1
00:00:00,738 --> 00:00:02,735
If you remember that first decade of the web,

2
00:00:02,735 --> 00:00:04,990
it was really a static place.

3
00:00:04,990 --> 00:00:07,235
You could go online, you could look at pages,

4
00:00:07,235 --> 00:00:09,748
and they were put up either by organizations

5
00:00:09,748 --> 00:00:11,269
who had teams to do it

6
00:00:11,269 --> 00:00:13,498
or by individuals who were really tech-savvy

7
00:00:13,498 --> 00:00:15,235
for the time.

8
00:00:15,235 --> 00:00:16,810
And with the rise of social media

9
00:00:16,810 --> 00:00:19,209
and social networks in the early 2000s,

10
00:00:19,209 --> 00:00:21,358
the web was completely changed

11
00:00:21,358 --> 00:00:24,966
to a place where now the vast majority of content

12
00:00:24,966 --> 00:00:28,278
we interact with is put up by average users,

13
00:00:28,278 --> 00:00:30,975
either in YouTube videos or blog posts

14
00:00:30,975 --> 00:00:34,290
or product reviews or social media postings.

15
00:00:34,290 --> 00:00:36,637
And it's also become a much more interactive place,

16
00:00:36,637 --> 00:00:39,274
where people are interacting with others,

17
00:00:39,274 --> 00:00:40,970
they're commenting, they're sharing,

18
00:00:40,970 --> 00:00:42,584
they're not just reading.

19
00:00:42,584 --> 00:00:44,450
So Facebook is not the only place you can do this,

20
00:00:44,450 --> 00:00:45,548
but it's the biggest,

21
00:00:45,548 --> 00:00:47,332
and it serves to illustrate the numbers.

22
00:00:47,332 --> 00:00:50,809
Facebook has 1.2 billion users per month.

23
00:00:50,809 --> 00:00:52,739
So half the Earth's Internet population

24
00:00:52,739 --> 00:00:54,392
is using Facebook.

25
00:00:54,392 --> 00:00:56,324
They are a site, along with others,

26
00:00:56,324 --> 00:00:59,543
that has allowed people to create an online persona

27
00:00:59,543 --> 00:01:01,325
with very little technical skill,

28
00:01:01,325 --> 00:01:03,801
and people responded by putting huge amounts

29
00:01:03,801 --> 00:01:05,784
of personal data online.

30
00:01:05,784 --> 00:01:08,327
So the result is that we have behavioral,

31
00:01:08,327 --> 00:01:10,313
preference, demographic data

32
00:01:10,313 --> 00:01:12,414
for hundreds of millions of people,

33
00:01:12,414 --> 00:01:14,440
which is unprecedented in history.

34
00:01:14,440 --> 00:01:17,000
And as a computer scientist, 
what this means is that

35
00:01:17,000 --> 00:01:18,664
I've been able to build models

36
00:01:18,664 --> 00:01:20,986
that can predict all sorts of hidden attributes

37
00:01:20,986 --> 00:01:23,270
for all of you that you don't even know

38
00:01:23,270 --> 00:01:25,472
you're sharing information about.

39
00:01:25,472 --> 00:01:27,854
As scientists, we use that to help

40
00:01:27,854 --> 00:01:29,968
the way people interact online,

41
00:01:29,968 --> 00:01:32,467
but there's less altruistic applications,

42
00:01:32,467 --> 00:01:34,848
and there's a problem in that users don't really

43
00:01:34,848 --> 00:01:37,318
understand these techniques and how they work,

44
00:01:37,318 --> 00:01:40,446
and even if they did, they don't
have a lot of control over it.

45
00:01:40,446 --> 00:01:41,936
So what I want to talk to you about today

46
00:01:41,936 --> 00:01:44,638
is some of these things that we're able to do,

47
00:01:44,638 --> 00:01:47,401
and then give us some ideas
of how we might go forward

48
00:01:47,401 --> 00:01:50,170
to move some control back into the hands of users.

49
00:01:50,170 --> 00:01:51,756
So this is Target, the company.

50
00:01:51,756 --> 00:01:53,080
I didn't just put that logo

51
00:01:53,080 --> 00:01:55,250
on this poor, pregnant woman's belly.

52
00:01:55,250 --> 00:01:57,090
You may have seen this anecdote that was printed

53
00:01:57,090 --> 00:01:59,151
in Forbes magazine where Target

54
00:01:59,151 --> 00:02:01,512
sent a flyer to this 15-year-old girl

55
00:02:01,512 --> 00:02:03,222
with advertisements and coupons

56
00:02:03,222 --> 00:02:05,776
for baby bottles and diapers and cribs

57
00:02:05,776 --> 00:02:07,460
two weeks before she told her parents

58
00:02:07,460 --> 00:02:09,324
that she was pregnant.

59
00:02:09,324 --> 00:02:12,028
Yeah, the dad was really upset.

60
00:02:12,028 --> 00:02:13,744
He said, "How did Target figure out

61
00:02:13,744 --> 00:02:15,568
that this high school girl was pregnant

62
00:02:15,568 --> 00:02:17,528
before she told her parents?"

63
00:02:17,528 --> 00:02:20,149
It turns out that they have the purchase history

64
00:02:20,149 --> 00:02:22,450
for hundreds of thousands of customers

65
00:02:22,450 --> 00:02:25,180
and they compute what they 
call a pregnancy score,

66
00:02:25,180 --> 00:02:27,512
which is not just whether or 
not a woman's pregnant,

67
00:02:27,512 --> 00:02:29,242
but what her due date is.

68
00:02:29,242 --> 00:02:30,546
And they compute that

69
00:02:30,546 --> 00:02:32,314
not by looking at the obvious things,

70
00:02:32,314 --> 00:02:34,826
like, she's buying a crib or baby clothes,

71
00:02:34,826 --> 00:02:37,769
but things like, she bought more vitamins

72
00:02:37,769 --> 00:02:39,486
than she normally had,

73
00:02:39,486 --> 00:02:40,950
or she bought a handbag

74
00:02:40,950 --> 00:02:42,661
that's big enough to hold diapers.

75
00:02:42,661 --> 00:02:44,571
And by themselves, those purchases don't seem

76
00:02:44,571 --> 00:02:47,040
like they might reveal a lot,

77
00:02:47,040 --> 00:02:49,018
but it's a pattern of behavior that,

78
00:02:49,018 --> 00:02:52,135
when you take it in the context 
of thousands of other people,

79
00:02:52,135 --> 00:02:54,892
starts to actually reveal some insights.

80
00:02:54,892 --> 00:02:56,685
So that's the kind of thing that we do

81
00:02:56,685 --> 00:02:59,252
when we're predicting stuff
about you on social media.

82
00:02:59,252 --> 00:03:02,048
We're looking for little
patterns of behavior that,

83
00:03:02,048 --> 00:03:04,730
when you detect them among millions of people,

84
00:03:04,730 --> 00:03:07,436
lets us find out all kinds of things.

85
00:03:07,436 --> 00:03:09,183
So in my lab and with colleagues,

86
00:03:09,183 --> 00:03:10,960
we've developed mechanisms where we can

87
00:03:10,960 --> 00:03:12,520
quite accurately predict things

88
00:03:12,520 --> 00:03:14,245
like your political preference,

89
00:03:14,245 --> 00:03:17,997
your personality score, gender, sexual orientation,

90
00:03:17,997 --> 00:03:20,870
religion, age, intelligence,

91
00:03:20,870 --> 00:03:22,264
along with things like

92
00:03:22,264 --> 00:03:24,201
how much you trust the people you know

93
00:03:24,201 --> 00:03:26,005
and how strong those relationships are.

94
00:03:26,005 --> 00:03:27,790
We can do all of this really well.

95
00:03:27,790 --> 00:03:29,987
And again, it doesn't come from what you might

96
00:03:29,987 --> 00:03:32,089
think of as obvious information.

97
00:03:32,089 --> 00:03:34,370
So my favorite example is from this study

98
00:03:34,370 --> 00:03:35,610
that was published this year

99
00:03:35,610 --> 00:03:37,405
in the Proceedings of the National Academies.

100
00:03:37,405 --> 00:03:38,690
If you Google this, you'll find it.

101
00:03:38,690 --> 00:03:40,562
It's four pages, easy to read.

102
00:03:40,562 --> 00:03:43,565
And they looked at just people's Facebook likes,

103
00:03:43,565 --> 00:03:45,485
so just the things you like on Facebook,

104
00:03:45,485 --> 00:03:47,623
and used that to predict all these attributes,

105
00:03:47,623 --> 00:03:49,268
along with some other ones.

106
00:03:49,268 --> 00:03:52,229
And in their paper they listed the five likes

107
00:03:52,229 --> 00:03:55,016
that were most indicative of high intelligence.

108
00:03:55,016 --> 00:03:57,340
And among those was liking a page

109
00:03:57,340 --> 00:03:59,245
for curly fries. (Laughter)

110
00:03:59,245 --> 00:04:01,338
Curly fries are delicious,

111
00:04:01,338 --> 00:04:03,868
but liking them does not necessarily mean

112
00:04:03,868 --> 00:04:05,948
that you're smarter than the average person.

113
00:04:05,948 --> 00:04:09,155
So how is it that one of the strongest indicators

114
00:04:09,155 --> 00:04:10,725
of your intelligence

115
00:04:10,725 --> 00:04:12,172
is liking this page

116
00:04:12,172 --> 00:04:14,424
when the content is totally irrelevant

117
00:04:14,424 --> 00:04:16,951
to the attribute that's being predicted?

118
00:04:16,951 --> 00:04:18,535
And it turns out that we have to look at

119
00:04:18,535 --> 00:04:20,153
a whole bunch of underlying theories

120
00:04:20,153 --> 00:04:22,722
to see why we're able to do this.

121
00:04:22,722 --> 00:04:25,635
One of them is a sociological
theory called homophily,

122
00:04:25,635 --> 00:04:28,727
which basically says people are
friends with people like them.

123
00:04:28,727 --> 00:04:30,741
So if you're smart, you tend to
be friends with smart people,

124
00:04:30,741 --> 00:04:33,371
and if you're young, you tend
to be friends with young people,

125
00:04:33,371 --> 00:04:34,998
and this is well established

126
00:04:34,998 --> 00:04:36,743
for hundreds of years.

127
00:04:36,743 --> 00:04:37,975
We also know a lot

128
00:04:37,975 --> 00:04:40,525
about how information spreads through networks.

129
00:04:40,525 --> 00:04:42,279
It turns out things like viral videos

130
00:04:42,279 --> 00:04:44,685
or Facebook likes or other information

131
00:04:44,685 --> 00:04:46,573
spreads in exactly the same way

132
00:04:46,573 --> 00:04:49,027
that diseases spread through social networks.

133
00:04:49,027 --> 00:04:50,818
So this is something we've studied for a long time.

134
00:04:50,818 --> 00:04:52,394
We have good models of it.

135
00:04:52,394 --> 00:04:54,551
And so you can put those things together

136
00:04:54,551 --> 00:04:57,639
and start seeing why things like this happen.

137
00:04:57,639 --> 00:04:59,453
So if I were to give you a hypothesis,

138
00:04:59,453 --> 00:05:02,680
it would be that a smart guy started this page,

139
00:05:02,680 --> 00:05:04,619
or maybe one of the first people who liked it

140
00:05:04,619 --> 00:05:06,355
would have scored high on that test.

141
00:05:06,355 --> 00:05:08,643
And they liked it, and their friends saw it,

142
00:05:08,643 --> 00:05:11,765
and by homophily, we know that
he probably had smart friends,

143
00:05:11,765 --> 00:05:14,821
and so it spread to them, 
and some of them liked it,

144
00:05:14,821 --> 00:05:16,010
and they had smart friends,

145
00:05:16,010 --> 00:05:16,817
and so it spread to them,

146
00:05:16,817 --> 00:05:18,790
and so it propagated through the network

147
00:05:18,790 --> 00:05:21,359
to a host of smart people,

148
00:05:21,359 --> 00:05:23,415
so that by the end, the action

149
00:05:23,415 --> 00:05:25,959
of liking the curly fries page

150
00:05:25,959 --> 00:05:27,574
is indicative of high intelligence,

151
00:05:27,574 --> 00:05:29,377
not because of the content,

152
00:05:29,377 --> 00:05:31,899
but because the actual action of liking

153
00:05:31,899 --> 00:05:33,799
reflects back the common attributes

154
00:05:33,799 --> 00:05:36,267
of other people who have done it.

155
00:05:36,267 --> 00:05:39,164
So this is pretty complicated stuff, right?

156
00:05:39,164 --> 00:05:41,363
It's a hard thing to sit down and explain

157
00:05:41,363 --> 00:05:44,211
to an average user, and even if you do,

158
00:05:44,211 --> 00:05:46,399
what can the average user do about it?

159
00:05:46,399 --> 00:05:48,447
How do you know that 
you've liked something

160
00:05:48,447 --> 00:05:49,939
that indicates a trait for you

161
00:05:49,939 --> 00:05:53,484
that's totally irrelevant to the
content of what you've liked?

162
00:05:53,484 --> 00:05:56,030
There's a lot of power that users don't have

163
00:05:56,030 --> 00:05:58,260
to control how this data is used.

164
00:05:58,260 --> 00:06:01,372
And I see that as a real 
problem going forward.

165
00:06:01,372 --> 00:06:03,349
So I think there's a couple paths

166
00:06:03,349 --> 00:06:04,350
that we want to look at

167
00:06:04,350 --> 00:06:06,260
if we want to give users some control

168
00:06:06,260 --> 00:06:08,000
over how this data is used,

169
00:06:08,000 --> 00:06:09,940
because it's not always going to be used

170
00:06:09,940 --> 00:06:11,321
for their benefit.

171
00:06:11,321 --> 00:06:12,743
An example I often give is that,

172
00:06:12,743 --> 00:06:14,389
if I ever get bored being a professor,

173
00:06:14,389 --> 00:06:16,042
I'm going to go start a company

174
00:06:16,042 --> 00:06:17,496
that predicts all of these attributes

175
00:06:17,496 --> 00:06:19,098
and things like how well you work in teams

176
00:06:19,098 --> 00:06:21,769
and if you're a drug user, if you're an alcoholic.

177
00:06:21,769 --> 00:06:23,209
We know how to predict all that.

178
00:06:23,209 --> 00:06:24,970
And I'm going to sell reports

179
00:06:24,970 --> 00:06:27,070
to H.R. companies and big businesses

180
00:06:27,070 --> 00:06:29,343
that want to hire you.

181
00:06:29,343 --> 00:06:30,520
We totally can do that now.

182
00:06:30,520 --> 00:06:32,308
I could start that business tomorrow,

183
00:06:32,308 --> 00:06:34,360
and you would have
absolutely no control

184
00:06:34,360 --> 00:06:36,498
over me using your data like that.

185
00:06:36,498 --> 00:06:38,790
That seems to me to be a problem.

186
00:06:38,790 --> 00:06:40,700
So one of the paths we can go down

187
00:06:40,700 --> 00:06:42,732
is the policy and law path.

188
00:06:42,732 --> 00:06:45,778
And in some respects, I think
that that would be most effective,

189
00:06:45,778 --> 00:06:48,534
but the problem is we'd
actually have to do it.

190
00:06:48,534 --> 00:06:51,314
Observing our political process in action

191
00:06:51,314 --> 00:06:53,693
makes me think it's highly unlikely

192
00:06:53,693 --> 00:06:55,290
that we're going to get a bunch of representatives

193
00:06:55,290 --> 00:06:57,276
to sit down, learn about this,

194
00:06:57,276 --> 00:06:59,382
and then enact sweeping changes

195
00:06:59,382 --> 00:07:01,539
to intellectual property law in the U.S.

196
00:07:01,539 --> 00:07:04,000
so users control their data.

197
00:07:04,000 --> 00:07:05,304
We could go the policy route,

198
00:07:05,304 --> 00:07:06,783
where social media companies say,

199
00:07:06,783 --> 00:07:08,185
you know what? You own your data.

200
00:07:08,185 --> 00:07:10,674
You have total control over how it's used.

201
00:07:10,674 --> 00:07:12,522
The problem is that the revenue models

202
00:07:12,522 --> 00:07:14,246
for most social media companies

203
00:07:14,246 --> 00:07:18,277
rely on sharing or exploiting 
users' data in some way.

204
00:07:18,277 --> 00:07:20,110
It's sometimes said of Facebook that the users

205
00:07:20,110 --> 00:07:22,638
aren't the customer, they're the product.

206
00:07:22,638 --> 00:07:25,352
And so how do you get a company

207
00:07:25,352 --> 00:07:27,910
to cede control of their main asset

208
00:07:27,910 --> 00:07:29,159
back to the users?

209
00:07:29,159 --> 00:07:30,860
It's possible, but I don't think it's something

210
00:07:30,860 --> 00:07:33,180
that we're going to see change quickly.

211
00:07:33,180 --> 00:07:34,680
So I think the other path

212
00:07:34,680 --> 00:07:36,968
that we can go down that's
going to be more effective

213
00:07:36,968 --> 00:07:38,476
is one of more science.

214
00:07:38,476 --> 00:07:40,986
It's doing science that allowed us to develop

215
00:07:40,986 --> 00:07:42,736
all these mechanisms for computing

216
00:07:42,736 --> 00:07:44,788
this personal data in the first place.

217
00:07:44,788 --> 00:07:46,894
And it's actually very similar research

218
00:07:46,894 --> 00:07:48,332
that we'd have to do

219
00:07:48,332 --> 00:07:50,718
if we want to develop mechanisms

220
00:07:50,718 --> 00:07:52,139
that can say to a user,

221
00:07:52,139 --> 00:07:54,368
"Here's the risk of that action you just took."

222
00:07:54,368 --> 00:07:56,448
By liking that Facebook page,

223
00:07:56,448 --> 00:07:58,983
or by sharing this piece of personal information,

224
00:07:58,983 --> 00:08:00,485
you've now improved my ability

225
00:08:00,485 --> 00:08:02,571
to predict whether or not you're using drugs

226
00:08:02,571 --> 00:08:05,433
or whether or not you get
along well in the workplace.

227
00:08:05,433 --> 00:08:07,281
And that, I think, can affect whether or not

228
00:08:07,281 --> 00:08:08,791
people want to share something,

229
00:08:08,791 --> 00:08:12,030
keep it private, or just keep it offline altogether.

230
00:08:12,030 --> 00:08:13,593
We can also look at things like

231
00:08:13,593 --> 00:08:16,321
allowing people to encrypt data that they upload,

232
00:08:16,321 --> 00:08:18,176
so it's kind of invisible and worthless

233
00:08:18,176 --> 00:08:19,607
to sites like Facebook

234
00:08:19,607 --> 00:08:22,236
or third party services that access it,

235
00:08:22,236 --> 00:08:25,483
but that select users who the person who posted it

236
00:08:25,483 --> 00:08:28,153
want to see it have access to see it.

237
00:08:28,153 --> 00:08:30,319
This is all super exciting research

238
00:08:30,319 --> 00:08:31,939
from an intellectual perspective,

239
00:08:31,939 --> 00:08:33,798
and so scientists are going to be willing to do it.

240
00:08:33,798 --> 00:08:37,408
So that gives us an advantage over the law side.

241
00:08:37,408 --> 00:08:39,133
One of the problems that people bring up

242
00:08:39,133 --> 00:08:40,728
when I talk about this is, they say,

243
00:08:40,728 --> 00:08:43,374
you know, if people start
keeping all this data private,

244
00:08:43,374 --> 00:08:45,487
all those methods that you've been developing

245
00:08:45,487 --> 00:08:48,140
to predict their traits are going to fail.

246
00:08:48,140 --> 00:08:51,660
And I say, absolutely, and for me, that's success,

247
00:08:51,660 --> 00:08:53,446
because as a scientist,

248
00:08:53,446 --> 00:08:57,134
my goal is not to infer information about users,

249
00:08:57,134 --> 00:08:59,901
it's to improve the way people interact online.

250
00:08:59,901 --> 00:09:03,119
And sometimes that involves
inferring things about them,

251
00:09:03,119 --> 00:09:06,141
but if users don't want me to use that data,

252
00:09:06,141 --> 00:09:08,179
I think they should have the right to do that.

253
00:09:08,179 --> 00:09:10,830
I want users to be informed and consenting

254
00:09:10,830 --> 00:09:12,942
users of the tools that we develop.

255
00:09:12,942 --> 00:09:15,894
And so I think encouraging this kind of science

256
00:09:15,894 --> 00:09:17,240
and supporting researchers

257
00:09:17,240 --> 00:09:20,263
who want to cede some of that control back to users

258
00:09:20,263 --> 00:09:22,574
and away from the social media companies

259
00:09:22,574 --> 00:09:25,245
means that going forward, as these tools evolve

260
00:09:25,245 --> 00:09:26,721
and advance,

261
00:09:26,721 --> 00:09:28,135
means that we're going to have an educated

262
00:09:28,135 --> 00:09:29,829
and empowered user base,

263
00:09:29,829 --> 00:09:30,929
and I think all of us can agree

264
00:09:30,929 --> 00:09:33,493
that that's a pretty ideal way to go forward.

265
00:09:33,493 --> 00:09:35,677
Thank you.

266
00:09:35,677 --> 00:09:38,757
(Applause)