1
00:00:33,250 --> 00:00:37,900
OK, everybody, thank you for coming this evening, welcome to the Oxford Martin School.

2
00:00:37,900 --> 00:00:43,300
I'm Alan Robin. I oversee the research portfolio of programmes that we fund here at the school.

3
00:00:43,300 --> 00:00:47,860
I'm delighted to host the Kerry Madsen to tonight.

4
00:00:47,860 --> 00:00:56,110
So Jens is a classic example of an interdisciplinary researcher. His research focuses on how people form beliefs and why they act the way they do.

5
00:00:56,110 --> 00:01:02,410
His main areas of interest are information integration, source credibility, dynamic decision making,

6
00:01:02,410 --> 00:01:08,920
the dissemination of misinformation, political persuasion and complex environmental information systems management.

7
00:01:08,920 --> 00:01:16,900
And he's taken those interests into a wide range of contexts. He's an Oxford Martin fellow in our Programme for Sustainable Oceans and a post-doctoral

8
00:01:16,900 --> 00:01:22,450
fellow at the Complex Human Environmental Systems Simulations Lab in the School of Geography.

9
00:01:22,450 --> 00:01:28,900
So he works on simulating the behaviour of fisheries that to best understand complex and dynamic models.

10
00:01:28,900 --> 00:01:34,870
But he's here tonight to discuss his book on the psychology of micro-targeted election campaigns,

11
00:01:34,870 --> 00:01:42,850
and this examines the capabilities of increasingly sophisticated models of persuasion in elections and its implications for democracy.

12
00:01:42,850 --> 00:01:47,860
So again, it's going to speak for about 40 minutes, and we'll move straight into a question and answer session,

13
00:01:47,860 --> 00:01:52,360
and we'll have a drinks reception in the cafe afterwards and you all warmly invited to it.

14
00:01:52,360 --> 00:01:59,680
And lastly, just to note that the lecture is videoed, as is the Q&A, so it is going to be filmed and live webcast.

15
00:01:59,680 --> 00:02:03,970
So bear that in mind. If you don't want to be filmed, don't answer it. Ask a question.

16
00:02:03,970 --> 00:02:15,040
So without further ado, let me welcome you to the stage.

17
00:02:15,040 --> 00:02:21,790
Yeah, hello, everyone, and thank you so much for coming on a Friday night instead of having a pint in the pub,

18
00:02:21,790 --> 00:02:29,500
quite impressed that I actually managed to get so many people. So I'm absolutely delighted and I've massively misled everyone.

19
00:02:29,500 --> 00:02:34,960
I am going to talk about the halibut. And it is going to be a long talk about sustainability of oceans as well.

20
00:02:34,960 --> 00:02:38,530
And without further ado, the tuna now.

21
00:02:38,530 --> 00:02:46,420
But all kidding aside, my job here at the University of Oxford and the Oxford Martin School is to understand what makes a

22
00:02:46,420 --> 00:02:52,750
person form a particular viewpoint of the world and what makes that person act in a particular way.

23
00:02:52,750 --> 00:03:01,450
And I've been doing this from a computational perspective, which means I develop mathematical models that I can test and try and refine in

24
00:03:01,450 --> 00:03:07,990
order to explain how beliefs and behaviours or form both individually and socially.

25
00:03:07,990 --> 00:03:14,440
And so while this is something that can be applied to fisheries, as we've been doing in my spare time,

26
00:03:14,440 --> 00:03:19,300
I wrote this book on the psychology of microtargeting campaigns.

27
00:03:19,300 --> 00:03:25,600
And so this is something that is near and dear to my heart because it's something fundamentally democratic and it's something

28
00:03:25,600 --> 00:03:32,350
where I could apply the same kind of modelling principles that I've been working with in the fisheries and perspective,

29
00:03:32,350 --> 00:03:38,260
but information theory and to do electoral decisions.

30
00:03:38,260 --> 00:03:45,850
And this sort of ties back to my original and true sort of foundation of where my undergraduate degree comes from,

31
00:03:45,850 --> 00:03:49,240
which is rhetorical theory, the art of persuasion.

32
00:03:49,240 --> 00:03:55,790
So basically, I've always been interested in what makes someone tick and what makes someone believe what they believe.

33
00:03:55,790 --> 00:04:00,500
So why did I write this book and all right.

34
00:04:00,500 --> 00:04:06,560
So first of all, I think we've all heard a lot about micro-targeted election campaigns in the last couple of years,

35
00:04:06,560 --> 00:04:08,510
in particular with Cambridge Analytica.

36
00:04:08,510 --> 00:04:16,100
I'm going to talk about this a tiny bit, but I'm going to try and veer away from Cambridge Analytica as much as possible as possible.

37
00:04:16,100 --> 00:04:20,460
Because while this is an interesting case, study might not tell the full story.

38
00:04:20,460 --> 00:04:26,540
And so I was really interested in figuring out an investigation of the underlying microtargeting principles.

39
00:04:26,540 --> 00:04:33,950
So basically, what is it that allows a model with data about specific people to build models of

40
00:04:33,950 --> 00:04:38,870
those people to persuade them of something or to make them act in a particular way?

41
00:04:38,870 --> 00:04:44,270
For me, that's a fundamental question to our democratic sustainability because if there

42
00:04:44,270 --> 00:04:49,520
are models out there and companies out there who can target of innermost desires,

43
00:04:49,520 --> 00:04:54,350
innermost beliefs and addresses in a way that is incredibly persuasive to us,

44
00:04:54,350 --> 00:05:04,280
and it is worth knowing how that amount of power can be can be used and puts it potentially abused in a democratic system.

45
00:05:04,280 --> 00:05:10,940
So in order to understand the degree to which our information systems are fragile and vulnerable to misinformation,

46
00:05:10,940 --> 00:05:19,460
we need to understand the psychology of individuals, as well as how these models can be used to generate these individualised models.

47
00:05:19,460 --> 00:05:22,880
So the intended use of the book is not academic per se,

48
00:05:22,880 --> 00:05:31,070
although obviously it builds on academic work that I've been doing for the past decade and work that I've been doing with colleagues,

49
00:05:31,070 --> 00:05:35,690
and I wanted to buy a book that was more geared as a broad perspective,

50
00:05:35,690 --> 00:05:38,270
which can be used as a friend owes journalists,

51
00:05:38,270 --> 00:05:45,620
students and even regulators and network providers to understand the type of modelling that they engage with,

52
00:05:45,620 --> 00:05:52,340
in particular, if you want to regulate the use of faith and politics. We have to understand what these models can and cannot do.

53
00:05:52,340 --> 00:06:00,440
And there's a lot of hyperbole around with a lot of newspaper articles saying data can predict everything or we,

54
00:06:00,440 --> 00:06:04,640
the human mind, is so infinitely complex that mathematical models can say nothing.

55
00:06:04,640 --> 00:06:09,650
Why do you like Nescafé coffee when Aly is better or whatever?

56
00:06:09,650 --> 00:06:13,700
Like, there's always going to be hyperbole on both sides of that aisle.

57
00:06:13,700 --> 00:06:18,140
And obviously, as with most things, the truth probably lies in the middle.

58
00:06:18,140 --> 00:06:23,270
And where data can say something about your psychology, I probably can't say everything.

59
00:06:23,270 --> 00:06:26,810
And so if you really want to understand how we can improve political discourse

60
00:06:26,810 --> 00:06:32,810
and which I seem to sort of feel this kind of at an all time low at the moment,

61
00:06:32,810 --> 00:06:36,740
unfortunately, but like, let's hope not.

62
00:06:36,740 --> 00:06:40,610
Well, let's hope it isn't. An all time low is bouncing back, I should say.

63
00:06:40,610 --> 00:06:46,940
But if you really want to understand how we can improve political discourse and the kind of democratic institutions that we have,

64
00:06:46,940 --> 00:06:51,980
we have to understand the fundamental nature of the psychology of the citizens and of the electorate,

65
00:06:51,980 --> 00:06:57,530
and also understand a how they can be modelled and persuaded by the people in power.

66
00:06:57,530 --> 00:07:06,140
And B, how we can set up institutions and protections to make sure that those are not being abused in an unlawful manner.

67
00:07:06,140 --> 00:07:14,150
And finally, I think it bears saying that if we don't do this at a university where all of our information is disseminated publicly,

68
00:07:14,150 --> 00:07:18,530
it will be done behind closed doors like what was done.

69
00:07:18,530 --> 00:07:26,930
Cambridge Analytica, for instance, was not publicly available and the whole principle behind doing a thorough academic study of persuasion,

70
00:07:26,930 --> 00:07:33,650
a thorough academic study of micro-targeting is that we can get the results into the hands of people who might

71
00:07:33,650 --> 00:07:40,850
want to consider who to vote for in an election and why we want to structure society in the way that we want to.

72
00:07:40,850 --> 00:07:46,370
So obviously the most famous case of micro-targeting in the last couple of years, I've been Cambridge Analytica.

73
00:07:46,370 --> 00:07:54,890
I'm sure that you will know it was the 2016 so-called hacking of the US presidential election,

74
00:07:54,890 --> 00:07:59,630
in which a London based company used personalised data to draw up individual

75
00:07:59,630 --> 00:08:06,100
profiles of citizens in the US to optimise their persuasion attempts to solve.

76
00:08:06,100 --> 00:08:13,860
Yeah, basically do a shrewd marketing campaign. So that's a really interesting case study, and it got a lot of attention.

77
00:08:13,860 --> 00:08:19,470
But I think it's an inadequate example and it's a good historical case,

78
00:08:19,470 --> 00:08:23,130
but it's inadequate to understand the fundamental principles that underpin

79
00:08:23,130 --> 00:08:29,550
these models and these ways of using data to to generate persuasion campaigns.

80
00:08:29,550 --> 00:08:33,120
Why? Well, first, because case studies can't be generalised.

81
00:08:33,120 --> 00:08:39,930
They represent a single data point, and the social cultural context and the candidates of any given election will differ.

82
00:08:39,930 --> 00:08:42,360
And they're necessarily unique to the situation.

83
00:08:42,360 --> 00:08:51,660
If you think that you can retrofit a model of what Cambridge Analytica did in 2016 or even 2014 when they were involved in the midterm elections,

84
00:08:51,660 --> 00:08:55,170
you are sadly mistaken because the situation is going to be different.

85
00:08:55,170 --> 00:09:01,810
It's not going to be Hillary Clinton versus Donald Trump, presumably, although some rumours has it that she will run.

86
00:09:01,810 --> 00:09:09,210
But let's see. And so you can't necessarily rely on that individual case study.

87
00:09:09,210 --> 00:09:13,440
Furthermore, you don't know exactly if Cambridge Analytica really influence the situation.

88
00:09:13,440 --> 00:09:18,870
Would Trump have won, regardless of the involvement of Cambridge Analytica is literally impossible to say,

89
00:09:18,870 --> 00:09:23,700
given the amount of and confounding variables that are in that situation.

90
00:09:23,700 --> 00:09:27,570
Secondly, the methods are very specific to the situation.

91
00:09:27,570 --> 00:09:33,630
So, for instance, the API that they use to generate personalities of individual citizens is no longer available.

92
00:09:33,630 --> 00:09:38,490
So if you're going to go into a big discussion about like, Oh, we should protect the API of the citizens.

93
00:09:38,490 --> 00:09:45,690
Done and done. Problem solved. But that doesn't mean that the intention of getting that data is not out there.

94
00:09:45,690 --> 00:09:51,090
So they appeared as well to have relied on some pop psychology. So the effectiveness of it,

95
00:09:51,090 --> 00:09:57,180
which is unknown and further journalists and people who have been involved in Cambridge

96
00:09:57,180 --> 00:10:01,500
Analytica may have an interest in beefing up the impact of Cambridge Analytica.

97
00:10:01,500 --> 00:10:07,650
So maybe a bit of sensationalism. So then further, in future campaigns,

98
00:10:07,650 --> 00:10:16,470
elections may have better with data because the type of data that we have available for any given person will constantly shift.

99
00:10:16,470 --> 00:10:22,170
Like, for instance, in 2016, no one was talking about Tik. That's a really important app right now.

100
00:10:22,170 --> 00:10:25,770
And for anyone who doesn't know what tick tock is, it's a massive app.

101
00:10:25,770 --> 00:10:30,930
Social media app that is out of China, which is collecting a lot of data on people.

102
00:10:30,930 --> 00:10:39,690
So if we don't understand what TikTok does in a fundamental principle about gathering data from individual citizens,

103
00:10:39,690 --> 00:10:46,980
we risk running into a retro sort of prospective way of doing regulation and analysis.

104
00:10:46,980 --> 00:10:53,070
So in that case, like, we need to understand the micro-targeting in principle rather than in cases.

105
00:10:53,070 --> 00:10:54,570
And that's a really key point.

106
00:10:54,570 --> 00:11:00,840
So while I applaud all the journalism and it is a vital piece of journalistic work that's been done on Cambridge Analytica,

107
00:11:00,840 --> 00:11:08,190
it would be a bit like saying, OK, how do we design the best football team to beat Real Madrid?

108
00:11:08,190 --> 00:11:12,480
Well, I will. I will show you because I am Danish.

109
00:11:12,480 --> 00:11:16,960
I am from a tiny little town called on site and in a glorious full day.

110
00:11:16,960 --> 00:11:21,900
Or maybe actually sorry, spring day of 1994,

111
00:11:21,900 --> 00:11:32,190
all into Bullet Club went to the mighty Bernabeu and beat Real Madrid two nil, knocking them out of the UEFA Cup Shimbun.

112
00:11:32,190 --> 00:11:33,180
So what does that mean?

113
00:11:33,180 --> 00:11:40,900
Clearly, if we just extrapolate from a single case study, it means that clearly we have to wear yellow if we want to beat Real Madrid.

114
00:11:40,900 --> 00:11:45,270
That's a really good explanation, and we probably have to have a lot of things.

115
00:11:45,270 --> 00:11:55,140
That's a really good explanation as well. But unfortunately, if any of you have ever followed the subsequent endeavours of my beloved hometown team,

116
00:11:55,140 --> 00:12:01,800
you will have realised that they are not at the pinnacle of the European scene lifting the Champions League trophy year after year.

117
00:12:01,800 --> 00:12:07,280
And so extrapolating from one case, maybe a slightly dicey thing to do.

118
00:12:07,280 --> 00:12:12,470
So what I want to do today is basically talk about three aspects of microtargeting,

119
00:12:12,470 --> 00:12:21,800
and I want to talk a bit about how you can model subjectivity because that seems like a nebulous thing, but it has been done and can be done.

120
00:12:21,800 --> 00:12:27,800
Then I will try and make this as a transition into what I call analytic micro-targeting.

121
00:12:27,800 --> 00:12:32,510
So this is reportedly what Cambridge Analytica did.

122
00:12:32,510 --> 00:12:36,860
And these are great models about segmentation of peoples.

123
00:12:36,860 --> 00:12:47,000
But we'll see some of the limitations as we go into the third section, which is dynamic microtargeting, and then we'll end with a bit of Q&A.

124
00:12:47,000 --> 00:12:52,880
So part one is models of subjectivity. So how does it capture your beliefs?

125
00:12:52,880 --> 00:13:00,470
Because obviously, we know that individual differences are absolutely critical. And obviously, everyone's political leanings differ.

126
00:13:00,470 --> 00:13:05,930
You can have a wide range and wide range of beliefs that barrier within the population,

127
00:13:05,930 --> 00:13:09,650
both in terms of what they think is important and what they think is credible or not.

128
00:13:09,650 --> 00:13:15,560
So, for instance, some people might think like, Oh, what really matters to me when I choose a candidate is the economy.

129
00:13:15,560 --> 00:13:21,740
And it is their stance on jobs and labour market policy.

130
00:13:21,740 --> 00:13:27,500
Whilst another person might think like what really matters to me is people who want to ensure women's rights or who want

131
00:13:27,500 --> 00:13:35,630
to understand socio economic consequences of poverty or who want to safeguard our society against environmental disaster.

132
00:13:35,630 --> 00:13:47,180
So. Those people will have different proclivities towards what they really value in terms of what's important in this election.

133
00:13:47,180 --> 00:13:51,350
But they will also have different personal beliefs about the likelihood of outcomes.

134
00:13:51,350 --> 00:14:02,810
So for instance, if you talk to a climate change denier and if you talk to someone who has looked at the data and is a climate change believer,

135
00:14:02,810 --> 00:14:08,720
you might have very different subjective perspectives on the likelihood of particular outcomes.

136
00:14:08,720 --> 00:14:09,260
Like, for instance,

137
00:14:09,260 --> 00:14:19,160
what's the likelihood that Miami will suffer from severe flooding and if weather conditions in the next decade or 20 years or 30 years?

138
00:14:19,160 --> 00:14:23,090
If you ask a climate change denier, they might say like, Well, do you know?

139
00:14:23,090 --> 00:14:27,710
I don't think that's very likely because like even though there's been a couple of like storms or whatever

140
00:14:27,710 --> 00:14:34,250
like weather comes and goes like there was an Ice Age once there was like a solar flare or whatever,

141
00:14:34,250 --> 00:14:40,550
and they're entitled to their opinions and they're entitled to their subjective perception of the reality.

142
00:14:40,550 --> 00:14:46,010
Similarly of comparatively. If you talk to someone who believes in climate change, they might say like, Yeah,

143
00:14:46,010 --> 00:14:54,020
I expect those kind of extreme weather patterns to increase with time because we're getting into a worse situation in terms of climate change.

144
00:14:54,020 --> 00:15:00,200
So one so a perspective of this is like all like human beings are too messy will subjective.

145
00:15:00,200 --> 00:15:05,780
We're all just like bulldozing down the road of of our own personal beliefs.

146
00:15:05,780 --> 00:15:10,280
And it's kind of impossible to to conceptualise and model.

147
00:15:10,280 --> 00:15:14,810
But there's been some strides in cognitive psychology to address this.

148
00:15:14,810 --> 00:15:19,190
So I'm going to simplify this by four pathways to personal beliefs.

149
00:15:19,190 --> 00:15:26,630
And so I should note that in all of these pathways and I along with colleagues,

150
00:15:26,630 --> 00:15:33,740
some of which are in the room today, Toby and we have been doing work on what is known as Bayesian modelling,

151
00:15:33,740 --> 00:15:39,710
and I'll explain what that is in a bid to try and account for how people use their subjective probability

152
00:15:39,710 --> 00:15:45,500
estimations of the world to predict how they will change their belief given new information.

153
00:15:45,500 --> 00:15:51,050
But there's no two ways you can get new information about the world. So, for instance, you can have an personal experience.

154
00:15:51,050 --> 00:16:01,940
So if you just fired from a job and you're asked to evaluate the volatility of the labour market, you might say like, it's pretty damn volatile.

155
00:16:01,940 --> 00:16:06,620
I am not feeling good and I'm not feeling good about the prospects of my future.

156
00:16:06,620 --> 00:16:13,760
That's an entirely sensible thing to say if you like she has been canned walking away as you do with your P45.

157
00:16:13,760 --> 00:16:20,120
So there's obviously personal experiences that will drive a subjective perception of how the world looks.

158
00:16:20,120 --> 00:16:23,930
But it's not the only way that we're getting information about the world.

159
00:16:23,930 --> 00:16:28,220
Probably most of the information we get about the world is from other people.

160
00:16:28,220 --> 00:16:33,590
So we read news reports and we read meteorological reports about whether or not it's going to rain

161
00:16:33,590 --> 00:16:39,140
in Oxford this afternoon and which I believed it was going to be because I looked at the BBC.

162
00:16:39,140 --> 00:16:46,490
And so I consulted a source who could give me some report about something that I was interested in some hypothesis.

163
00:16:46,490 --> 00:16:51,260
Now I can have a greater or lesser extent to which I believe in that source.

164
00:16:51,260 --> 00:17:00,440
So for instance, if a drunk person stops me in the street and goes like the economy is going to crash tomorrow, I'm going to go like Hamlet.

165
00:17:00,440 --> 00:17:05,300
That's that's that's a really solid prediction, man. And like pop back to the pub, please.

166
00:17:05,300 --> 00:17:07,820
And I might not revise my beliefs about it.

167
00:17:07,820 --> 00:17:15,110
Like significantly, comparatively, if the Bank of England comes out with a report saying the economy is going to crash tomorrow,

168
00:17:15,110 --> 00:17:23,450
and I'll probably take that a bit more seriously. Why? Because I trust they are going to disseminate information to the best of their knowledge,

169
00:17:23,450 --> 00:17:27,290
and I trust that they have some expertise in that particular area.

170
00:17:27,290 --> 00:17:32,360
So it makes sense that I would skew my belief or adjust my belief provision differently,

171
00:17:32,360 --> 00:17:40,610
depending on whether or not I trust that particular source to be a credible source of information for that particular domain.

172
00:17:40,610 --> 00:17:44,330
So trustworthiness, for instance, is something that cuts across domains,

173
00:17:44,330 --> 00:17:52,010
and I would trust my brother in almost any situation, like following maybe playing a game with him.

174
00:17:52,010 --> 00:17:56,750
But would I have a high rate of expertise for everything with my brother?

175
00:17:56,750 --> 00:18:01,040
No, that's how I dependent on some things I know that he knows or not.

176
00:18:01,040 --> 00:18:03,410
And on some things, I know that he doesn't know that much.

177
00:18:03,410 --> 00:18:08,930
So if he advises me on something where I know like, OK, he's got a lot of expertise in this, he's got my best interests at heart.

178
00:18:08,930 --> 00:18:11,900
So I trust what he's saying. I'm going to listen to him.

179
00:18:11,900 --> 00:18:18,980
But conversely, if I consult him on something where I know he knows nothing, then even though I trust him,

180
00:18:18,980 --> 00:18:27,200
I'm not really going to trust my body so much because my subjective perception of his expertise will differ in that case.

181
00:18:27,200 --> 00:18:32,450
So you can see that the subjective assignments of probability that we can give to

182
00:18:32,450 --> 00:18:37,770
people can start influencing how are we going to treat information from that source?

183
00:18:37,770 --> 00:18:48,240
So this is just to sort of say in mostly kind of just words rather than in maths, that we have our subjective estimation,

184
00:18:48,240 --> 00:18:52,590
both of the particular case that we're considering like the hypothesis in question.

185
00:18:52,590 --> 00:18:59,980
How likely do I think this is? Then you've got some information coming in which you can think of as more or less strong.

186
00:18:59,980 --> 00:19:11,160
And so like you can think like, Oh, if I'm going to if every billionaire like I report came out saying every billionaire has sold the house in London,

187
00:19:11,160 --> 00:19:15,060
like this week, you go like, Oh, there's something going on here, like something.

188
00:19:15,060 --> 00:19:18,870
Something is afoot in terms of London's housing market.

189
00:19:18,870 --> 00:19:24,120
And you would probably adjust your belief in the crack in the volatility of that housing market accordingly.

190
00:19:24,120 --> 00:19:31,170
But that's because the strength of the relationship between the evidence and the hypothesis is quite strong, comparatively.

191
00:19:31,170 --> 00:19:37,050
If I see a bunch of semi random stuff that has little to no relationship to that hypothesis,

192
00:19:37,050 --> 00:19:41,310
I can see a lot of information like butterflies floating the wrong wings or something of that.

193
00:19:41,310 --> 00:19:47,010
But I'm not really going to adjust my belief in that hypothesis because I don't think there's any relationship between those things.

194
00:19:47,010 --> 00:19:53,370
Now the key difference here is I may think that there's a strong causal relationship

195
00:19:53,370 --> 00:20:01,380
between two pieces of information or a piece of information and an hypothesis. But another person might think that that is a spurious causality.

196
00:20:01,380 --> 00:20:10,170
So we may disagree subjectively on the prior belief as in like, how likely do I think this is before I listen to any information?

197
00:20:10,170 --> 00:20:19,460
Yes, then we can disagree on the degree to which this evidence is strong or weak, or positively or negatively correlated with that hypothesis.

198
00:20:19,460 --> 00:20:23,330
But then we can also disagree on whether or not that source is trustworthy.

199
00:20:23,330 --> 00:20:28,940
So for instance, let's imagine that you have a climate change denier who thinks like, you know what?

200
00:20:28,940 --> 00:20:34,520
Climate scientists are paid by universities to tow a particular company line,

201
00:20:34,520 --> 00:20:42,830
much in the same way that people from Exxon and show paid to tow a particular party line or company line.

202
00:20:42,830 --> 00:20:44,360
They may then think like, Well,

203
00:20:44,360 --> 00:20:54,530
there's no operational difference between the credibility of a scientist and someone from Shell or of other petrol companies,

204
00:20:54,530 --> 00:21:00,440
in which case they should operationally adjust their beliefs in the exact same manner.

205
00:21:00,440 --> 00:21:09,020
Given reports from scientists and reports from oil company executives, and those can be subjective again.

206
00:21:09,020 --> 00:21:13,850
Thirdly, you may think that people are dependent on each other as in life.

207
00:21:13,850 --> 00:21:20,300
I think these guys have talked to each other before giving this report, or you can have different perceptions of causality.

208
00:21:20,300 --> 00:21:27,710
So all of this, these kind of different perspectives on how evidence links to hypotheses, how sources are related to each other, like, for instance,

209
00:21:27,710 --> 00:21:37,280
if you see three people and they all say the economy is going to boom after Brexit or something or the economy is going to crash after Brexit,

210
00:21:37,280 --> 00:21:43,290
then it makes a [INAUDIBLE] of a lot of difference. If you learn that those three people are drawing the same,

211
00:21:43,290 --> 00:21:49,580
drawing their conclusions from the exact same report because they're basically just all saying the exact same thing from the exact same report,

212
00:21:49,580 --> 00:21:54,890
in which case the real thing that you should worry about is what's the credibility of that report?

213
00:21:54,890 --> 00:22:00,260
Not the amount of people who independently say what they say or independently in this case.

214
00:22:00,260 --> 00:22:09,710
So the computational impact to people's been subjective beliefs has been fleshed out by this wonderful priest, Thomas Bayes.

215
00:22:09,710 --> 00:22:15,590
There's also a French mathematician called Laplace, and he actually formalised it a bit better.

216
00:22:15,590 --> 00:22:21,050
So. But I mean, seeing that face is buried in London, I'm going to go with bass.

217
00:22:21,050 --> 00:22:28,940
So what this basically tells you is that you can conceptualise people's subjective beliefs as a spectrum between zero and one on

218
00:22:28,940 --> 00:22:36,800
hypotheses prior to any kind of evidence as related to the strength of the diagnostic city of the evidence as related to that hypothesis.

219
00:22:36,800 --> 00:22:40,700
And you can integrate those two to figure out, given your prior belief.

220
00:22:40,700 --> 00:22:44,380
And this information, what should you now believe?

221
00:22:44,380 --> 00:22:52,840
Depending on your subjective understanding of that relationship, so to give you an example of this, his model that I've been working with for a while,

222
00:22:52,840 --> 00:22:58,870
which tries to conceptualise exactly what I was talking about earlier and the probability of trustworthiness,

223
00:22:58,870 --> 00:23:05,050
probability of expertise, PVH is the probability of the hypothesis before you hear a report from that source.

224
00:23:05,050 --> 00:23:13,700
And h given is what's the likelihood of this hypothesis, given that I'm that this source is giving me a positive report about this?

225
00:23:13,700 --> 00:23:16,220
And we've done loads of testing on this.

226
00:23:16,220 --> 00:23:26,270
And just to give you a couple of examples, this is us trying to predict and posterior degrees of belief after someone gets a report from a source.

227
00:23:26,270 --> 00:23:35,210
As you can see, the dotted line and the solid line are almost identical, and those are the predictions versus observations.

228
00:23:35,210 --> 00:23:40,130
Similarly, I did a study in 2016 on on political candidates,

229
00:23:40,130 --> 00:23:45,770
and we could we could predict for Republicans and Democrats alike exactly how they would change their beliefs,

230
00:23:45,770 --> 00:23:53,210
given reports from candidates given their subjective perceptions of the credibility of that person.

231
00:23:53,210 --> 00:23:57,050
So it is possible to do mathematical modelling of people subjective beliefs.

232
00:23:57,050 --> 00:24:03,440
But the question is how do I start to populate those models? Like, how can I start doing my analytic microtargeting?

233
00:24:03,440 --> 00:24:07,550
So I have segmentation of that population? And there's loads of ways that you can do this.

234
00:24:07,550 --> 00:24:14,990
So one thing is traditional segmentation. So for instance, if you know that certain demographic information or certain information about

235
00:24:14,990 --> 00:24:20,030
the population is correlated with a particular thing like so for instance,

236
00:24:20,030 --> 00:24:25,700
these are income, education, gender shopping habits or hobbies.

237
00:24:25,700 --> 00:24:33,350
So for instance, if you are an American and you buy a foreign beer, you are eighty three percent likely to be a Democrat.

238
00:24:33,350 --> 00:24:44,690
So if I know if I get access to your target loyalty card, which may be sold and I go through your shopping list,

239
00:24:44,690 --> 00:24:48,200
I can see like, Oh, this person's buying a lot of organic products.

240
00:24:48,200 --> 00:24:53,870
This person is buying a lot of foreign beer, like slightly troubling amount of foreign beer,

241
00:24:53,870 --> 00:24:57,860
but you get my drift that you can start building up models of.

242
00:24:57,860 --> 00:24:59,330
What is this person like?

243
00:24:59,330 --> 00:25:06,910
What is this person probably going to value in terms of priorities, beliefs like is this person conscious about environment or not?

244
00:25:06,910 --> 00:25:15,680
And like, if you just have like a slab of red meat, you're probably not like the massive environmentalist that you would otherwise proposed to be.

245
00:25:15,680 --> 00:25:20,990
Same thing you can do digital segmentation. So like, for instance, what kind of apps do you like?

246
00:25:20,990 --> 00:25:29,690
How how do you use your phone and where do you move in the city, if I can to you locate you, I can see like, Oh, you went to the extinction.

247
00:25:29,690 --> 00:25:37,460
Extinction Rebellion every time they were in town, you probably care about the environment and we can do word clouds.

248
00:25:37,460 --> 00:25:43,790
So, for instance, this is a cloud that has been segmented in between Democrats and Republicans.

249
00:25:43,790 --> 00:25:49,880
So it turns out that in this analysis, if you say Democrats, a lot of the tweets, you're probably a Republican.

250
00:25:49,880 --> 00:25:53,270
But if you say Democratic, you're probably a Democrat.

251
00:25:53,270 --> 00:26:00,590
So imagine that I'm going to scrape your Twitter account for all the words that you've ever posted in your thousands and thousands of tweets,

252
00:26:00,590 --> 00:26:05,510
and I can start doing analysis of what you probably think about things.

253
00:26:05,510 --> 00:26:12,770
What do you care about? How do you write about it? And you can start correlating those with subjective perspectives on the likelihood of these things.

254
00:26:12,770 --> 00:26:18,470
Like, for instance, if you like, we're all going to die after, like in 10 years.

255
00:26:18,470 --> 00:26:25,550
Hashtag environment. The likelihood is that you probably have a pretty strong perception on the prior on on climate change.

256
00:26:25,550 --> 00:26:32,690
Yes. So you can train models to exactly do this same thing with your social position in the network.

257
00:26:32,690 --> 00:26:39,020
And you can also use psychometrics such as personality measures and the wonderful media.

258
00:26:39,020 --> 00:26:45,340
It is also from the University of Cambridge. Feel free to boo, though.

259
00:26:45,340 --> 00:26:50,600
And from Cambridge, he's done a lot of stellar work on personality pressures, but equally,

260
00:26:50,600 --> 00:26:56,210
this moral foundations, heuristics and biases like confirmation bias and the need for closure.

261
00:26:56,210 --> 00:27:02,360
As in like, how willing are you to accept that something is open ended and not sort of settled?

262
00:27:02,360 --> 00:27:06,170
So you can imagine that you can start drawing a lot of data from people,

263
00:27:06,170 --> 00:27:14,630
either by asking a representative sample and then correlating those responses with other people in the population who exhibit the same kind of traits.

264
00:27:14,630 --> 00:27:17,750
So if you learn that like, for instance, to be an example,

265
00:27:17,750 --> 00:27:25,760
that people who buy foreign beer are more likely to be Democrats than if you can get hands your hands on a lot of data about people's shopping habits,

266
00:27:25,760 --> 00:27:28,670
you're going to have one data point that pushes them towards the Democrats.

267
00:27:28,670 --> 00:27:33,290
But then they also own SUV, which maybe puts them in a different direction.

268
00:27:33,290 --> 00:27:41,570
So you can get that you can start building up these individualised models where you build up a perception of that person,

269
00:27:41,570 --> 00:27:45,080
which you can then later use to segment the population.

270
00:27:45,080 --> 00:27:51,170
And importantly as well, you can separate your campaign into what is like beliefs and behaviour phases.

271
00:27:51,170 --> 00:27:56,450
So like the beliefs is how do you change your belief, which is dependent on how you see the credibility of the source,

272
00:27:56,450 --> 00:28:01,310
the likelihood of that evidence, your prior priorities for policies?

273
00:28:01,310 --> 00:28:04,610
But it's also important to know that some people need to go out and vote for your

274
00:28:04,610 --> 00:28:08,780
candidate so you want to have a campaign phase as well and get out the vote phase,

275
00:28:08,780 --> 00:28:14,270
which basically says, I know that you're in my base is I'm not going to contact you through the persuasion phase,

276
00:28:14,270 --> 00:28:23,300
but I am damn sure going to sound like a person to your car, to your house to remind you, is Election Day because I want you to turn up.

277
00:28:23,300 --> 00:28:27,950
So this requires persuasion. This requires get out the vote.

278
00:28:27,950 --> 00:28:39,620
Finally, you can in the constituency based elections like the UK and the US, you can start segmenting people along the lines of Do you even matter?

279
00:28:39,620 --> 00:28:44,420
So for instance, if I find a Republican in California or a Democrat for that matter,

280
00:28:44,420 --> 00:28:48,440
it doesn't really matter because that state is going to be democratic no matter what.

281
00:28:48,440 --> 00:28:53,540
So you're not really going to you're going to waste your money if you're going to go and campaign in California.

282
00:28:53,540 --> 00:28:57,080
So in in the states and the UK as well,

283
00:28:57,080 --> 00:29:02,630
we have a first past the post constituency based system where you can weed out the people

284
00:29:02,630 --> 00:29:07,940
who don't really matter either people in the same states or the safe seats in America.

285
00:29:07,940 --> 00:29:15,550
This has the added wrinkle. A wrinkle of not every state is equal in the eyes of the founders.

286
00:29:15,550 --> 00:29:24,080
And so I took the liberty of calculating how many people you need per electoral vote in in the various states.

287
00:29:24,080 --> 00:29:33,650
And that looks like this. And it turns out that you need like, I think it's like a hundred and eighty thousand people per electoral vote in Wyoming,

288
00:29:33,650 --> 00:29:39,080
but you need to convince seven hundred forty four thousand Texans to get a single electoral vote.

289
00:29:39,080 --> 00:29:46,880
So that basically says some states are more important than others. So armed with this information, you can start doing analytic voter segmentation.

290
00:29:46,880 --> 00:29:51,980
So if I really want to figure out who am I going to talk to, what am I going to talk to them?

291
00:29:51,980 --> 00:29:59,990
And what am I going to say to them? You're going to throw all your data into this and model generation machine, which basically, as I said,

292
00:29:59,990 --> 00:30:06,070
just trying to draw correlations between what a person is like to believe in what a person is likely to find important.

293
00:30:06,070 --> 00:30:13,740
And this can be done by trawling through your digital data, trawling through your shopping lists, your Google searches,

294
00:30:13,740 --> 00:30:20,540
your activities on on apps like how many TikToks of little means you post today,

295
00:30:20,540 --> 00:30:24,950
for instance, and maybe inversely correlated with your likelihood of voting.

296
00:30:24,950 --> 00:30:31,880
I would imagine. And so you can start drawing up these segmentation lists.

297
00:30:31,880 --> 00:30:38,180
So let's say that you want to figure out like what is the subjective belief of any kind of person, according to the model that I've been drawing up?

298
00:30:38,180 --> 00:30:44,860
And bear in mind, you can have a really bad model in which case you're going to be really bad at predicting people.

299
00:30:44,860 --> 00:30:53,300
And this is why I'm interested in the principles underpinning micro-targeting rather than the specific case of, say, Cambridge Analytica,

300
00:30:53,300 --> 00:30:58,580
because it's really difficult to tell how good their models were, partly because it's a single case study,

301
00:30:58,580 --> 00:31:02,720
but also partly because they probably haven't published exactly how they did everything.

302
00:31:02,720 --> 00:31:06,830
So there's a lot of hearsay, there's a lot of sort of rumours and stuff going around.

303
00:31:06,830 --> 00:31:14,510
So fundamentally, what this kind of a perspective means is that you can start drawing up these Bayesian belief networks,

304
00:31:14,510 --> 00:31:18,350
which here I'm just going to take focus on a single subjective belief.

305
00:31:18,350 --> 00:31:25,320
So like, for instance, what's the likelihood that I'm going to vote for this guy or that guy or this candidate or that candidate?

306
00:31:25,320 --> 00:31:30,120
So in this case, I have segmented first to the left.

307
00:31:30,120 --> 00:31:38,620
Anyone who has a very strong preference for my candidate, so above point seven, according to where my model deems that person to be.

308
00:31:38,620 --> 00:31:45,420
Or in other words, that's my base. And like, those guys are pretty safe in my own modelling perspective.

309
00:31:45,420 --> 00:31:54,300
Conversely, to the far right, no pun intended and you've got people who are definitely not in my base,

310
00:31:54,300 --> 00:31:59,250
but who are definitely, definitely in the other person's base in the opposing candidate's base.

311
00:31:59,250 --> 00:32:04,890
They're not really that attractive either, because it's going to take a [INAUDIBLE] of a lot of effort to draw them to my side.

312
00:32:04,890 --> 00:32:09,480
But then you've got all the delicious people in the middle and all the swing voters,

313
00:32:09,480 --> 00:32:14,100
the people who haven't really decided, am I going to vote for a candidate, a candidate B?

314
00:32:14,100 --> 00:32:21,000
And then you can start winnowing away what your model thinks they they think about your candidate.

315
00:32:21,000 --> 00:32:26,970
Credibility wise. So for instance, on the left hand side, do they think positively of my candidate?

316
00:32:26,970 --> 00:32:32,460
Will they respond positively to reports from my candidate or negatively about my candidate?

317
00:32:32,460 --> 00:32:35,760
In which case you probably shouldn't contact them even though they're swing voters,

318
00:32:35,760 --> 00:32:43,380
because you might just annoy them, like Ted Cruz going door to door like an unlikeable guy?

319
00:32:43,380 --> 00:32:50,250
So in the middle, for instance, you can see that you can segment it into a more positive versus less positive,

320
00:32:50,250 --> 00:32:55,470
and in the base that is adversely position to what your candidate,

321
00:32:55,470 --> 00:33:03,540
you may want to segment it not as a 50 50 as in like, oh, just mildly positive, but like wildly positive, like above point seventy five.

322
00:33:03,540 --> 00:33:12,090
So this may be something where say you want to persuade someone that climate change is happening and you know that they strongly disbelieve it.

323
00:33:12,090 --> 00:33:16,470
But you know that they really respect you as a as a person or a scientist or as a source,

324
00:33:16,470 --> 00:33:23,070
then you may want to actually contact those guys, even though they are already pretty far on the other side.

325
00:33:23,070 --> 00:33:27,030
So, for instance, you can ask yourself to eye contact this person the base.

326
00:33:27,030 --> 00:33:29,660
Not so much because they're already in your pocket.

327
00:33:29,660 --> 00:33:35,640
And here you should probably forgo the people who hate your candidate candidate or who doesn't really like your candidate,

328
00:33:35,640 --> 00:33:40,680
and you should only contact the people who kind of like your candidate. Same thing here,

329
00:33:40,680 --> 00:33:49,050
where I have beefed up the proportion to which the degree to which they really should like your candidate before you should contact them.

330
00:33:49,050 --> 00:33:55,890
So now you can see, like we segment of the population into six bins, two of which are now going to be contacted.

331
00:33:55,890 --> 00:34:00,090
Um, do they have an influence on the election? Well, yes or no.

332
00:34:00,090 --> 00:34:07,570
So like, let's say that this guy in the middle who actually likes my candidate, but he lives in California.

333
00:34:07,570 --> 00:34:11,230
Who cares? And in that case, I'm going to leave him.

334
00:34:11,230 --> 00:34:18,400
Same thing, obviously, for the people who strongly like your candidate, but who was marginally towards the other person's base.

335
00:34:18,400 --> 00:34:22,840
And what's it like to have a voting fun? You can buy that in America.

336
00:34:22,840 --> 00:34:27,520
You can't buy what they voted because obviously that's secret.

337
00:34:27,520 --> 00:34:33,190
But there is records as to how many times you participated in elections that you were eligible to participate in.

338
00:34:33,190 --> 00:34:36,400
So I can get a pretty decent estimate of that.

339
00:34:36,400 --> 00:34:45,460
So let's say that anyone who is more than 75 percent likely to vote, I can segment my population into that as well.

340
00:34:45,460 --> 00:34:53,350
And anyone who is less than that. So now I think, OK, who should I try contacting to get out the vote for us?

341
00:34:53,350 --> 00:34:58,300
So in the last couple of days before the election, who should I really try and motivate to get out of the House to vote?

342
00:34:58,300 --> 00:35:05,500
Well, you definitely like reach out to my base now because they're the ones that I really want to make sure turns out on Election Day.

343
00:35:05,500 --> 00:35:14,620
So, for instance, in in 2008, and one of the strategies that Obama invoked was a pretty clever one.

344
00:35:14,620 --> 00:35:22,210
So there was some communities of African-American citizens in not Illinois that was a safe state.

345
00:35:22,210 --> 00:35:27,460
I think it was Iowa, one of the neighbouring states. And obviously, they've done the legwork.

346
00:35:27,460 --> 00:35:34,030
So they knew that if they get a thousand African-American people to turn out for them to vote for in the election,

347
00:35:34,030 --> 00:35:40,900
probably nine hundred and ninety seven of them are going to vote for Obama and like three of them, are going to vote for John McCain.

348
00:35:40,900 --> 00:35:46,540
And so what do they do? They didn't contact them in the persuasion phase, but they just drove in buses after buses,

349
00:35:46,540 --> 00:35:54,650
and they didn't bother bother persuading people or like and convincing people like, vote for Obama is just like, Do you register to vote?

350
00:35:54,650 --> 00:36:01,810
Great. That's driving to the election polls right now because they know from the modelling that most likely,

351
00:36:01,810 --> 00:36:05,740
if you just get a thousand people, that's going to be a really, really key thing.

352
00:36:05,740 --> 00:36:11,980
So that's a kind of get out the vote strategy that separates clearly from a persuasion strategy, a similar thing.

353
00:36:11,980 --> 00:36:17,830
You may want to focus on the people who are unlikely to vote and forgo the people who are very likely to vote.

354
00:36:17,830 --> 00:36:24,730
So if I know like, OK, I've already persuaded you to vote for my candidate and you are ninety three percent likely to vote,

355
00:36:24,730 --> 00:36:30,530
I don't really want to bother knocking on your door come Election Day because you'll probably be there anyways,

356
00:36:30,530 --> 00:36:35,860
because the one election you missed was because you were on a holiday or something on you.

357
00:36:35,860 --> 00:36:44,800
And then you can use the further subdivision of the psychometrics individual's psychological differences like, for instance, personality measures.

358
00:36:44,800 --> 00:36:49,240
So you can see how this creates this beautiful patchwork of further segmentation,

359
00:36:49,240 --> 00:36:53,920
further further segmentation where you can figure out what does this person believe in?

360
00:36:53,920 --> 00:36:58,330
What's the likelihood of credibility of that guy? And should I contact that guy for persuasion?

361
00:36:58,330 --> 00:37:03,350
Should I contact that guy for behaviour, like for the get out to vote? And what can you do with this?

362
00:37:03,350 --> 00:37:07,960
Let's say that you generate like 80 bins of people that you can then filter

363
00:37:07,960 --> 00:37:13,690
them into four personality four beliefs for priorities of political leanings,

364
00:37:13,690 --> 00:37:19,690
but you can start AB testing that so you can start running experiments on what is the most effective way of

365
00:37:19,690 --> 00:37:28,450
contacting that subdivision that steps up steps up segment of the population isn't using a particular type of word.

366
00:37:28,450 --> 00:37:31,360
Is it to use like a particular message type?

367
00:37:31,360 --> 00:37:40,930
So one example is like if you have three guys in America and they all kinds of Democratic Republican leaning,

368
00:37:40,930 --> 00:37:45,670
but maybe slightly on the fence and they all care deeply about guns.

369
00:37:45,670 --> 00:37:53,950
If you have someone who is strongly conscientious, you might contact them and say, like, Americans are law abiding people.

370
00:37:53,950 --> 00:37:59,050
We keep our country safe. We we know how to follow, like and adhere to rules.

371
00:37:59,050 --> 00:38:03,100
So guns are protected by the Second Amendment. [INAUDIBLE] obviously stand.

372
00:38:03,100 --> 00:38:07,420
I think that's a very simplistic so like civil kind of an argument.

373
00:38:07,420 --> 00:38:15,520
Then you may have a pro-gun message for extroverted people, which is more like guns, a part of American way of life.

374
00:38:15,520 --> 00:38:23,860
We hunt together, we barbecue together with or like you or Becks or hunter like hunters of frontier kind of people.

375
00:38:23,860 --> 00:38:30,550
And that might work really well for them in the AB testing that you continuously trying to refine to make sure that the way that you are contacting

376
00:38:30,550 --> 00:38:41,620
that subset of that voter will be as as sort of spot on for their proclivities psychologically and belief wise as you as you possibly can.

377
00:38:41,620 --> 00:38:49,830
Finally, like if you have a deeply neurotic person, you can go like, oh, like the Democrats coming for your guns, which is obviously flare up.

378
00:38:49,830 --> 00:38:56,230
And so imagine that you want to sue in state here, Columbus, Ohio,

379
00:38:56,230 --> 00:39:07,650
and you go on a street level so you can start drawing maps of this because obviously this personal data is linked to your personal profile on a live.

380
00:39:07,650 --> 00:39:13,680
Because those registries of that. So what I would do, that's an analytic microtargeting and targeting campaign person.

381
00:39:13,680 --> 00:39:21,600
I would then start to figure out like, OK, there's a bunch of houses who are relevant and irrelevant because they hate my candidate or

382
00:39:21,600 --> 00:39:26,820
they or they they really hate my issue and they hate my candidate and they're unlikely to vote.

383
00:39:26,820 --> 00:39:34,590
And like, forget about these guys. Then there's a bunch of people where you may want to think about like, Oh, these guys on the fence.

384
00:39:34,590 --> 00:39:39,300
So for instance, you've got the Smith household where my model says, like all probability, you're voting like pretty high.

385
00:39:39,300 --> 00:39:46,530
Ninety three percent, my data is kind of distributed and as illustrated by this graph,

386
00:39:46,530 --> 00:39:49,620
like you have like 200 data points in that person and they they're not like really

387
00:39:49,620 --> 00:39:54,360
clear on whether or not this person is a Democrat or a Republican from Twitter.

388
00:39:54,360 --> 00:39:59,010
I know that you talk a lot about economy, environment and women's rights and some psychometrics,

389
00:39:59,010 --> 00:40:03,480
in which case I'm going to try and persuade you with something that is specifically designed

390
00:40:03,480 --> 00:40:10,080
for that sub segment of the population as driven by the data analysis that I've been doing.

391
00:40:10,080 --> 00:40:13,860
Similarly, when it comes to get out the vote, you get these guys.

392
00:40:13,860 --> 00:40:20,640
So like the Taliad House, where the probability of voting is pretty low, but all for all intents purposes,

393
00:40:20,640 --> 00:40:25,320
the data analysis that you've been doing suggests that the person is strongly in your camp.

394
00:40:25,320 --> 00:40:28,740
So you know that this is someone who probably liked your candidate but is unlikely to vote,

395
00:40:28,740 --> 00:40:36,010
maybe because of despite discouragement of disinterest in the political system or something of that in which case that person gets it,

396
00:40:36,010 --> 00:40:41,190
gets a get out the vote message so you can start building up this.

397
00:40:41,190 --> 00:40:49,080
So a fascinating picture of a whole of a nation. You can zoom in on street level and figure out exactly how you're going to approach that person

398
00:40:49,080 --> 00:40:54,720
who you're not going to approach and who you're going to approach for the get out the vote part.

399
00:40:54,720 --> 00:41:04,260
So this is kind of reportedly what Cambridge Analytica did, they sort of mined deeper and deeper and deeper, which is great.

400
00:41:04,260 --> 00:41:12,060
It's a very strong perspective on basically marketing purposes like it's not anything more fancy than sort of swanky.

401
00:41:12,060 --> 00:41:19,260
Marketing is basically just like, what is this person like? And then throwing a dash of psychological profiling.

402
00:41:19,260 --> 00:41:27,540
Again, I will say this doesn't detract from detract from the idea that you can get increasingly sophisticated models as we go along.

403
00:41:27,540 --> 00:41:35,450
But as we all know, human beings are not stuck in a well. Rather, we are interconnected, so we constantly talk with each other.

404
00:41:35,450 --> 00:41:42,960
And what's the main danger of increasingly just digging down into a well, metaphorically speaking,

405
00:41:42,960 --> 00:41:52,800
is imagine like for the completely soul of imagine the example that you would have a presidential candidate who had, let's say, a racist or a segment.

406
00:41:52,800 --> 00:41:55,620
I know it's a completely made up example, obviously, but like,

407
00:41:55,620 --> 00:42:03,330
go with me for for for the for the purpose of illustration and then you want to figure out how am I really

408
00:42:03,330 --> 00:42:08,160
going to motivate this guy like these guys to come out and vote like I really need to motivate my base?

409
00:42:08,160 --> 00:42:15,570
Well, that may be to send an incredibly like you can sort of refine and refine and refine and optimise and optimise and optimise and figure out like,

410
00:42:15,570 --> 00:42:23,280
Oh, I just need to send them a supremely bigoted and mixed message about Mexicans or something that that's going to get them out of the door.

411
00:42:23,280 --> 00:42:29,580
But what happens then in this imagine example, if one of these guys goes on Twitter and goes like,

412
00:42:29,580 --> 00:42:36,450
Oh, the president is doing something really great here? Hashtag get on my country kind of a thing.

413
00:42:36,450 --> 00:42:43,500
And it goes viral. Then everyone in the middle of the political spectrum who might find that to be abhorrent are now pushed away from the candidate,

414
00:42:43,500 --> 00:42:47,490
and it works counterproductive. And so the deeper you go,

415
00:42:47,490 --> 00:42:54,030
the larger you increase your probability of hitting something that might not be palatable for someone who's not as strong you within your base.

416
00:42:54,030 --> 00:42:58,440
And the problem in this modern day of age is that we can talk to each other via Twitter,

417
00:42:58,440 --> 00:43:02,820
via Facebook, via Messenger and stuff that so we have agency as citizens.

418
00:43:02,820 --> 00:43:08,340
So what? I want to end the loss of like five 10 minutes of the talk about is how is it perspective?

419
00:43:08,340 --> 00:43:11,790
I see this as a dynamic thing rather than as analytic thing.

420
00:43:11,790 --> 00:43:22,590
So information flow on social networks, and I've been doing some work on this with Typekit Hilditch and Richard Bailey and me or may not be here,

421
00:43:22,590 --> 00:43:27,630
but I've been doing work on this on echo chamber formation, on micro-targeted and campaigns.

422
00:43:27,630 --> 00:43:32,730
And the idea here is that you can simulate individual citizens and their connectivity

423
00:43:32,730 --> 00:43:38,460
and see how things play out once you release your campaign into that environment.

424
00:43:38,460 --> 00:43:46,110
So they take this takes into account the agency of citizens to share information with each other as well as receive information,

425
00:43:46,110 --> 00:43:53,130
which means that there's less like there's more of a give and take. It also takes into account people's social network position.

426
00:43:53,130 --> 00:44:03,360
So who is really important here? So one excruciatingly important person in the recent midterm election was today.

427
00:44:03,360 --> 00:44:10,800
Hey, hey, hey. And Taylor Swift, for people who don't know her, is an American singer songwriter.

428
00:44:10,800 --> 00:44:18,600
And she tweeted out, Remember to register to vote? And within twenty four hours, forty nine thousand people had registered to vote.

429
00:44:18,600 --> 00:44:24,300
That's a pretty good return, but she lives in California.

430
00:44:24,300 --> 00:44:32,400
Any kind of analytical marketing and micro-targeting segmentation model worth their salt would have disregarded her any day of the week.

431
00:44:32,400 --> 00:44:40,620
But there's a huge power to be had from understanding the perspective of what is a social network and how information flows,

432
00:44:40,620 --> 00:44:46,950
how it has influence so levels on information networks.

433
00:44:46,950 --> 00:44:51,180
So one way of modelling this is what is known as agent based modelling.

434
00:44:51,180 --> 00:44:53,010
I'm not going to bore you with the details,

435
00:44:53,010 --> 00:44:59,400
but suffice to say that you simulate a bunch of individual agents and you can make them as heterogeneous as you want.

436
00:44:59,400 --> 00:45:06,480
And so you can populate them with whatever kind of models that your analytic, micro-targeted campaign says that they have.

437
00:45:06,480 --> 00:45:09,390
Then you contact them with each other and you set them into social networks.

438
00:45:09,390 --> 00:45:13,920
So you have interactions and you let it loose and you see, how are they going to share information?

439
00:45:13,920 --> 00:45:17,940
How are they going to act? And once the election campaign is over?

440
00:45:17,940 --> 00:45:25,050
So an example of this is we want to see if we can simulate echo chamber formation and we could.

441
00:45:25,050 --> 00:45:32,870
I'm going to skip slightly over this. But the main points taken away from this paper is that even rational agents can become stuck in echo chambers.

442
00:45:32,870 --> 00:45:38,310
And when we increase connectivity in the social network, the problem becomes worse, not better.

443
00:45:38,310 --> 00:45:46,020
So suck on that sucker back, and education broadcasts get less effective in dispelling HQ chambers with time.

444
00:45:46,020 --> 00:45:49,530
And so this is just an example where you can show that you can use these dynamic

445
00:45:49,530 --> 00:45:55,420
models to test some really fundamental problems of democracy and information.

446
00:45:55,420 --> 00:45:59,420
So one way of conceptualising an election instead of conceptualising it from

447
00:45:59,420 --> 00:46:03,920
this tree branch that goes down is to conceptualise it as a dynamic model.

448
00:46:03,920 --> 00:46:08,570
We have some voters who talk to each other. You've got some news pundits trying to influence them.

449
00:46:08,570 --> 00:46:11,220
You've got some other campaigns trying to influence them as well.

450
00:46:11,220 --> 00:46:16,550
It's all groups and they are also responsive to what goes on within that social network.

451
00:46:16,550 --> 00:46:23,720
So let's say that there's a shift in in the perspectives of the citizens that may cause the other campaigns to shift their strategies,

452
00:46:23,720 --> 00:46:27,590
which may mean that you have to switch your trustees to a conference sort give and

453
00:46:27,590 --> 00:46:33,320
take and a constant interaction between like those peoples and campaign strategies.

454
00:46:33,320 --> 00:46:43,340
So and this also points towards a completely different perspective where instead of constantly pouring down into even deeper wells deeper and deeper,

455
00:46:43,340 --> 00:46:47,360
like stronger and stronger isolation between the bins of citizens,

456
00:46:47,360 --> 00:46:52,820
you conceptualise the heterogeneity and you populated and stimulated within a model.

457
00:46:52,820 --> 00:47:00,950
You give them some possibilities for interacting with each other. And you throw them like some campaign strategies and you see, how do they react?

458
00:47:00,950 --> 00:47:08,000
How do they respond to this? Like, if I go out and send that like bigoted message, how does the electorate so form around that?

459
00:47:08,000 --> 00:47:11,090
How what's this off vibe on Twitter?

460
00:47:11,090 --> 00:47:18,710
There's no way you can do this as an athletic model, but there's every which way that you can do this mathematically through simulated dynamic models.

461
00:47:18,710 --> 00:47:25,460
Further, if you really want to use sure about it, you can wrap this kind of a model in what is known as an optimisation,

462
00:47:25,460 --> 00:47:32,600
so you can find the optimal way of managing that complex human human environment system.

463
00:47:32,600 --> 00:47:41,120
This is something that we've been doing in fisheries, where we are trying to figure out what is the best way of managing these dynamic systems.

464
00:47:41,120 --> 00:47:48,650
So if we have some closing reflections on what do we need to think about when we think about microtargeting?

465
00:47:48,650 --> 00:47:54,230
And first of all, we need to understand that the main modify as to whether or not these models are

466
00:47:54,230 --> 00:47:59,030
effective is access to data skewed resources and access to experts and modelling.

467
00:47:59,030 --> 00:48:05,150
Because if you have really bad models with really bad data and and almost no money to run it,

468
00:48:05,150 --> 00:48:09,740
like good luck in your micro-targeted campaign, it's got to be shite.

469
00:48:09,740 --> 00:48:17,330
But if you have really good data like extraordinarily good data, we can get modellers and all the time in the world to segment that electorate,

470
00:48:17,330 --> 00:48:22,670
including their social positions and the dynamics going beyond pretty good standing

471
00:48:22,670 --> 00:48:29,840
as data becomes more abundant and psychological models become more precise. We may expect candidates to increasing use of this.

472
00:48:29,840 --> 00:48:33,800
And so what can we do? One thing maybe to modify campaigns access to data.

473
00:48:33,800 --> 00:48:41,240
One proposal I have is make any data available to any campaign publicly available for any other campaign.

474
00:48:41,240 --> 00:48:47,030
Because then if you collaborate with, say, external lobbying groups who give you data,

475
00:48:47,030 --> 00:48:50,330
then everyone gets that data and they can all do the same modelling of that.

476
00:48:50,330 --> 00:49:01,490
At least that would limit the degree to which you can buy sensible, like sensitive data and get an advantage in a strategic edge from doing that.

477
00:49:01,490 --> 00:49:06,050
Also, obviously, we can enforce limits on campaign spending. In the UK, we have limits.

478
00:49:06,050 --> 00:49:13,190
Thank God. But they don't in America. But this should also include any collaborators because otherwise I would just go like, Oh,

479
00:49:13,190 --> 00:49:19,880
my buddy over here has a billion dollars, but it's not me, like, it's my superPAC back.

480
00:49:19,880 --> 00:49:26,080
So this would also limit obviously access to development expertise and forced people to think about their money.

481
00:49:26,080 --> 00:49:32,440
And also when considering all these negative aspects that are often been chucked out when we're discussing these microtargeting campaigns,

482
00:49:32,440 --> 00:49:36,020
while also think about what we can do with these methods proactively.

483
00:49:36,020 --> 00:49:40,700
So, for instance, optimise communication of scientific findings to people who don't believe in climate change,

484
00:49:40,700 --> 00:49:44,120
and we can generate better campaigns for public health and environmental issues

485
00:49:44,120 --> 00:49:50,210
like recycling because that underpins or relies on a fundamental understanding,

486
00:49:50,210 --> 00:49:54,590
both on the psychology of the persons who are the targets of those campaigns,

487
00:49:54,590 --> 00:50:00,600
but also the social fabric that binds them and what causes that behaviour.

488
00:50:00,600 --> 00:50:07,620
And finally, we can use these models to measure the vulnerability of information systems and misinformation.

489
00:50:07,620 --> 00:50:13,740
So to conclude based him can enable modellers to model subjective beliefs.

490
00:50:13,740 --> 00:50:17,520
We've been doing this with prior beliefs, with perception of causal relationships,

491
00:50:17,520 --> 00:50:21,570
with perceived credibility and how they see the dependency of various sources.

492
00:50:21,570 --> 00:50:30,390
And we're getting increasingly good at predicting what people will do once they are exposed to particular parts of particular types of information.

493
00:50:30,390 --> 00:50:34,440
Second, we can use personal data to populate so like Bayesian models.

494
00:50:34,440 --> 00:50:40,050
This includes tomography, digital traces, psychometrics influence, an election and anything else you can imagine.

495
00:50:40,050 --> 00:50:45,240
This kind of a list will grow as the models become more sophisticated.

496
00:50:45,240 --> 00:50:49,260
Campaigns obviously need to understand beliefs as well as behaviour changes because

497
00:50:49,260 --> 00:50:53,070
you can't just assume behaviour following deterministic form of her beliefs.

498
00:50:53,070 --> 00:50:55,980
So these are two very separate things and also social things.

499
00:50:55,980 --> 00:51:03,570
I should say analytic models fundamentally can't account for pushback from citizens agency.

500
00:51:03,570 --> 00:51:08,520
And so precision can backfire if you increasingly borrow your way down,

501
00:51:08,520 --> 00:51:15,390
which is what Cambridge Analytica reportedly did and believe information and behaviour on social networks are complex and non-linear.

502
00:51:15,390 --> 00:51:19,590
Is this really crucial and the heterogeneous, dynamic adaptive?

503
00:51:19,590 --> 00:51:24,720
But we can use these agent based models that we've been using to capture this.

504
00:51:24,720 --> 00:51:33,570
And this is why pre-emptively said before that point, which is basically that we can optimise over the management of these kind of models.

505
00:51:33,570 --> 00:51:39,540
And finally, it's worth remembering is always the amount of data that can save a bad candidate,

506
00:51:39,540 --> 00:51:46,800
Jeb, because he had the access to the most data in 2016's primary election.

507
00:51:46,800 --> 00:52:02,240
And look where that got him. So thanks very much for listening, and I will take some Q&A.

508
00:52:02,240 --> 00:52:11,820
And so if you raise your hand, if you have a question or just if no one has questions, why oh yes.

509
00:52:11,820 --> 00:52:20,440
I should say as well, I was asked to say, wait until you have a microphone before you ask the question, because otherwise they won't be recorded.

510
00:52:20,440 --> 00:52:26,080
You talk about dynamic and analytic ways of targeting campaigns.

511
00:52:26,080 --> 00:52:30,670
How much do you think the analytic campaigns are being used and if they're not being used very much at the minute?

512
00:52:30,670 --> 00:52:40,780
How much of a threat does this put on, like the um, like how moral campaigns are in the future?

513
00:52:40,780 --> 00:52:46,660
So how much are analytical models being used right now? Yeah. Oh, I would.

514
00:52:46,660 --> 00:52:51,130
I would hope that every kind of every political campaign uses analytic modelling,

515
00:52:51,130 --> 00:52:57,640
because if you think about analytic modelling is kind of just the degree to which you can segment and understand the electorate.

516
00:52:57,640 --> 00:53:02,230
And that could just be like James Carville have the most simple analytic model.

517
00:53:02,230 --> 00:53:07,360
Is the economy stupid? Like, that's that's an analytic model of just like a one issue thing.

518
00:53:07,360 --> 00:53:10,630
You're going to vote on that. But then you can splice that into two.

519
00:53:10,630 --> 00:53:17,710
So any kind of any kind of data driven segmentation or ideologically driven segmentation is kind of like model.

520
00:53:17,710 --> 00:53:27,910
Now how mobile it is, I think, depends entirely on obviously the campaign in the sense of like if the analytic campaign that's being done is to say,

521
00:53:27,910 --> 00:53:35,020
decrease criminality or increase recycling, I think that's a pretty good thing.

522
00:53:35,020 --> 00:53:41,200
But if it's to like, sell more coal, I might not personally think of that as a small thing.

523
00:53:41,200 --> 00:53:46,450
But I'm I'm kind of keen that this should be disassociated in principle from those

524
00:53:46,450 --> 00:53:51,700
in the sense that like this is a tool through which you can sort of like segment

525
00:53:51,700 --> 00:53:57,160
and optimise persuasion campaigns both analytically and also simplistic way or

526
00:53:57,160 --> 00:54:03,220
dynamically in a more interactive way and in whichever purpose that you use it.

527
00:54:03,220 --> 00:54:09,910
But it can be good or bad. It's a bit like studying medicine in a way like who better to poison you than a doctor?

528
00:54:09,910 --> 00:54:17,860
Like, we know, like they really know. But who better to say to you as well?

529
00:54:17,860 --> 00:54:28,120
Yeah. To down. And then some APC things genes for great and scary talk about the macro data management.

530
00:54:28,120 --> 00:54:35,260
My question is, what experience do you have with that dynamic micro targeting models like the AVMs that you

531
00:54:35,260 --> 00:54:41,860
have used with data because the one you mentioned was about more empirical and stylised facts,

532
00:54:41,860 --> 00:54:46,830
I think. Yeah. Have you used any of the dynamical models?

533
00:54:46,830 --> 00:54:51,270
With actually with training them on actual data. Not yet.

534
00:54:51,270 --> 00:54:57,180
And but. And spoiler, I am a in the process of doing so.

535
00:54:57,180 --> 00:55:01,050
And I have just submitted a grant application, which would exactly do that.

536
00:55:01,050 --> 00:55:06,540
So take stuff like digital traces and calibrate and validate dynamic models to figure out how, for instance,

537
00:55:06,540 --> 00:55:14,460
the networked self prunes as in like, who do I follow and who do I unfollow as a product of like whatever they say on Twitter or something?

538
00:55:14,460 --> 00:55:20,280
And one more time then then I think maybe some up here. Thank you for the talk.

539
00:55:20,280 --> 00:55:25,320
I appreciate that you weren't trying to sort of get to right or wrong on it.

540
00:55:25,320 --> 00:55:36,030
But if I could draw you slightly on one of your suggestions, which was that we might open up whatever data one campaign has to every campaign.

541
00:55:36,030 --> 00:55:45,480
Do you not think that that's perhaps overreach in terms of what an electoral commission or whatever sort of body is actually doing?

542
00:55:45,480 --> 00:55:53,010
For example, one party or campaign might have newspapers on their side, and we wouldn't enforce that.

543
00:55:53,010 --> 00:56:02,070
A newspaper gives equal coverage to all campaigns by enforcing sort of the data that each campaign can have access to.

544
00:56:02,070 --> 00:56:07,170
Are you sort of artificially levelling the playing field? So like, I mean,

545
00:56:07,170 --> 00:56:10,440
those are two separate things as in like the degree to which a newspaper is on the

546
00:56:10,440 --> 00:56:14,370
side of the candidate or party is more about like the general support from say,

547
00:56:14,370 --> 00:56:19,170
like the Telegraph, the Tories or the Guardian for Labour.

548
00:56:19,170 --> 00:56:24,090
But what I'm talking about is more about if I get as a campaign,

549
00:56:24,090 --> 00:56:34,290
if I get socioeconomic data because I've got like a bucket load of money and I can disproportionately unfairly sharpen my models to the extent that,

550
00:56:34,290 --> 00:56:41,880
like in this old marketplace of ideas, I get all silly things else being equal, an advantage over my slightly less rich candidate.

551
00:56:41,880 --> 00:56:47,070
And the proposal would be to force that campaign, to share that data,

552
00:56:47,070 --> 00:56:53,730
to limit the incentive to buy incredibly precise data that costs umpteen millions.

553
00:56:53,730 --> 00:56:57,600
And so it is an artificial leveller of the playing field.

554
00:56:57,600 --> 00:57:03,030
But that's kind of the point is to ensure that there's more of a level playing field for getting your ideas out there.

555
00:57:03,030 --> 00:57:09,840
And so I think we should take some questions up here. The two of you sorry, I understood what you were saying about I'm sorry.

556
00:57:09,840 --> 00:57:14,710
I understood what you were saying about the economics of campaigning.

557
00:57:14,710 --> 00:57:19,260
And you you had these various people with various ideas. But how did you get them in there?

558
00:57:19,260 --> 00:57:25,870
How did you know? How did you know how to address me? Or how did you know how to address the kind of audience that is here?

559
00:57:25,870 --> 00:57:32,500
You were you were very, very specific about what you do with the information when you've got it.

560
00:57:32,500 --> 00:57:37,200
But less clear about how this information arrives in forms which could be manipulated.

561
00:57:37,200 --> 00:57:41,760
Yeah. What did I say to my friends last night that would change your thinking and so on?

562
00:57:41,760 --> 00:57:48,870
So. And so I, for instance, and the things I mentioned about your demographic data.

563
00:57:48,870 --> 00:58:02,130
So if I can get access to your your shopping habits or your Googling habits, and if I have how the dosh money like a lot of this is for sale.

564
00:58:02,130 --> 00:58:09,220
So in particular, in America, a lot of these personal personal data is for sale.

565
00:58:09,220 --> 00:58:19,380
So I don't know if everyone does it, but like some companies where they have loyalty cards, when you shop, that's pretty valuable data.

566
00:58:19,380 --> 00:58:25,140
So yeah, that's one way. Then a second method of doing it is scraping.

567
00:58:25,140 --> 00:58:34,200
So if I go into someone's Twitter profile, if it's public and I just scrape all the tweets that they've ever done and I run it through a

568
00:58:34,200 --> 00:58:39,180
machine optimisation algorithm that picks out whether or not they are more environmental words,

569
00:58:39,180 --> 00:58:44,490
also a Democratic virtue Republican birds, I can then again get a version of that person.

570
00:58:44,490 --> 00:58:48,600
So let's say that I got your Twitter data, I've got your shopping data, I've got your Google data,

571
00:58:48,600 --> 00:58:53,400
I've got a whole host of different types of data sets, and I merged those two.

572
00:58:53,400 --> 00:58:58,230
So I'll try and figure out what are you roughly like where you roughly in terms of your political position?

573
00:58:58,230 --> 00:59:03,560
That's how they that's how they can be used for that. Thanks for.

574
00:59:03,560 --> 00:59:13,220
And thank you for your talk. So you mentioned at one point that Google searches could essentially be bought.

575
00:59:13,220 --> 00:59:20,570
My impression is that I don't know if they can. But in principle, and that can be themed.

576
00:59:20,570 --> 00:59:26,750
I don't know if they actually said it. I can say my impression is that they sell very broad categories.

577
00:59:26,750 --> 00:59:35,030
Yeah, probably. And so if. Someone told me recently that it was as few as 12 categories,

578
00:59:35,030 --> 00:59:40,190
but if you could help us understand a little bit more about that, I think it would be helpful.

579
00:59:40,190 --> 00:59:49,760
I don't know how much data Google sells, I'm afraid. But again, for me, that's a case specific thing because like I now say,

580
00:59:49,760 --> 00:59:56,720
the sucker may like for Facebook may hold off from selling particular data points.

581
00:59:56,720 --> 01:00:04,940
That is in no way a conceptual guarantee that the next big company because Facebook is dying of easy.

582
01:00:04,940 --> 01:00:09,470
But the next big company is social media. There's no guarantee that they would have the same principles.

583
01:00:09,470 --> 01:00:13,790
So like Facebook, post that API once they realised it was there.

584
01:00:13,790 --> 01:00:21,080
But that in in in principle doesn't safeguard us in any kind of way, which is my point exactly,

585
01:00:21,080 --> 01:00:30,890
is that we have to understand what is the data that is out there about each person, how is it regulated and how is it potentially sold?

586
01:00:30,890 --> 01:00:40,760
And if we don't understand what that fundamental marketplace of data acquisition is, we are just like struggling.

587
01:00:40,760 --> 01:00:44,900
We're blind in that way in terms of regulating this kind of stuff.

588
01:00:44,900 --> 01:00:50,240
And I should say I promised at six o'clock to say that there was a wine reception.

589
01:00:50,240 --> 01:00:53,750
And so I'm going to take two more questions. Sorry, it's three past six.

590
01:00:53,750 --> 01:01:01,220
I would love to take more questions, but yes, you and then the lady in the front.

591
01:01:01,220 --> 01:01:08,090
Just quick, quick, quick question at the you have mentioned that you can use this model for positive reasons like we go coverage.

592
01:01:08,090 --> 01:01:16,730
Do you know any cases that are actually use except for election campaign and the question that is U.K.,

593
01:01:16,730 --> 01:01:21,470
you mention that in the US, the more data is available. Yeah.

594
01:01:21,470 --> 01:01:25,670
Where does the UK stand in terms of election? How susceptible?

595
01:01:25,670 --> 01:01:36,650
And we until tomorrow we're standing pretty well because there's EU regulations on GDPR, which limits the transaction of personal data.

596
01:01:36,650 --> 01:01:44,450
And yeah, so ask me again on Monday, it would be my mind depressing answer.

597
01:01:44,450 --> 01:01:50,960
So right now, like in Europe, there's a lot more protection. So you can probably do what they do in America for two reasons.

598
01:01:50,960 --> 01:01:56,150
One, because the access to data is more restricted and more protected,

599
01:01:56,150 --> 01:02:02,300
but also be because literally the campaign finances are restricted much more severely.

600
01:02:02,300 --> 01:02:08,660
And I can't remember exactly how much each campaign is allowed to spend, but it's in the tens of millions,

601
01:02:08,660 --> 01:02:12,710
like maybe 40 million or something for a political campaign and that's across the board.

602
01:02:12,710 --> 01:02:20,180
You cannot spend anymore. That's it. Comparatively, oh, is it $5.6 billion dollars spent on the last midterm election?

603
01:02:20,180 --> 01:02:26,810
Like Pfizer as a magnitude, it's not just like the size of the country, and that's just an open flood.

604
01:02:26,810 --> 01:02:34,730
It's not the case for the size of this. Oh, sorry that any other cases that are used this kind of money.

605
01:02:34,730 --> 01:02:42,680
Oh, and I mean, again, in the same way that I responded to that woman down there, I think every model,

606
01:02:42,680 --> 01:02:46,670
every campaign must have some kind of an idea of the kind of audience they're talking to.

607
01:02:46,670 --> 01:02:54,230
But I would doubt that many campaigns that are publicly funded would have this amount of of sophistication.

608
01:02:54,230 --> 01:02:59,840
I doubt it. But again, I haven't run in, so I can't say.

609
01:02:59,840 --> 01:03:06,110
I've been wondering whether devices like Alexa play a significant role in information gathering,

610
01:03:06,110 --> 01:03:12,650
and they do so like Alexa, for instance, and trained like.

611
01:03:12,650 --> 01:03:21,380
Like, I mean, the specific again, I hasten to not make too big a point of specific things, like specific things.

612
01:03:21,380 --> 01:03:31,310
But like an Alexa, you have to go into a subset of your settings to allow it to not use your vocal data to train it on itself.

613
01:03:31,310 --> 01:03:39,620
As I was constantly listening, but it's for the purpose of machine learning, optimisation of picking out exactly your voice.

614
01:03:39,620 --> 01:03:44,360
And so again, I don't think Amazon is selling that data,

615
01:03:44,360 --> 01:03:48,170
but there's nothing in principle that would stop them unless we have strong

616
01:03:48,170 --> 01:03:53,840
regulation and strong top down management from a from a societal perspective,

617
01:03:53,840 --> 01:04:02,630
which is exactly why I think this is a key thing. Like I hasten to say, I don't want to be doom and gloom because like I am,

618
01:04:02,630 --> 01:04:07,490
micro-targeting campaigns and dynamic microtargeting campaigns are great in getting your message out.

619
01:04:07,490 --> 01:04:12,290
And if you have a good message, that's fine. But the uneven playing field,

620
01:04:12,290 --> 01:04:20,330
the impact of money on elections and the skewed access to data gives a disproportionately unfair advantage to people who are moneyed,

621
01:04:20,330 --> 01:04:22,070
which is a democratic problem.

622
01:04:22,070 --> 01:04:29,840
So as a society, we shouldn't think like, Oh, these data hungry machines can do everything like they can't predict what shoes I'm going to wear today.

623
01:04:29,840 --> 01:04:36,290
Maybe, maybe they can, but probably not. And so they can't predict everything, but they also can't.

624
01:04:36,290 --> 01:04:40,460
They're not like nothing, so they can predict something for sure.

625
01:04:40,460 --> 01:04:45,650
And the degree to which those models are sophisticated is just increasing what they can predict.

626
01:04:45,650 --> 01:04:50,840
And this is why I think this is a key area of research.

627
01:04:50,840 --> 01:05:00,590
But yeah, I think, um. Hello. And that was a great conclusion and a really informative but unsettling talk.

628
01:05:00,590 --> 01:05:07,640
And my prior my prior belief about Danish football teams has changed so much.

629
01:05:07,640 --> 01:05:14,630
So thank you all very much for coming. There is the wine reception. I'd also like to draw everyone's attention to a talk coming up next Monday,

630
01:05:14,630 --> 01:05:21,780
28th on the geographies of the platform economy with Professor Mark Harrison, which proves to be a very interesting one.

631
01:05:21,780 --> 01:05:42,214
So once again, thank heavens. And that's my.