1
00:00:01,548 --> 00:00:04,064
No matter who you are or where you live,

2
00:00:04,088 --> 00:00:06,444
I'm guessing that you have
at least one relative

3
00:00:06,468 --> 00:00:08,802
that likes to forward those emails.

4
00:00:09,206 --> 00:00:11,106
You know the ones I'm talking about --

5
00:00:11,130 --> 00:00:13,984
the ones with dubious claims
or conspiracy videos.

6
00:00:14,315 --> 00:00:16,983
And you've probably
already muted them on Facebook

7
00:00:17,007 --> 00:00:19,355
for sharing social posts like this one.

8
00:00:19,379 --> 00:00:20,783
It's an image of a banana

9
00:00:20,807 --> 00:00:23,474
with a strange red cross
running through the center.

10
00:00:23,498 --> 00:00:25,637
And the text around it is warning people

11
00:00:25,661 --> 00:00:27,819
not to eat fruits that look like this,

12
00:00:27,843 --> 00:00:29,871
suggesting they've been
injected with blood

13
00:00:29,895 --> 00:00:32,025
contaminated with the HIV virus.

14
00:00:32,049 --> 00:00:34,652
And the social share message
above it simply says,

15
00:00:34,676 --> 00:00:36,857
"Please forward to save lives."

16
00:00:37,672 --> 00:00:40,966
Now, fact-checkers have been debunking
this one for years,

17
00:00:40,990 --> 00:00:43,799
but it's one of those rumors
that just won't die.

18
00:00:43,823 --> 00:00:45,094
A zombie rumor.

19
00:00:45,513 --> 00:00:47,606
And, of course, it's entirely false.

20
00:00:48,180 --> 00:00:51,139
It might be tempting to laugh
at an example like this, to say,

21
00:00:51,163 --> 00:00:53,047
"Well, who would believe this, anyway?"

22
00:00:53,419 --> 00:00:55,045
But the reason it's a zombie rumor

23
00:00:55,069 --> 00:00:58,958
is because it taps into people's
deepest fears about their own safety

24
00:00:58,982 --> 00:01:01,157
and that of the people they love.

25
00:01:01,783 --> 00:01:05,056
And if you spend as enough time
as I have looking at misinformation,

26
00:01:05,080 --> 00:01:07,500
you know that this is just
one example of many

27
00:01:07,524 --> 00:01:10,571
that taps into people's deepest
fears and vulnerabilities.

28
00:01:11,214 --> 00:01:15,583
Every day, across the world,
we see scores of new memes on Instagram

29
00:01:15,607 --> 00:01:18,646
encouraging parents
not to vaccinate their children.

30
00:01:18,670 --> 00:01:23,202
We see new videos on YouTube
explaining that climate change is a hoax.

31
00:01:23,226 --> 00:01:27,528
And across all platforms, we see
endless posts designed to demonize others

32
00:01:27,552 --> 00:01:31,053
on the basis of their race,
religion or sexuality.

33
00:01:32,314 --> 00:01:35,344
Welcome to one of the central
challenges of our time.

34
00:01:35,647 --> 00:01:39,672
How can we maintain an internet
with freedom of expression at the core,

35
00:01:39,696 --> 00:01:42,999
while also ensuring that the content
that's being disseminated

36
00:01:43,023 --> 00:01:46,909
doesn't cause irreparable harms
to our democracies, our communities

37
00:01:46,933 --> 00:01:49,171
and to our physical and mental well-being?

38
00:01:49,998 --> 00:01:52,085
Because we live in the information age,

39
00:01:52,109 --> 00:01:55,656
yet the central currency
upon which we all depend -- information --

40
00:01:55,680 --> 00:01:58,037
is no longer deemed entirely trustworthy

41
00:01:58,061 --> 00:02:00,389
and, at times, can appear
downright dangerous.

42
00:02:00,811 --> 00:02:04,748
This is thanks in part to the runaway
growth of social sharing platforms

43
00:02:04,772 --> 00:02:06,414
that allow us to scroll through,

44
00:02:06,438 --> 00:02:08,660
where lies and facts sit side by side,

45
00:02:08,684 --> 00:02:11,755
but with none of the traditional
signals of trustworthiness.

46
00:02:12,268 --> 00:02:15,887
And goodness -- our language around this
is horribly muddled.

47
00:02:15,911 --> 00:02:19,014
People are still obsessed
with the phrase "fake news,"

48
00:02:19,038 --> 00:02:21,569
despite the fact that
it's extraordinarily unhelpful

49
00:02:21,593 --> 00:02:25,053
and used to describe a number of things
that are actually very different:

50
00:02:25,077 --> 00:02:28,463
lies, rumors, hoaxes,
conspiracies, propaganda.

51
00:02:28,911 --> 00:02:31,823
And I really wish
we could stop using a phrase

52
00:02:31,847 --> 00:02:34,709
that's been co-opted by politicians
right around the world,

53
00:02:34,733 --> 00:02:36,204
from the left and the right,

54
00:02:36,228 --> 00:02:39,450
used as a weapon to attack
a free and independent press.

55
00:02:40,307 --> 00:02:45,009
(Applause)

56
00:02:45,033 --> 00:02:48,495
Because we need our professional
news media now more than ever.

57
00:02:48,882 --> 00:02:52,255
And besides, most of this content
doesn't even masquerade as news.

58
00:02:52,279 --> 00:02:54,921
It's memes, videos, social posts.

59
00:02:54,945 --> 00:02:58,398
And most of it is not fake;
it's misleading.

60
00:02:58,422 --> 00:03:01,437
We tend to fixate on what's true or false.

61
00:03:01,461 --> 00:03:05,493
But the biggest concern is actually
the weaponization of context.

62
00:03:06,855 --> 00:03:08,823
Because the most effective disinformation

63
00:03:08,847 --> 00:03:11,895
has always been that
which has a kernel of truth to it.

64
00:03:11,919 --> 00:03:15,395
Let's take this example
from London, from March 2017,

65
00:03:15,419 --> 00:03:16,959
a tweet that circulated widely

66
00:03:16,983 --> 00:03:20,570
in the aftermath of a terrorist incident
on Westminster Bridge.

67
00:03:20,594 --> 00:03:23,022
This is a genuine image, not fake.

68
00:03:23,046 --> 00:03:26,215
The woman who appears in the photograph
was interviewed afterwards,

69
00:03:26,239 --> 00:03:28,648
and she explained that
she was utterly traumatized.

70
00:03:28,672 --> 00:03:30,410
She was on the phone to a loved one,

71
00:03:30,434 --> 00:03:33,052
and she wasn't looking
at the victim out of respect.

72
00:03:33,076 --> 00:03:37,036
But it still was circulated widely
with this Islamophobic framing,

73
00:03:37,060 --> 00:03:40,106
with multiple hashtags,
including: #BanIslam.

74
00:03:40,425 --> 00:03:42,823
Now, if you worked at Twitter,
what would you do?

75
00:03:42,847 --> 00:03:45,409
Would you take that down,
or would you leave it up?

76
00:03:46,553 --> 00:03:49,982
My gut reaction, my emotional reaction,
is to take this down.

77
00:03:50,006 --> 00:03:52,148
I hate the framing of this image.

78
00:03:52,585 --> 00:03:54,973
But freedom of expression
is a human right,

79
00:03:54,997 --> 00:03:58,222
and if we start taking down speech
that makes us feel uncomfortable,

80
00:03:58,246 --> 00:03:59,476
we're in trouble.

81
00:03:59,500 --> 00:04:01,794
And this might look like a clear-cut case,

82
00:04:01,818 --> 00:04:03,516
but, actually, most speech isn't.

83
00:04:03,540 --> 00:04:05,976
These lines are incredibly
difficult to draw.

84
00:04:06,000 --> 00:04:08,281
What's a well-meaning
decision by one person

85
00:04:08,305 --> 00:04:10,382
is outright censorship to the next.

86
00:04:10,759 --> 00:04:13,688
What we now know is that
this account, Texas Lone Star,

87
00:04:13,712 --> 00:04:16,942
was part of a wider Russian
disinformation campaign,

88
00:04:16,966 --> 00:04:19,117
one that has since been taken down.

89
00:04:19,141 --> 00:04:20,704
Would that change your view?

90
00:04:21,322 --> 00:04:22,481
It would mine,

91
00:04:22,505 --> 00:04:24,806
because now it's a case
of a coordinated campaign

92
00:04:24,830 --> 00:04:26,045
to sow discord.

93
00:04:26,069 --> 00:04:28,030
And for those of you who'd like to think

94
00:04:28,054 --> 00:04:30,885
that artificial intelligence
will solve all of our problems,

95
00:04:30,909 --> 00:04:33,134
I think we can agree
that we're a long way away

96
00:04:33,158 --> 00:04:35,745
from AI that's able to make sense
of posts like this.

97
00:04:36,856 --> 00:04:39,363
So I'd like to explain
three interlocking issues

98
00:04:39,387 --> 00:04:41,760
that make this so complex

99
00:04:41,784 --> 00:04:44,906
and then think about some ways
we can consider these challenges.

100
00:04:45,348 --> 00:04:49,238
First, we just don't have
a rational relationship to information,

101
00:04:49,262 --> 00:04:50,730
we have an emotional one.

102
00:04:50,754 --> 00:04:54,548
It's just not true that more facts
will make everything OK,

103
00:04:54,572 --> 00:04:57,672
because the algorithms that determine
what content we see,

104
00:04:57,696 --> 00:05:00,823
well, they're designed to reward
our emotional responses.

105
00:05:00,847 --> 00:05:02,228
And when we're fearful,

106
00:05:02,252 --> 00:05:05,426
oversimplified narratives,
conspiratorial explanations

107
00:05:05,450 --> 00:05:08,868
and language that demonizes others
is far more effective.

108
00:05:09,538 --> 00:05:11,412
And besides, many of these companies,

109
00:05:11,436 --> 00:05:13,982
their business model
is attached to attention,

110
00:05:14,006 --> 00:05:17,696
which means these algorithms
will always be skewed towards emotion.

111
00:05:18,371 --> 00:05:22,669
Second, most of the speech
I'm talking about here is legal.

112
00:05:23,081 --> 00:05:24,527
It would be a different matter

113
00:05:24,551 --> 00:05:26,892
if I was talking about
child sexual abuse imagery

114
00:05:26,916 --> 00:05:28,843
or content that incites violence.

115
00:05:28,867 --> 00:05:32,137
It can be perfectly legal
to post an outright lie.

116
00:05:33,130 --> 00:05:37,164
But people keep talking about taking down
"problematic" or "harmful" content,

117
00:05:37,188 --> 00:05:39,797
but with no clear definition
of what they mean by that,

118
00:05:39,821 --> 00:05:41,085
including Mark Zuckerberg,

119
00:05:41,109 --> 00:05:44,521
who recently called for global
regulation to moderate speech.

120
00:05:44,870 --> 00:05:47,085
And my concern is that
we're seeing governments

121
00:05:47,109 --> 00:05:48,401
right around the world

122
00:05:48,425 --> 00:05:51,101
rolling out hasty policy decisions

123
00:05:51,125 --> 00:05:53,871
that might actually trigger
much more serious consequences

124
00:05:53,895 --> 00:05:55,609
when it comes to our speech.

125
00:05:56,006 --> 00:05:59,712
And even if we could decide
which speech to take up or take down,

126
00:05:59,736 --> 00:06:01,910
we've never had so much speech.

127
00:06:01,934 --> 00:06:04,065
Every second, millions
of pieces of content

128
00:06:04,089 --> 00:06:06,196
are uploaded by people
right around the world

129
00:06:06,220 --> 00:06:07,388
in different languages,

130
00:06:07,412 --> 00:06:10,180
drawing on thousands
of different cultural contexts.

131
00:06:10,204 --> 00:06:12,736
We've simply never had
effective mechanisms

132
00:06:12,760 --> 00:06:14,498
to moderate speech at this scale,

133
00:06:14,522 --> 00:06:17,323
whether powered by humans
or by technology.

134
00:06:18,284 --> 00:06:22,228
And third, these companies --
Google, Twitter, Facebook, WhatsApp --

135
00:06:22,252 --> 00:06:25,093
they're part of a wider
information ecosystem.

136
00:06:25,117 --> 00:06:28,469
We like to lay all the blame
at their feet, but the truth is,

137
00:06:28,493 --> 00:06:32,323
the mass media and elected officials
can also play an equal role

138
00:06:32,347 --> 00:06:35,260
in amplifying rumors and conspiracies
when they want to.

139
00:06:35,800 --> 00:06:40,744
As can we, when we mindlessly forward
divisive or misleading content

140
00:06:40,768 --> 00:06:42,053
without trying.

141
00:06:42,077 --> 00:06:43,877
We're adding to the pollution.

142
00:06:45,236 --> 00:06:47,854
I know we're all looking for an easy fix.

143
00:06:47,878 --> 00:06:49,545
But there just isn't one.

144
00:06:49,950 --> 00:06:54,395
Any solution will have to be rolled out
at a massive scale, internet scale,

145
00:06:54,419 --> 00:06:57,680
and yes, the platforms,
they're used to operating at that level.

146
00:06:57,704 --> 00:07:01,176
But can and should we allow them
to fix these problems?

147
00:07:01,668 --> 00:07:02,900
They're certainly trying.

148
00:07:02,924 --> 00:07:07,010
But most of us would agree that, actually,
we don't want global corporations

149
00:07:07,034 --> 00:07:09,366
to be the guardians of truth
and fairness online.

150
00:07:09,390 --> 00:07:11,927
And I also think the platforms
would agree with that.

151
00:07:12,257 --> 00:07:15,138
And at the moment,
they're marking their own homework.

152
00:07:15,162 --> 00:07:16,360
They like to tell us

153
00:07:16,384 --> 00:07:18,963
that the interventions
they're rolling out are working,

154
00:07:18,987 --> 00:07:21,527
but because they write
their own transparency reports,

155
00:07:21,551 --> 00:07:25,169
there's no way for us to independently
verify what's actually happening.

156
00:07:26,431 --> 00:07:29,773
(Applause)

157
00:07:29,797 --> 00:07:32,749
And let's also be clear
that most of the changes we see

158
00:07:32,773 --> 00:07:35,767
only happen after journalists
undertake an investigation

159
00:07:35,791 --> 00:07:37,402
and find evidence of bias

160
00:07:37,426 --> 00:07:40,255
or content that breaks
their community guidelines.

161
00:07:40,815 --> 00:07:45,410
So yes, these companies have to play
a really important role in this process,

162
00:07:45,434 --> 00:07:46,994
but they can't control it.

163
00:07:47,855 --> 00:07:49,373
So what about governments?

164
00:07:49,863 --> 00:07:52,959
Many people believe
that global regulation is our last hope

165
00:07:52,983 --> 00:07:55,863
in terms of cleaning up
our information ecosystem.

166
00:07:55,887 --> 00:07:59,053
But what I see are lawmakers
who are struggling to keep up to date

167
00:07:59,077 --> 00:08:01,418
with the rapid changes in technology.

168
00:08:01,442 --> 00:08:03,346
And worse, they're working in the dark,

169
00:08:03,370 --> 00:08:05,191
because they don't have access to data

170
00:08:05,215 --> 00:08:07,865
to understand what's happening
on these platforms.

171
00:08:08,260 --> 00:08:11,331
And anyway, which governments
would we trust to do this?

172
00:08:11,355 --> 00:08:14,125
We need a global response,
not a national one.

173
00:08:15,419 --> 00:08:17,696
So the missing link is us.

174
00:08:17,720 --> 00:08:20,843
It's those people who use
these technologies every day.

175
00:08:21,260 --> 00:08:25,851
Can we design a new infrastructure
to support quality information?

176
00:08:26,371 --> 00:08:27,601
Well, I believe we can,

177
00:08:27,625 --> 00:08:30,982
and I've got a few ideas about
what we might be able to actually do.

178
00:08:31,006 --> 00:08:34,109
So firstly, if we're serious
about bringing the public into this,

179
00:08:34,133 --> 00:08:36,514
can we take some inspiration
from Wikipedia?

180
00:08:36,538 --> 00:08:38,362
They've shown us what's possible.

181
00:08:38,386 --> 00:08:39,537
Yes, it's not perfect,

182
00:08:39,561 --> 00:08:42,195
but they've demonstrated
that with the right structures,

183
00:08:42,219 --> 00:08:44,854
with a global outlook
and lots and lots of transparency,

184
00:08:44,878 --> 00:08:47,974
you can build something
that will earn the trust of most people.

185
00:08:47,998 --> 00:08:51,160
Because we have to find a way
to tap into the collective wisdom

186
00:08:51,184 --> 00:08:53,493
and experience of all users.

187
00:08:53,517 --> 00:08:56,163
This is particularly the case
for women, people of color

188
00:08:56,187 --> 00:08:57,533
and underrepresented groups.

189
00:08:57,557 --> 00:08:58,723
Because guess what?

190
00:08:58,747 --> 00:09:01,482
They are experts when it comes
to hate and disinformation,

191
00:09:01,506 --> 00:09:05,022
because they have been the targets
of these campaigns for so long.

192
00:09:05,046 --> 00:09:07,396
And over the years,
they've been raising flags,

193
00:09:07,420 --> 00:09:09,085
and they haven't been listened to.

194
00:09:09,109 --> 00:09:10,389
This has got to change.

195
00:09:10,807 --> 00:09:15,133
So could we build a Wikipedia for trust?

196
00:09:15,157 --> 00:09:19,346
Could we find a way that users
can actually provide insights?

197
00:09:19,370 --> 00:09:23,067
They could offer insights around
difficult content-moderation decisions.

198
00:09:23,091 --> 00:09:24,554
They could provide feedback

199
00:09:24,578 --> 00:09:27,619
when platforms decide
they want to roll out new changes.

200
00:09:28,241 --> 00:09:32,403
Second, people's experiences
with the information is personalized.

201
00:09:32,427 --> 00:09:35,070
My Facebook news feed
is very different to yours.

202
00:09:35,094 --> 00:09:37,839
Your YouTube recommendations
are very different to mine.

203
00:09:37,863 --> 00:09:40,355
That makes it impossible for us
to actually examine

204
00:09:40,379 --> 00:09:42,402
what information people are seeing.

205
00:09:42,815 --> 00:09:44,204
So could we imagine

206
00:09:44,228 --> 00:09:49,006
developing some kind of centralized
open repository for anonymized data,

207
00:09:49,030 --> 00:09:51,894
with privacy and ethical
concerns built in?

208
00:09:52,220 --> 00:09:53,998
Because imagine what we would learn

209
00:09:54,022 --> 00:09:57,283
if we built out a global network
of concerned citizens

210
00:09:57,307 --> 00:10:00,601
who wanted to donate
their social data to science.

211
00:10:01,141 --> 00:10:02,863
Because we actually know very little

212
00:10:02,887 --> 00:10:05,768
about the long-term consequences
of hate and disinformation

213
00:10:05,792 --> 00:10:07,767
on people's attitudes and behaviors.

214
00:10:08,236 --> 00:10:09,403
And what we do know,

215
00:10:09,427 --> 00:10:11,620
most of that has been
carried out in the US,

216
00:10:11,644 --> 00:10:14,025
despite the fact that
this is a global problem.

217
00:10:14,049 --> 00:10:15,684
We need to work on that, too.

218
00:10:16,192 --> 00:10:17,342
And third,

219
00:10:17,366 --> 00:10:19,676
can we find a way to connect the dots?

220
00:10:19,700 --> 00:10:23,138
No one sector, let alone nonprofit,
start-up or government,

221
00:10:23,162 --> 00:10:24,584
is going to solve this.

222
00:10:24,608 --> 00:10:27,172
But there are very smart people
right around the world

223
00:10:27,196 --> 00:10:28,577
working on these challenges,

224
00:10:28,601 --> 00:10:32,177
from newsrooms, civil society,
academia, activist groups.

225
00:10:32,201 --> 00:10:34,099
And you can see some of them here.

226
00:10:34,123 --> 00:10:37,050
Some are building out indicators
of content credibility.

227
00:10:37,074 --> 00:10:38,320
Others are fact-checking,

228
00:10:38,344 --> 00:10:41,935
so that false claims, videos and images
can be down-ranked by the platforms.

229
00:10:41,959 --> 00:10:44,172
A nonprofit I helped
to found, First Draft,

230
00:10:44,196 --> 00:10:47,164
is working with normally competitive
newsrooms around the world

231
00:10:47,188 --> 00:10:50,691
to help them build out investigative,
collaborative programs.

232
00:10:51,231 --> 00:10:53,540
And Danny Hillis, a software architect,

233
00:10:53,564 --> 00:10:55,945
is designing a new system
called The Underlay,

234
00:10:55,969 --> 00:10:58,744
which will be a record
of all public statements of fact

235
00:10:58,768 --> 00:11:00,097
connected to their sources,

236
00:11:00,121 --> 00:11:03,775
so that people and algorithms
can better judge what is credible.

237
00:11:04,800 --> 00:11:08,156
And educators around the world
are testing different techniques

238
00:11:08,180 --> 00:11:11,664
for finding ways to make people
critical of the content they consume.

239
00:11:12,633 --> 00:11:15,774
All of these efforts are wonderful,
but they're working in silos,

240
00:11:15,798 --> 00:11:18,478
and many of them are woefully underfunded.

241
00:11:18,502 --> 00:11:20,555
There are also hundreds
of very smart people

242
00:11:20,579 --> 00:11:22,231
working inside these companies,

243
00:11:22,255 --> 00:11:24,580
but again, these efforts
can feel disjointed,

244
00:11:24,604 --> 00:11:28,541
because they're actually developing
different solutions to the same problems.

245
00:11:29,205 --> 00:11:31,474
How can we find a way
to bring people together

246
00:11:31,498 --> 00:11:34,776
in one physical location
for days or weeks at a time,

247
00:11:34,800 --> 00:11:37,196
so they can actually tackle
these problems together

248
00:11:37,220 --> 00:11:39,038
but from their different perspectives?

249
00:11:39,062 --> 00:11:40,402
So can we do this?

250
00:11:40,426 --> 00:11:43,665
Can we build out a coordinated,
ambitious response,

251
00:11:43,689 --> 00:11:47,398
one that matches the scale
and the complexity of the problem?

252
00:11:47,819 --> 00:11:49,192
I really think we can.

253
00:11:49,216 --> 00:11:52,175
Together, let's rebuild
our information commons.

254
00:11:52,819 --> 00:11:54,009
Thank you.

255
00:11:54,033 --> 00:11:57,761
(Applause)