1
00:00:02,430 --> 00:00:08,730
So I'm they are not I didn't have work as a postdoctoral research fellow researcher at the Oxford Internet Institute.

2
00:00:08,730 --> 00:00:16,620
And so I work in the Computational Propaganda Project, which is headed by Professor Phil Howard.

3
00:00:16,620 --> 00:00:25,330
So we're a group of political scientists, media scholars and computer scientists say training.

4
00:00:25,330 --> 00:00:36,990
So it's a highly interdisciplinary project and we track the spread of fake news, misinformation, disinformation on social media.

5
00:00:36,990 --> 00:00:48,600
So I think I would like to describe my work as mapping the impact of A.I. on our democratic processes.

6
00:00:48,600 --> 00:00:55,590
So computational propaganda is within the project. We have a working definition of computational propaganda.

7
00:00:55,590 --> 00:01:03,600
So this is about how algorithms and the affordances of social media platforms come

8
00:01:03,600 --> 00:01:11,340
together to enable malicious actors to spread fake news and to influence public opinion,

9
00:01:11,340 --> 00:01:19,020
manipulate elections. So you would notice that I've put down jump use here instead of fake news,

10
00:01:19,020 --> 00:01:25,320
and this was a deliberate attempt to move away from the term fake news because it was

11
00:01:25,320 --> 00:01:31,710
being weaponized and being used by authoritarian figures to to attack publications,

12
00:01:31,710 --> 00:01:36,450
mainstream media that were critical of government policy and so on.

13
00:01:36,450 --> 00:01:43,170
So when when we first started this project, we were looking, we were looking after Twitter,

14
00:01:43,170 --> 00:01:48,240
so we would track politically relevant hashtags on Twitter as we would study elections around the world.

15
00:01:48,240 --> 00:01:57,660
We started with, I think, us 2016, the presidential elections that, you know some of you might have heard about.

16
00:01:57,660 --> 00:02:05,340
So, you know, under Brexit. So Brexit before Brexit and 2016 elections.

17
00:02:05,340 --> 00:02:12,540
So we would track politically relevant hashtags on Twitter extract to extract tweets that were

18
00:02:12,540 --> 00:02:19,080
posted with these with these hashtags and then extract links to news sources from these tweets.

19
00:02:19,080 --> 00:02:23,490
And then we had a classification classification scheme.

20
00:02:23,490 --> 00:02:29,370
And we would try to separate our professional news sources from more problematic use.

21
00:02:29,370 --> 00:02:34,580
So, I mean, there was nothing that, you know, we were not using advanced air.

22
00:02:34,580 --> 00:02:45,870
This was all done. It was just we use it mainly to track the spread of problematic new sources of junk users on social media platforms.

23
00:02:45,870 --> 00:02:52,990
But but just so you know, we would classify them into professional news sources and junk news sources.

24
00:02:52,990 --> 00:02:57,270
Then we would track the spread of job sources on social media platforms,

25
00:02:57,270 --> 00:03:03,960
and we had a detailed we had five criteria to determine if a news source was in fact problematic.

26
00:03:03,960 --> 00:03:09,630
So this would be, you know, professionalism if adapted to professional standards.

27
00:03:09,630 --> 00:03:14,800
If there were open about editorial policy, who their funders were science.

28
00:03:14,800 --> 00:03:21,090
So this could include, you know, ad hominem attacks, excessive use of capitalisation,

29
00:03:21,090 --> 00:03:25,830
the type of stuff, you know, the type of sensationalism and so on.

30
00:03:25,830 --> 00:03:31,320
And then you have credibility and then bias that could be hyperpartisan.

31
00:03:31,320 --> 00:03:39,120
You could, you know, you have these screaming news headlines purporting to be news, but it's actually commentary and so on.

32
00:03:39,120 --> 00:03:47,820
And then to come to feed new sources, they would mimic the branding and other features more well established news organisations.

33
00:03:47,820 --> 00:03:56,250
So if they fail, if a new source fail, three of these five criteria that we would classify them as junk news sources and then not and

34
00:03:56,250 --> 00:04:03,390
then find out in our sample what percentage of news sources were junk professional and so on.

35
00:04:03,390 --> 00:04:07,710
So this served us quite well when we were studying Western democracies.

36
00:04:07,710 --> 00:04:19,830
But when we move to Latin America in 2017, you know, 2000 18 and then ended early 2019.

37
00:04:19,830 --> 00:04:21,990
Things change fairly dramatically.

38
00:04:21,990 --> 00:04:30,990
First of all, I think, you know, it coincided with the rise of encrypted platforms like in messaging apps like WhatsApp.

39
00:04:30,990 --> 00:04:39,690
So people. So I think, you know, that is it speaks to different cultures of internet use and so on.

40
00:04:39,690 --> 00:04:48,730
So platforms like WhatsApp afforded a level of privacy that this absent and not absent on Twitter, Facebook and so on.

41
00:04:48,730 --> 00:04:56,310
We saw that we saw these apps grow in popularity, both in Brazil, particularly in Brazil and in India,

42
00:04:56,310 --> 00:05:01,820
and these were being used by political actors to directly to bypass traditional filters direct.

43
00:05:01,820 --> 00:05:11,600
You reach water populations. And with the rise of these messaging apps sort of need to to use dubious new sources

44
00:05:11,600 --> 00:05:18,890
to push propaganda out reduced and people are increasingly using visual media.

45
00:05:18,890 --> 00:05:32,000
For instance, names, images and data and videos to to convey that to reach audience groups and to manipulate voters.

46
00:05:32,000 --> 00:05:40,490
So for the first time when we studied the Brazilian election, we found very low levels of junk use on Twitter using our classifications and so on.

47
00:05:40,490 --> 00:05:44,960
And of course, that doesn't mean that there was no junk use in the Brazilian elections.

48
00:05:44,960 --> 00:05:48,590
It simply meant that it had migrated to other platforms.

49
00:05:48,590 --> 00:05:57,620
So, of course, that there are a number of issues with accessing information with that same data on what's meant to be an encrypted platform.

50
00:05:57,620 --> 00:06:00,980
So we had to actually ask to join these groups.

51
00:06:00,980 --> 00:06:02,480
So we rejoined these.

52
00:06:02,480 --> 00:06:10,880
First of all, they had to be public WhatsApp groups, which meant that we could only join using a link that was published on the on the web.

53
00:06:10,880 --> 00:06:12,890
And then once we joined these groups,

54
00:06:12,890 --> 00:06:21,840
we had to announce that we had researchers and we had to be very upfront about that and seek that consent of consent before using using that data.

55
00:06:21,840 --> 00:06:24,860
And of course, we've had run ins with WhatsApp,

56
00:06:24,860 --> 00:06:33,500
and they would question if if this was ethical because it was designed as safe as a private, encrypted platform.

57
00:06:33,500 --> 00:06:39,800
And they would also question whether this was representative perfectly valid questions, except that,

58
00:06:39,800 --> 00:06:47,210
you know, WhatsApp, that they knew very well that we had no other way of accessing the data.

59
00:06:47,210 --> 00:06:54,650
And, you know, we were trying to shed light on what was going going on within these groups that.

60
00:06:54,650 --> 00:07:02,660
But that raises some very significant ethical questions about, you know, privacy and data protection.

61
00:07:02,660 --> 00:07:06,980
And how do we and whether researchers have an automatic right?

62
00:07:06,980 --> 00:07:13,640
I mean, in the interests of serving public good and so on to access the sector.

63
00:07:13,640 --> 00:07:21,170
But coming back to this before we were because we saw an increasing amount of propaganda taking the shape of means and images,

64
00:07:21,170 --> 00:07:28,560
we developed different visual typology to to classify the kind of material that we were seeing propaganda material.

65
00:07:28,560 --> 00:07:40,610
So because that was the standard campaign stuff that was satire, religion played a big part that was junk, you know, similarly styled credibility.

66
00:07:40,610 --> 00:07:48,140
But crucially, there was no more this distinction between professional versus non-professional or dubious news sources.

67
00:07:48,140 --> 00:07:52,160
Because, you know, this was this was moving away from attribution.

68
00:07:52,160 --> 00:07:59,150
People did know that was not from a professional news source, this particular image or meme, and that was part of the appeal.

69
00:07:59,150 --> 00:08:04,850
It was still talking about was making that, you know, it was either critiquing policy decisions,

70
00:08:04,850 --> 00:08:13,880
it was making fun of politicians and they seemed to strike a chord with audience groups, although people knew that it was not professionally produced.

71
00:08:13,880 --> 00:08:19,640
So our entire definition of junk news versus professional use and our methodology had to

72
00:08:19,640 --> 00:08:28,160
be adapted to be able to describe what was meant as propaganda and this in this context.

73
00:08:28,160 --> 00:08:33,020
Yeah. And so in the Brazilian context, again, we saw a lot of hate gold and porn,

74
00:08:33,020 --> 00:08:40,790
and that raised a whole different set of questions ethical questions about what do we do when we see,

75
00:08:40,790 --> 00:08:47,100
you know, say, call to action within a WhatsApp group? What is our responsibility as far as researchers?

76
00:08:47,100 --> 00:08:49,160
I mean, we've promised not to intervene,

77
00:08:49,160 --> 00:08:57,560
but if we if we thought that there was actually some danger to a group or to an individual, what do we do in that case and so on?

78
00:08:57,560 --> 00:09:05,240
And also, you know, our responsibility to our own research students and other researchers because they would be

79
00:09:05,240 --> 00:09:10,490
exposed to fairly dire material on social media and how would it impact their well-being?

80
00:09:10,490 --> 00:09:17,240
So these are some of the questions that we have to grapple with when we were doing this research.

81
00:09:17,240 --> 00:09:24,320
So in the Indian context, of course, there are research period coincided with the Pulwama attack.

82
00:09:24,320 --> 00:09:32,450
The Indian Pakistan almost were on the brink of war and so on the exchange of hostilities.

83
00:09:32,450 --> 00:09:42,230
So, you know, nationalism and support for armed forces, we saw a whole lot of pictures celebrating the Indian Army and patriotism and so on.

84
00:09:42,230 --> 00:09:54,630
And so so both. So the key takeaway for us as a research group was that propaganda had shifted from from open platforms

85
00:09:54,630 --> 00:10:01,930
to more encrypted platforms like WhatsApp and from using attribution and dubious new sources to spread.

86
00:10:01,930 --> 00:10:11,170
Gander malicious actors were increasingly using means and images and visual media to influence to reach audience groups.

87
00:10:11,170 --> 00:10:15,940
So into this mix, OK, so I could skip this. This is just WhatsApp.

88
00:10:15,940 --> 00:10:21,580
These are just some results from the Indian elections. So into this week.

89
00:10:21,580 --> 00:10:30,850
So it's already it's already tricky to classify these, these images and so on and then a whole host of ethical questions already.

90
00:10:30,850 --> 00:10:42,370
But then, you know, I think some some of our speakers have already referred to generated adversarial networks and deepfakes and so on.

91
00:10:42,370 --> 00:10:52,870
And you know, the potential that they could have on future misinformation disinformation campaigns was what was beginning to exercise our interest.

92
00:10:52,870 --> 00:10:59,950
And because, you know, my work above, I've done some work on reinforcement learning and so on.

93
00:10:59,950 --> 00:11:07,150
This was that this was something that was of huge interest to me personally as a researcher as well.

94
00:11:07,150 --> 00:11:12,640
So I mean, we saw we saw Katrina talking about GP2, for instance.

95
00:11:12,640 --> 00:11:17,140
And you know, you have speech to text, you have image to image,

96
00:11:17,140 --> 00:11:27,130
you have you have a whole host of advanced A.I. techniques that good that can generate misinformation and disinformation and in the wrong hands,

97
00:11:27,130 --> 00:11:31,660
one could only imagine the impact that would have on future elections.

98
00:11:31,660 --> 00:11:43,720
And in fact, in our even in the recent UK elections, we saw future advocacy produce videos of Boris Johnson and and Jeremy Corbyn each

99
00:11:43,720 --> 00:11:51,550
endorsing the other just to demonstrate the power of these these technologies.

100
00:11:51,550 --> 00:11:56,890
So, yeah, of course, this was researchers. The intention is not to generate misinformation,

101
00:11:56,890 --> 00:12:04,780
but it's to advance the field of A.I. itself and to endow computers with an intrinsic understanding of the world,

102
00:12:04,780 --> 00:12:07,700
which is what I suppose is all about it.

103
00:12:07,700 --> 00:12:17,500
And if you can generate new datasets from existing datasets, that means your reliance on data itself produces.

104
00:12:17,500 --> 00:12:26,830
So this is an example of a progressive Ghana's as they're and this company, Nvidia in the Silicon Valley managed to perfect this technique.

105
00:12:26,830 --> 00:12:32,530
So this is a completely Computer-Generated image, and it's indistinguishable from a human image.

106
00:12:32,530 --> 00:12:36,670
And it's it's not possible to detect that it's it's an art.

107
00:12:36,670 --> 00:12:44,530
It's an artificially generated image using, you know, current techniques because by definition, it is an image.

108
00:12:44,530 --> 00:12:56,650
It's just not a real person. So. So yeah, it completely alters our perception of reality, particularly amongst groups that are that are vulnerable,

109
00:12:56,650 --> 00:13:03,630
that have come that have little or no technical know how so.

110
00:13:03,630 --> 00:13:06,670
So it process a lot of questions for us as a society,

111
00:13:06,670 --> 00:13:16,390
and we do identify if there are groups that are more vulnerable to this kind of generated misinformation and what is as as stakeholders,

112
00:13:16,390 --> 00:13:22,990
what can we do about research as well as governments and the Big Tech companies?

113
00:13:22,990 --> 00:13:29,450
So with that, I'd like to continue. Thank you. Very.