•ืข.
07:36
the criminality of individuals.
158
456111
1770
ืืช ื ื˜ื™ื™ืชื ืฉืœ ืื ืฉื™ื ืžืกื•ื™ื™ืžื™ื ืœืคืฉื•ืข.
07:38
The news organization ProPublica recently looked into
159
458792
3963
ืกื•ื›ื ื•ืช ื”ื—ื“ืฉื•ืช "ืคืจื•ืคื‘ืœื™ืงื”" ื‘ื—ื ื” ืœืื—ืจื•ื ื”
07:42
one of those "recidivism risk" algorithms,
160
462779
2024
ืื—ื“ ืื•ืชื ืืœื’ื•ืจื™ืชืžื™ื ืœ"ื ื™ื‘ื•ื™ ื”ื™ืฉื ื•ืช ืคืฉื™ืขื”"
07:44
as they're called,
161
464827
1163
ื›ืžื• ืฉืงื•ืจืื™ื ืœื”ื.
07:46
being used in Florida during sentencing by judges.
162
466014
3194
ื‘ืคืœื•ืจื™ื“ื” ืžืฉืชืžืฉื™ื ื‘ื”ื ืฉื•ืคื˜ื™ื ื‘ื–ืžืŸ ื—ืจื™ืฆืช ื’ื–ืจ ื”ื“ื™ืŸ.
07:50
Bernard, on the left, the black man, was scored a 10 out of 10.
163
470231
3585
ื‘ืจื ืจื“, ืžืฉืžืืœ, ื”ื’ื‘ืจ ื”ืฉื—ื•ืจ ืงื™ื‘ืœ 10 ื ืงื•ื“ื•ืช ืžืชืš 10.
07:54
Dylan, on the right, 3 out of 10.
164
474999
2007
ื“ื™ืœืŸ, ืžื™ืžื™ืŸ - 3 ืžืชื•ืš 10.
07:57
10 out of 10, high risk. 3 out of 10, low risk.
165
477030
2501
10 ืžืชื•ืš 10 - ืกื™ื›ื•ืŸ ื’ื‘ื•ื”. 3 ืžืชื•ืš 10 - ืกื™ื›ื•ืŸ ื ืžื•ืš.
08:00
They were both brought in for drug possession.
166
480418
2385
ืฉื ื™ื”ื ื ืขืฆืจื• ืขืœ ื”ื—ื–ืงืช ืกืžื™ื.
08:02
They both had records,
167
482827
1154
ืœืฉื ื™ื”ื ื”ื™ื” ื›ื‘ืจ ืชื™ืง.
08:04
but Dylan had a felony
168
484005
2806
ืื‘ืœ ื“ื™ืœืŸ ืขื‘ืจ ืขื‘ื™ืจื”
08:06
but Bernard didn't.
169
486835
1176
ื•ื‘ืจื ืจื“ - ืœื.
08:09
This matters, because the higher score you are,
170
489638
3066
ื–ื” ืžืฉื ื”, ื›ื™ ื›ื›ืœ ืฉืชืงื‘ืœ ื ื™ืงื•ื“ ื™ื•ืชืจ ื’ื‘ื•ื”,
08:12
the more likely you're being given a longer sentence.
171
492728
3473
ื’ื•ื‘ืจ ื”ืกื™ื›ื•ื™ ืฉืชืงื‘ืœ ืขื•ื ืฉ ืžืืกืจ ืืจื•ืš ื™ื•ืชืจ.
08:18
What's going on?
172
498114
1294
ืžื” ืงื•ืจื” ืคื”?
08:20
Data laundering.
173
500346
1332
ื”ืœื‘ื ืช ื ืชื•ื ื™ื.
08:22
It's a process by which technologists hide ugly truths
174
502750
4427
ื–ื”ื• ืชื”ืœื™ืš ืฉื‘ื• ืื ืฉื™ ื”ื˜ื›ื ื•ืœื•ื’ื™ื” ืžืกืชื™ืจื™ื ืืžื™ืชื•ืช ืžื›ื•ืขืจื•ืช
08:27
inside black box algorithms
175
507201
1821
ื‘ืชื•ืš ืืœื’ื•ืจื™ืชืžื™ื ื—ืชื•ืžื™ื
ื•ืื•ืžืจื™ื ืฉื”ื "ืื•ื‘ื™ื™ืงื˜ื™ื‘ื™ื™ื",
08:29
and call them objective;
176
509046
1290
08:31
call them meritocratic.
177
511140
1568
ืฉื–ืืช ืžืจื™ื˜ื•ืงืจื˜ื™ื”.
08:34
When they're secret, important and destructive,
178
514938
2385
ื˜ื‘ืขืชื™ ื›ื™ื ื•ื™ ืœืืœื’ื•ืจื™ืชืžื™ื ืกื•ื“ื™ื™ื, ื—ืฉื•ื‘ื™ื ื•ื”ืจืกื ื™ื™ื ืืœื•:
08:37
I've coined a term for these algorithms:
179
517347
2487
08:39
"weapons of math destruction."
180
519858
1999
"ื ืฉืง ืœื”ืฉืžื“ื” ืžืชืžื˜ื™ืช".
08:41
(Laughter)
181
521881
1564
(ืฆื—ื•ืง)
08:43
(Applause)
182
523469
3054
(ืžื—ื™ืื•ืช ื›ืคื™ื™ื)
08:46
They're everywhere, and it's not a mistake.
183
526547
2354
ื”ื ื‘ื›ืœ ืžืงื•ื ื•ื–ื• ืœื ื˜ืขื•ืช:
08:49
These are private companies building private algorithms
184
529515
3723
ืžื“ื•ื‘ืจ ื‘ื—ื‘ืจื•ืช ืคืจื˜ื™ื•ืช ืฉื›ื•ืชื‘ื•ืช ืืœื’ื•ืจื™ืชืžื™ื ืคืจื˜ื™ื™ื
08:53
for private ends.
185
533262
1392
ืœืฆืจื›ื™ื”ืŸ ื”ืคืจื˜ื™ื™ื.
08:55
Even the ones I talked about for teachers and the public police,
186
535034
3214
ืืคื™ืœื• ืืœื• ืฉื”ื–ื›ืจืชื™, ืฉืžืฉืžืฉื™ื ืœื”ืขืจื›ื” ืฉืœ ืžื•ืจื™ื ื•ืœืฉื™ื˜ื•ืจ
08:58
those were built by private companies
187
538272
1869
ื ื›ืชื‘ื• ื‘ื™ื“ื™ ื—ื‘ืจื•ืช ืคืจื˜ื™ื•ืช ื•ื ืžื›ืจื• ืœืžื•ืกื“ื•ืช ืžืžืฉืœืชื™ื™ื.
09:00
and sold to the government institutions.
188
540165
2231
09:02
They call it their "secret sauce" --
189
542420
1873
ื”ื ืื•ืžืจื™ื ืฉื–ื” "ื”ืจื•ื˜ื‘ ื”ืกื•ื“ื™" ืฉืœื”ื
09:04
that's why they can't tell us about it.
190
544317
2128
ื•ืœื›ืŸ ืื™ื ื ื™ื›ื•ืœื™ื ืœื—ืฉื•ืฃ ืื•ืชื•.
09:06
It's also private power.
191
546469
2220
ื–ื”ื• ื’ื ื›ื•ื— ืคืจื˜ื™.
09:09
They are profiting for wielding the authority of the inscrutable.
192
549744
4695
ื”ื ืžืจื•ื•ื™ื—ื™ื ืžื”ืคืขืœืช ื›ื•ื— ื”ืขืžื™ืžื•ืช.
09:16
Now you might think, since all this stuff is private
193
556934
2934
ืืชื ื™ื›ื•ืœื™ื ืœื—ืฉื•ื‘, "ื‘ื’ืœืœ ืฉื›ืœ ื–ื” ืคืจื˜ื™
09:19
and there's competition,
194
559892
1158
"ื•ื™ืฉื ื” ืชื—ืจื•ืช,
"ื”ืฉื•ืง ื”ื—ื•ืคืฉื™ ืื•ืœื™ ื™ืคืชื•ืจ ืืช ื”ื‘ืขื™ื”."
09:21
maybe the free market will solve this problem.
195
561074
2306
09:23
It won't.
196
563404
1249
ืœื ื ื›ื•ืŸ.
09:24
There's a lot of money to be made in unfairness.
197
564677
3120
ืืคืฉืจ ืœื”ืจื•ื•ื™ื— ื”ืจื‘ื” ื›ืกืฃ ืžื—ื•ืกืจ ื”ื•ื’ื ื•ืช.
09:28
Also, we're not economic rational agents.
198
568947
3369
ืื ื—ื ื• ื’ื ืœื ื™ืฆื•ืจื™ื ืจืฆื™ื•ื ืœื™ื™ื ืžื‘ื—ื™ื ื” ื›ืœื›ืœื™ืช
09:32
We all are biased.
199
572851
1292
ืœื›ื•ืœื ื• ื“ืขื•ืช ืงื“ื•ืžื•ืช.
09:34
We're all racist and bigoted in ways that we wish we weren't,
200
574780
3377
ื›ื•ืœื ื• ื’ื–ืขื ื™ื ื•ืžื•ื˜ื™ื ืœืžืจื•ืช ืฉื”ื™ื™ื ื• ืžืขื“ื™ืคื™ื ืœื ืœื”ื™ื•ืช ื›ืืœื”,
09:38
in ways that we don't even know.
201
578181
2019
ื•ื‘ื“ืจื›ื™ื ืฉืื™ื ื ื• ืืคื™ืœื• ื™ื•ื“ืขื™ื.
09:41
We know this, though, in aggregate,
202
581172
3081
ืื‘ืœ ืื ื—ื ื• ื™ื•ื“ืขื™ื ืฉื‘ืžืฆื˜ื‘ืจ,
09:44
because sociologists have consistently demonstrated this
203
584277
3220
ื‘ื’ืœืœ ืฉืกื•ืฆื™ื•ืœื•ื’ื™ื ืžืจืื™ื ื‘ืื•ืคืŸ ืขืงื‘ื™
09:47
with these experiments they build,
204
587521
1665
ื‘ื ื™ืกื•ื™ื™ื ืฉื”ื ืขื•ืจื›ื™ื,
09:49
where they send a bunch of applications to jobs out,
205
589210
2568
ืฉื‘ื”ื ื”ื ืฉื•ืœื—ื™ื ืœืžืขืกื™ืงื™ื ื”ืจื‘ื” ืงื•ืจื•ืช ื—ื™ื™ื
09:51
equally qualified but some have white-sounding names
206
591802
2501
ืขื ื›ื™ืฉื•ืจื™ื ื–ื”ื™ื, ื›ืฉื—ืœืง ืžื”ืฉืžื•ืช ื ืฉืžืขื™ื "ืœื‘ื ื™ื",
ื•ืฉืžื•ืช ืื—ืจื™ื ื ืฉืžืขื™ื "ืฉื—ื•ืจื™ื",
09:54
and some have black-sounding names,
207
594327
1706
09:56
and it's always disappointing, the results -- always.
208
596057
2694
ื•ื”ืชื•ืฆืื•ืช ืฉืœ ื”ื ื™ืกื•ื™ื™ื ืชืžื™ื“ ืžืื›ื–ื‘ื•ืช, ืชืžื™ื“.
09:59
So we are the ones that are biased,
209
599330
1771
ืื– ืื ื—ื ื• ื‘ืขืœื™ ื”ื“ืขื•ืช ื”ืงื“ื•ืžื•ืช,
10:01
and we are injecting those biases into the algorithms
210
601125
3429
ื•ืื ื—ื ื• ืžื—ื“ื™ืจื™ื ืืช ื”ื”ื˜ื™ื•ืช ื”ืืœื• ืœืชื•ืš ื”ืืœื’ื•ืจื™ืชืžื™ื
10:04
by choosing what data to collect,
211
604578
1812
ื‘ื›ืš ืฉืื ื• ื‘ื•ื—ืจื™ื ืื™ืœื• ื ืชื•ื ื™ื ื™ืฉ ืœืืกื•ืฃ,
10:06
like I chose not to think about ramen noodles --
212
606414
2743
ื›ืžื• ืฉืื ื™ ื”ื—ืœื˜ืชื™ ืœื ืœื”ืชื™ื™ื—ืก ืœ"ืžื ื” ื—ืžื”"-
10:09
I decided it was irrelevant.
213
609181
1625
ื”ื—ืœื˜ืชื™ ืฉื”ื™ื ืื™ื ื ื” ืจืœื•ื•ื ื˜ื™ืช.
10:10
But by trusting the data that's actually picking up on past practices
214
610830
5684
ืื‘ืœ ืื ืื ื—ื ื• ื‘ื•ื˜ื—ื™ื ื‘ื ืชื•ื ื™ื ื•ื‘ื”ื’ื“ืจืช ื”ื”ืฆืœื—ื” ืขืœ ื™ืกื•ื“ ื’ื™ืฉื•ืช ืงื•ื“ืžื•ืช,
10:16
and by choosing the definition of success,
215
616538
2014
10:18
how can we expect the algorithms to emerge unscathed?
216
618576
3983
ืื™ืš ืื ื—ื ื• ื™ื›ื•ืœื™ื ืœืฆืคื•ืช ืฉื”ืืœื’ื•ืจื™ืชืžื™ื ื™ื™ืฆืื• ืœืœื ืคื’ืข?
10:22
We can't. We have to check them.
217
622583
2356
ืžืžืฉ ืœื. ืื ื—ื ื• ืžื•ื›ืจื—ื™ื ืœื‘ื“ื•ืง ืื•ืชื.
10:25
We have to check them for fairness.
218
625985
1709
ืื ื—ื ื• ืžื•ื›ืจื—ื™ื ืœื•ื•ื“ื ืฉื”ื ื”ื•ื’ื ื™ื.
10:27
The good news is, we can check them for fairness.
219
627718
2711
ื”ื—ื“ืฉื•ืช ื”ื˜ื•ื‘ื•ืช ื”ืŸ: ื–ื” ืืคืฉืจื™.
10:30
Algorithms can be interrogated,
220
630453
3352
ืืคืฉืจ ืœื—ืงื•ืจ ืืœื’ื•ืจื™ืชืžื™ื
10:33
and they will tell us the truth every time.
221
633829
2034
ื•ื”ื ื™ื’ื™ื“ื• ืœื ื• ืชืžื™ื“ ืืช ื”ืืžืช.
10:35
And we can fix them. We can make them better.
222
635887
2493
ื•ืื ื—ื ื• ื™ื›ื•ืœื™ื ืœืชืงืŸ ื•ืœืฉืคืจ ืื•ืชื.
10:38
I call this an algorithmic audit,
223
638404
2375
ืื ื™ ืงื•ืจืืช ืœื–ื” "ื‘ื“ื™ืงืช ืืœื’ื•ืจื™ืชื"
10:40
and I'll walk you through it.
224
640803
1679
ืืกื‘ื™ืจ ืœื›ื ืื™ืš ื–ื” ื ืขืฉื”.
10:42
First, data integrity check.
225
642506
2196
ืจืืฉื™ืช ืžื•ื•ื“ืื™ื ืืช ืฉืœืžื•ืช ื”ื ืชื•ื ื™ื.
10:45
For the recidivism risk algorithm I talked about,
226
645952
2657
ื‘ืืœื’ื•ืจื™ืชื "ื”ื™ืฉื ื•ืช ื”ืคืฉื™ืขื”" ืฉื”ื–ื›ืจืชื™,
10:49
a data integrity check would mean we'd have to come to terms with the fact
227
649402
3573
ื‘ื“ื™ืงืช ืฉืœืžื•ืช ื”ื ืชื•ื ื™ื ืคื™ืจื•ืฉื” ืฉืžื•ื›ืจื—ื™ื ืœื”ืฉืœื™ื ืขื ื”ืขื•ื‘ื“ื”
10:52
that in the US, whites and blacks smoke pot at the same rate
228
652999
3526
ืฉื‘ืืจื”"ื‘, ื”ืœื‘ื ื™ื ื•ื”ืฉื—ื•ืจื™ื ืžืขืฉื ื™ื ืžืจื™ื—ื•ืื ื” ื‘ืื•ืชื” ืžื™ื“ื”
10:56
but blacks are far more likely to be arrested --
229
656549
2485
ืื‘ืœ ืœืฉื—ื•ืจื™ื ื™ืฉ ืกื™ื›ื•ื™ ื’ื‘ื•ื” ื™ื•ืชืจ ืœื”ื™ืขืฆืจ -
10:59
four or five times more likely, depending on the area.
230
659058
3184
ืกื™ื›ื•ื™ ื’ื‘ื•ื” ืคื™ ืืจื‘ืขื” ืื• ื—ืžื™ืฉื”, ืชืœื•ื™ ื‘ืื™ื–ื•ืจ.
11:03
What is that bias looking like in other crime categories,
231
663137
2826
ืื™ืš ื ืจืื™ืช ื”ื”ื˜ื™ื” ื‘ืชื—ื•ืžื™ ืคืฉืข ืื—ืจื™ื,
11:05
and how do we account for it?
232
665987
1451
ื•ืื™ืš ืื ื—ื ื• ืžืกื‘ื™ืจื™ื ืื•ืชื”?
11:07
Second, we should think about the definition of success,
233
667982
3039
ืฉื ื™ืช, ืื ื—ื ื• ืฆืจื™ื›ื™ื ืœื”ื’ื“ื™ืจ ืžื—ื“ืฉ ืžื”ื™ ื”ืฆืœื—ื”.
11:11
audit that.
234
671045
1381
ืœื‘ื“ื•ืง ืืช ื”ื ื•ืฉื.
11:12
Remember -- with the hiring algorithm? We talked about it.
235
672450
2752
ื–ื•ื›ืจื™ื ืืช ื”ืืœื’ื•ืจื™ืชื ืœืฉื›ื™ืจืช ืขื•ื‘ื“ื™ื? ื“ื™ื‘ืจื ื• ืขืœ ื–ื”.
11:15
Someone who stays for four years and is promoted once?
236
675226
3165
ืขื•ื‘ื“ ื”ืžื•ืขืกืง ื›ื‘ืจ ืืจื‘ืข ืฉื ื™ื ื•ืงื•ื“ื ืคืขื ืื—ืช?
11:18
Well, that is a successful employee,
237
678415
1769
ื–ื” ื‘ืืžืช ืขื•ื‘ื“ ืžืฆืœื™ื—,
11:20
but it's also an employee that is supported by their culture.
238
680208
3079
ืื‘ืœ ื–ื” ื’ื ืขื•ื‘ื“ ืฉื”ืกื‘ื™ื‘ื” ื”ืชืจื‘ื•ืชื™ืช ืชื•ืžื›ืช ื‘ื•.
11:23
That said, also it can be quite biased.
239
683909
1926
ืื‘ืœ ื’ื ื›ืืŸ ื™ื›ื•ืœื•ืช ืœื”ื™ื•ืช ื“ืขื•ืช ืงื“ื•ืžื•ืช.
11:25
We need to separate those two things.
240
685859
2065
ืฆืจื™ืš ืœื”ืคืจื™ื“ ื‘ื™ืŸ ืฉื ื™ ื”ื“ื‘ืจื™ื.
11:27
We should look to the blind orchestra audition
241
687948
2426
ืœืžืฉืœ ื‘ื‘ื—ื™ื ื•ืช ืงื‘ืœื” ืขื™ื•ื•ืจื•ืช,
11:30
as an example.
242
690398
1196
11:31
That's where the people auditioning are behind a sheet.
243
691618
2756
ื›ืฉื”ื‘ื•ื—ื ื™ื ื ืžืฆืื™ื ืžืื—ื•ืจื™ ืžืกืš.
11:34
What I want to think about there
244
694766
1931
ืื ื™ ืจื•ืฆื” ืœื—ืฉื•ื‘ ืฉื›ืืŸ,
11:36
is the people who are listening have decided what's important
245
696721
3417
ื”ืื ืฉื™ื ื”ืžืงืฉื™ื‘ื™ื ื”ื ืฉื”ื—ืœื™ื˜ื• ืžื” ื—ืฉื•ื‘ ื•ืžื” ืœื,
11:40
and they've decided what's not important,
246
700162
2029
ื•ื“ืขืชื ืœื ืžื•ืกื—ืช ืข"ื™ ื–ื”.
11:42
and they're not getting distracted by that.
247
702215
2059
11:44
When the blind orchestra auditions started,
248
704781
2749
ื›ืฉื”ืชื—ื™ืœื• ื”ืžื‘ื—ื ื™ื ื”ืขื™ื•ื•ืจื™ื,
11:47
the number of women in orchestras went up by a factor of five.
249
707554
3444
ืžืกืคืจ ื”ื ืฉื™ื ื”ืžื ื’ื ื•ืช ื‘ืชื–ืžื•ืจืช ื’ื“ืœ ืคื™ ื—ืžืฉ.
11:52
Next, we have to consider accuracy.
250
712073
2015
ื”ื‘ื ื‘ืชื•ืจ ื”ื•ื ื”ื“ื™ื•ืง.
11:55
This is where the value-added model for teachers would fail immediately.
251
715053
3734
ื›ืืŸ ืืœื’ื•ืจื™ืชื ื”ืขืจืš ื”ืžื•ืกืฃ ืœื“ื™ืจื•ื’ ืžื•ืจื™ื ื™ื™ื›ืฉืœ ืžื™ื“.
11:59
No algorithm is perfect, of course,
252
719398
2162
ืื™ืŸ ืืœื’ื•ืจื™ืชื ืžื•ืฉืœื, ื›ืžื•ื‘ืŸ,
12:02
so we have to consider the errors of every algorithm.
253
722440
3605
ืื– ืฆืจื™ืš ืœืงื—ืช ื‘ื—ืฉื‘ื•ืŸ ืืช ื”ืฉื’ื™ืื•ืช ืฉืœ ื›ืœ ืืœื’ื•ืจื™ืชื:
12:06
How often are there errors, and for whom does this model fail?
254
726656
4359
ื›ืžื” ื•ืžืชื™ ื”ืŸ ืงื•ืจื•ืช ื•ืขื ืžื™ ื”ืžื•ื“ืœ ื”ื–ื” ื ื›ืฉืœ?
12:11
What is the cost of that failure?
255
731670
1718
ืžื”ื• ื”ืžื—ื™ืจ ืฉืœ ื”ื›ืฉืœื•ืŸ ื”ื–ื”?
12:14
And finally, we have to consider
256
734254
2207
ื•ืœืกื™ื•ื, ืื ื—ื ื• ืžื•ื›ืจื—ื™ื ืœืงื—ืช ื‘ื—ืฉื‘ื•ืŸ
12:17
the long-term effects of algorithms,
257
737793
2186
ืืช ื”ื”ืฉืคืขื•ืช ืืจื•ื›ื•ืช ื”ื˜ื•ื•ื— ืฉืœ ื”ืืœื’ื•ืจื™ืชืžื™ื,
12:20
the feedback loops that are engendering.
258
740686
2207
ืฉืœ ืœื•ืœืื•ืช ื”ืžืฉื•ื‘ ืฉื ื•ืฆืจื•ืช.
12:23
That sounds abstract,
259
743406
1236
ื–ื” ื ืฉืžืข ืžื•ืคืฉื˜,
12:24
but imagine if Facebook engineers had considered that
260
744666
2664
ืื‘ืœ ืžื” ืื ืžื”ื ื“ืกื™ "ืคื™ื™ืกื‘ื•ืง" ื”ื™ื• ืœื•ืงื—ื™ื ื–ืืช ื‘ื—ืฉื‘ื•ืŸ
12:28
before they decided to show us only things that our friends had posted.
261
748090
4855
ื‘ื˜ืจื ื”ื—ืœื™ื˜ื• ืœื”ืจืื•ืช ืœื ื• ืจืง ืžื” ืฉืฉื™ืชืคื• ื”ื—ื‘ืจื™ื ืฉืœื ื•.
12:33
I have two more messages, one for the data scientists out there.
262
753581
3234
ื™ืฉ ืœื™ ืขื•ื“ ืฉื ื™ ืžืกืจื™ื, ืื—ื“ ืœืžืชื›ื ืชื™ื ื‘ืืฉืจ ื”ื:
12:37
Data scientists: we should not be the arbiters of truth.
263
757270
3409
ืžืชื›ื ืชื™ื: ืืกื•ืจ ืœื ื• ืœืชื•ื•ืš ืืช ื”ืืžืช.
12:41
We should be translators of ethical discussions that happen
264
761340
3783
ืื ื—ื ื• ืฆืจื™ื›ื™ื ืœืชืช ื‘ื™ื˜ื•ื™ ืœื“ื™ื•ื ื™ ืžื•ืกืจ ืฉืžืชืงื™ื™ืžื™ื
12:45
in larger society.
265
765147
1294
ื‘ื—ื‘ืจื” ื›ื•ืœื”.
12:47
(Applause)
266
767399
2133
(ืžื—ื™ืื•ืช ื›ืคื™ื™ื)
12:49
And the rest of you,
267
769556
1556
ื•ืœืฉืืจ ื”ืื ืฉื™ื,
12:51
the non-data scientists:
268
771831
1396
ืืœื• ืฉืื™ื ื ืขื•ืกืงื™ื ื‘ืžื™ื“ืข:
12:53
this is not a math test.
269
773251
1498
ืœื ืžื“ื•ื‘ืจ ื‘ืžื‘ื—ืŸ ื‘ืžืชืžื˜ื™ืงื”,
12:55
This is a political fight.
270
775452
1348
ืืœื ื‘ืžืื‘ืง ืคื•ืœื™ื˜ื™.
12:58
We need to demand accountability for our algorithmic overlords.
271
778407
3907
ืื ื—ื ื• ืฆืจื™ื›ื™ื ืœื“ืจื•ืฉ ืžืฉืœื™ื˜ื™ ื”ืืœื’ื•ืจื™ืชืžื™ื ืœืงื—ืช ืื—ืจื™ื•ืช.
13:03
(Applause)
272
783938
1499
(ืžื—ื™ืื•ืช ื›ืคื™ื™ื)
13:05
The era of blind faith in big data must end.
273
785461
4225
ืขื™ื“ืŸ ื”ืืžื•ืŸ ื”ืขื™ื•ื•ืจ ื‘ื ืชื•ื ื™ื ื—ื™ื™ื‘ ืœื”ืกืชื™ื™ื.
13:09
Thank you very much.
274
789710
1167
ืชื•ื“ื” ืจื‘ื”.
13:10
(Applause)
275
790901
5303
(ืžื—ื™ืื•ืช ื›ืคื™ื™ื)
ืขืœ ืืชืจ ื–ื”

ืืชืจ ื–ื” ื™ืฆื™ื’ ื‘ืคื ื™ื›ื ืกืจื˜ื•ื ื™ YouTube ื”ืžื•ืขื™ืœื™ื ืœืœื™ืžื•ื“ ืื ื’ืœื™ืช. ืชื•ื›ืœื• ืœืจืื•ืช ืฉื™ืขื•ืจื™ ืื ื’ืœื™ืช ื”ืžื•ืขื‘ืจื™ื ืขืœ ื™ื“ื™ ืžื•ืจื™ื ืžื”ืฉื•ืจื” ื”ืจืืฉื•ื ื” ืžืจื—ื‘ื™ ื”ืขื•ืœื. ืœื—ืฅ ืคืขืžื™ื™ื ืขืœ ื”ื›ืชื•ื‘ื™ื•ืช ื‘ืื ื’ืœื™ืช ื”ืžื•ืฆื’ื•ืช ื‘ื›ืœ ื“ืฃ ื•ื™ื“ืื• ื›ื“ื™ ืœื”ืคืขื™ืœ ืืช ื”ืกืจื˜ื•ืŸ ืžืฉื. ื”ื›ืชื•ื‘ื™ื•ืช ื’ื•ืœืœื•ืช ื‘ืกื ื›ืจื•ืŸ ืขื ื”ืคืขืœืช ื”ื•ื•ื™ื“ืื•. ืื ื™ืฉ ืœืš ื”ืขืจื•ืช ืื• ื‘ืงืฉื•ืช, ืื ื ืฆื•ืจ ืื™ืชื ื• ืงืฉืจ ื‘ืืžืฆืขื•ืช ื˜ื•ืคืก ื™ืฆื™ืจืช ืงืฉืจ ื–ื”.

https://forms.gle/WvT1wiN1qDtmnspy7