1
00:00:01,860 --> 00:00:10,710
I'm very excited to be able to talk to you a little bit about the research that I do with a couple of my colleagues here with you.

2
00:00:10,710 --> 00:00:18,020
And as you mentioned, I'm a lawyer, so I usually look at things from a legal perspective and.

3
00:00:18,020 --> 00:00:24,680
But what I want to do is actually talk about free aspects where I think that is disrupting legal concepts,

4
00:00:24,680 --> 00:00:28,670
and I think the only way to fix that problem is actually via.

5
00:00:28,670 --> 00:00:33,890
I mean, that is a America approach. So I'm actually arguing that the law might not be fit for purpose,

6
00:00:33,890 --> 00:00:41,690
which is why we need ethicists and technologies to work together to actually give new strength to the law that we currently have.

7
00:00:41,690 --> 00:00:49,580
And yeah, I want to give you three examples where I think we need to focus more on one of the areas that I'm focussing on as well.

8
00:00:49,580 --> 00:00:54,690
One is accountability, one in Spanish and one is privacy. So let's start with accountability.

9
00:00:54,690 --> 00:01:00,500
And I've chosen the example of loan applications and how we have done this previously in the past.

10
00:01:00,500 --> 00:01:06,230
So usually when you go in a human human setting, you go to a bank and you apply for a loan.

11
00:01:06,230 --> 00:01:14,570
The loan officer will ask you a couple of questions. They might ask you what your income is if you have in savings, your employment history.

12
00:01:14,570 --> 00:01:16,260
All these questions make sense.

13
00:01:16,260 --> 00:01:22,680
There is an intuitive link workers and ask you about your financial status in order to find out if you can't afford that one or not.

14
00:01:22,680 --> 00:01:25,910
Right is an intuitive connexion there.

15
00:01:25,910 --> 00:01:32,120
What we see now is that we are moving away from those traditional data sources to make decisions we use and traditional data sources.

16
00:01:32,120 --> 00:01:39,980
So a couple of examples here where credit institutions are now using Facebook profiles to decide if somebody should get a loan.

17
00:01:39,980 --> 00:01:45,650
So they look at you profile pictures, the friends that you have on Facebook, the groups that you join,

18
00:01:45,650 --> 00:01:52,670
the things that you post, the things that you like and they infer if you're a reliable borrower, not.

19
00:01:52,670 --> 00:01:56,750
Similar things happen in the insurance space, where also, for example,

20
00:01:56,750 --> 00:02:05,750
profile social profile network information is being used to define if somebody should get insurance and premiums should get.

21
00:02:05,750 --> 00:02:14,120
So but that's not the only thing. In general, whenever you make decisions, you start to use very untraditional data.

22
00:02:14,120 --> 00:02:21,740
That includes clicking behaviour, typing behaviour, geolocation, ice and ice tracking.

23
00:02:21,740 --> 00:02:26,030
All of that to infer certain things about you, things that can be very privacy,

24
00:02:26,030 --> 00:02:33,230
invasive and unintentional and very counter-intuitive where I don't really understand what my data says about me.

25
00:02:33,230 --> 00:02:39,770
I have no idea how my browsing behaviour will affect my credit rating in the future, and this information can be enhanced.

26
00:02:39,770 --> 00:02:45,230
We have that replicated and shared with a lot of people and not just in the financial services.

27
00:02:45,230 --> 00:02:51,560
Also, important decisions who has to go to jail, who can go to university and if you get hired, fired, promoted,

28
00:02:51,560 --> 00:02:58,010
all of this information of big data is being used for that and that poses questions for accountability because if in the future,

29
00:02:58,010 --> 00:03:02,780
if I don't get a loan, I don't get a job. The first thing that comes to mind is to ask why?

30
00:03:02,780 --> 00:03:06,680
Why did I not get a job? What happened there? I want to have an explanation,

31
00:03:06,680 --> 00:03:12,680
and this is why I started thinking about this topic and look at it at first from a legal perspective to figure out

32
00:03:12,680 --> 00:03:18,080
if we do have a right that algorithm physicians are being explained to you because that would make a lot of sense.

33
00:03:18,080 --> 00:03:20,120
And just looking at from a legal perspective,

34
00:03:20,120 --> 00:03:25,730
what I found is that I don't think we actually have a legally binding right that things up being explained to you.

35
00:03:25,730 --> 00:03:28,400
I was not very happy with the outcome of my paper.

36
00:03:28,400 --> 00:03:33,620
It's very hard to say from an academic perspective to actually admit to that, but it wasn't very happy with the outcome of the paper.

37
00:03:33,620 --> 00:03:38,630
But there was a starting point where I started to think about I don't actually care so much about what the law says.

38
00:03:38,630 --> 00:03:42,950
Just because something is not legally required doesn't mean it's not ethically desired,

39
00:03:42,950 --> 00:03:50,480
which was the reason why Rent and Chris Russo and I got together and we now have research programme at D-I,

40
00:03:50,480 --> 00:03:54,990
which is called the governance of Emerging Technologies. So Brando said, we'll be talking a couple of minutes.

41
00:03:54,990 --> 00:04:00,080
Is an ethicist. I'm Chris Russell is a machine learning expert.

42
00:04:00,080 --> 00:04:08,600
So, yeah, what basically starts like a joke in the sense that a lawyer or a philosopher and a scientist walk into a bar, which we did,

43
00:04:08,600 --> 00:04:15,290
and we spent three hours screaming at each other because we wanted to figure out what a good explanation would actually look like.

44
00:04:15,290 --> 00:04:19,220
What do we think of good explanations because we set the legal question of what I wanted to know?

45
00:04:19,220 --> 00:04:24,350
What is it that you want to know? So Frank was very much interested in the trust side of things.

46
00:04:24,350 --> 00:04:29,540
For him, it was very important to figure out what is justified to believe what a good argument is,

47
00:04:29,540 --> 00:04:34,970
what make arguments valid, and I didn't care about that at all. I went to a justice and accountability.

48
00:04:34,970 --> 00:04:38,510
I wanted to be able to contest the decision, but I'm not happy with it.

49
00:04:38,510 --> 00:04:44,810
And Curtis, a computer scientist, didn't care about each of those things. Just wanting parties could understand what's going on in the black box.

50
00:04:44,810 --> 00:04:48,620
So even though we were all very, very passionate about explanations,

51
00:04:48,620 --> 00:04:55,880
we started to understand that we think very different things about that and see the explanations in very different ways.

52
00:04:55,880 --> 00:04:59,750
But we finally found the solution and actually wrote a paper together,

53
00:04:59,750 --> 00:05:06,830
and we just can't come to factual explanations to try to reconcile those two approaches.

54
00:05:06,830 --> 00:05:14,240
And our method actually allows you to understand a little bit of what's going on in the black box also gives you more trust in a system,

55
00:05:14,240 --> 00:05:20,720
and it gives you accountability because you would be able to, for example, a decision not happy with it.

56
00:05:20,720 --> 00:05:27,740
So the exciting part of that is that it actually got off traction and Google has implemented our our method.

57
00:05:27,740 --> 00:05:34,100
Now, last year in TensorFlow Zoo, nine can actually play with algorithms and understand what's going on in a black box.

58
00:05:34,100 --> 00:05:42,530
And Google is not the only one who is very fond of our idea because IBM, Microsoft and FLoC also were our idea.

59
00:05:42,530 --> 00:05:48,110
So what that means is if you scream a lot of each other, it actually pays off at some ways.

60
00:05:48,110 --> 00:05:56,630
But we we also learnt that explanations are only one facet of true accountability because explanations are not justification.

61
00:05:56,630 --> 00:06:05,060
I can tell you, for example, I'm not giving you the loan because I don't like you face explanations.

62
00:06:05,060 --> 00:06:07,010
It doesn't mean it's justified.

63
00:06:07,010 --> 00:06:14,840
So actually, what you also want to look at is the inference as to predictions and the opinions that algorithms have about you, right?

64
00:06:14,840 --> 00:06:19,580
All that big data is being collected about you, and very sensitive things can be inferred.

65
00:06:19,580 --> 00:06:24,650
For example, if you're pregnant, if you're liable more, if you work like, you should get promotion.

66
00:06:24,650 --> 00:06:30,230
If you have undiagnosed disorders, all of that information can be very privacy invasive.

67
00:06:30,230 --> 00:06:33,650
And that is the actual thing that you're concerned about.

68
00:06:33,650 --> 00:06:38,930
So again, I look at it from a from a legal perspective and try to look at data protection law because,

69
00:06:38,930 --> 00:06:46,310
you know, this is very close to your private life. So if inferences, for example, of personal data, you would get a lot of data protection rights.

70
00:06:46,310 --> 00:06:52,250
And again, what I found is that I actually disrupts the law in a very untraditional.

71
00:06:52,250 --> 00:06:55,740
Where we have to think creatively about the problems.

72
00:06:55,740 --> 00:07:02,880
So, Brendan, I wrote a paper together, which is called a rights to reasonable inferences rethinking self-protection law,

73
00:07:02,880 --> 00:07:11,760
the age of data and actually calling for new standards because we found that the law, as it currently stands, is not good enough to protect us.

74
00:07:11,760 --> 00:07:16,590
As you can see, it's 130 pages trying to make the point that law is not good enough.

75
00:07:16,590 --> 00:07:22,710
It was actually planned to be a hundred and six, but threatened to work with me again if I don't stop writing.

76
00:07:22,710 --> 00:07:29,700
So it's just on 30 pages, but it's very important to point out that we found a lot of loopholes in the law,

77
00:07:29,700 --> 00:07:34,290
some of which are that we don't even know if inference is a personal data.

78
00:07:34,290 --> 00:07:37,890
So all the assumptions and algorithms are making a be made about you.

79
00:07:37,890 --> 00:07:42,900
It's not clear if it actually found a good PR, even if they do fall under the data protection law.

80
00:07:42,900 --> 00:07:48,250
What the purpose of data protection lawyers is not to regulate how we make decisions about you.

81
00:07:48,250 --> 00:07:52,380
So if you have a problem over how you are being evaluated, how you being seen,

82
00:07:52,380 --> 00:07:56,340
how you're being assessed, you would need to find another law that helps you do that.

83
00:07:56,340 --> 00:08:02,280
Data protection law doesn't give you a remedy over how you being seen by algorithms,

84
00:08:02,280 --> 00:08:06,570
and that is quite problematic because as was already mentioned very often,

85
00:08:06,570 --> 00:08:11,040
we don't have standards for accurate and reasonable decision making because in essence,

86
00:08:11,040 --> 00:08:16,470
a lot of times we don't have a right to get a job or insurance, although we'll go to university.

87
00:08:16,470 --> 00:08:22,380
So if you don't have any loss and regulate harmless decisions have to be made and data protection doesn't give you that.

88
00:08:22,380 --> 00:08:26,550
Maybe it's the ethical thing to think about what would be reasonable.

89
00:08:26,550 --> 00:08:30,180
And that leads me to the current research project that we just started,

90
00:08:30,180 --> 00:08:35,580
which is a right to reasonable inferences in online advertisement and in financial services.

91
00:08:35,580 --> 00:08:42,330
And we just had to docs starting with as one, as a philosopher, the other one a legal scholar.

92
00:08:42,330 --> 00:08:49,500
And we're trying to figure out what would be a reasonable standard for international ethics in those two sectors in the future.

93
00:08:49,500 --> 00:08:57,150
We're going to look at other sectors as well that will include, for example, health, employment and criminal justice.

94
00:08:57,150 --> 00:09:08,430
But what is the current focus of our project? The last area where I see that is actually disrupting the law has to do with non-discrimination law.

95
00:09:08,430 --> 00:09:12,210
Again, that's something that we all are aware of whenever we talk about it.

96
00:09:12,210 --> 00:09:15,690
We also have to think about bias and discrimination. So again,

97
00:09:15,690 --> 00:09:20,880
I did look at non-discrimination law and try to figure out if the law is actually good

98
00:09:20,880 --> 00:09:25,740
enough and it's good enough to protect us against those unintended and undesirable risks.

99
00:09:25,740 --> 00:09:29,580
And again, I have to say it doesn't, and it doesn't.

100
00:09:29,580 --> 00:09:34,470
And actually, it's not really the law's fault or technology fault. It's just something that happen.

101
00:09:34,470 --> 00:09:40,260
If you're an Oscar nomination, law is an answer to undesired behaviour of people, right?

102
00:09:40,260 --> 00:09:43,320
That means that very often we want to regulate.

103
00:09:43,320 --> 00:09:50,580
It prohibits that people are unethical to each other or being racist or sexist or homophobic to each other.

104
00:09:50,580 --> 00:09:55,230
But the way that algorithms do that is very, very different than humans.

105
00:09:55,230 --> 00:10:00,780
So, for example, if you think about price discrimination,

106
00:10:00,780 --> 00:10:09,810
suppress discrimination means that the same product is offered to different people at different prices in the offline world.

107
00:10:09,810 --> 00:10:12,120
You will be able to get easy for you to figure it out.

108
00:10:12,120 --> 00:10:20,670
You can just go to different stores and compare fruits and their prices and pick the products that you think is fair or appropriate.

109
00:10:20,670 --> 00:10:26,880
If, for example, Tesco doesn't let you into the store anymore, you know that you have been excluded from the market.

110
00:10:26,880 --> 00:10:32,670
How do you know that in an online world, right? How do you know if you actually being offered the best price?

111
00:10:32,670 --> 00:10:38,580
Do you know the precious other people get offered? And do you know the advertisements that you don't see, right?

112
00:10:38,580 --> 00:10:43,980
So that kind of discrimination, you don't feel it anymore, whereas in the offline world,

113
00:10:43,980 --> 00:10:48,990
there will be often the ability of the complainant to see that something's off.

114
00:10:48,990 --> 00:10:56,310
Something feels unfair, unjust, but that sense a feeling of unfairness might not be as easy to grasp anymore.

115
00:10:56,310 --> 00:11:07,440
And similarly, here with the law, if you look at the classic that we want to protect, that relates to sex and gender, ethnicity,

116
00:11:07,440 --> 00:11:15,930
religious beliefs because we had historical experiences where people have used that to discriminate against people, right?

117
00:11:15,930 --> 00:11:23,190
So this is how we constructed the world, how we group people. But algorithms might prove people completely differently, right?

118
00:11:23,190 --> 00:11:31,140
It could be that the algorithms start to discriminate against people who are born on a Tuesday, have brown socks and like dogs.

119
00:11:31,140 --> 00:11:40,650
But this class of people doesn't find recognition in non-discrimination law because we never have it or similar here with video gamers,

120
00:11:40,650 --> 00:11:45,780
which, for example, in China, is something that could cause your credit score to drop.

121
00:11:45,780 --> 00:11:52,190
Obviously, that class of people doesn't find any protection in non-discrimination law, but maybe.

122
00:11:52,190 --> 00:12:00,260
And maybe, too, so this is the the project that we also are working on trying to figure out if the law doesn't give you enough protection,

123
00:12:00,260 --> 00:12:06,650
what would be the ethical thing to do? How can we increase and protect fairness in algorithmic world?

124
00:12:06,650 --> 00:12:12,500
So, yeah, basically does the the the programme that you're currently working on.

125
00:12:12,500 --> 00:12:17,540
And I actually think that the one of the reasons why I'm so excited about all this,

126
00:12:17,540 --> 00:12:24,050
what's happening because I think it finally is clear to everyone that if you want to,

127
00:12:24,050 --> 00:12:31,850
you know, use, if you need to think of it, at least three perspectives need to think about is the law actually helping you isn't all good enough.

128
00:12:31,850 --> 00:12:33,990
And if it's not, what would be desirable?

129
00:12:33,990 --> 00:12:39,710
Well, the ethical thing to do and then actually match it against technology and figure out if it's actually feasible.

130
00:12:39,710 --> 00:12:45,410
And if you do that, you can harness the full potential of it, but make sure that you protect human rights.

131
00:12:45,410 --> 00:12:47,765
Thank you.