1
00:00:00,210 --> 00:00:11,610
It's my great pleasure to welcome as our first speaker, appropriately so Professor Sir Nigel Shadbolt, principal of Jesus College,

2
00:00:11,610 --> 00:00:18,030
who's a professorial research fellow in the Department of Computer Science and chairman of the Open Data Institute,

3
00:00:18,030 --> 00:00:23,130
which he co-founded with Tim Berners Lee, amongst many other distinctions.

4
00:00:23,130 --> 00:00:30,810
Nigel is a fellow of the Royal Society. And in 2013 was knighted for services to science and engineering.

5
00:00:30,810 --> 00:00:40,860
He's published over 500 articles on topics ranging from cognitive psychology to computational neuroscience and from A.I. to the Semantic Web.

6
00:00:40,860 --> 00:00:50,970
And Nigel has been fronting the universities in Connexions with the Schmall Schwartzman Foundation to bring about this new institute.

7
00:00:50,970 --> 00:01:04,980
So, Nigel, over to you. Thank you. So I just wanted to set the scene really for today's wonderful set of talks,

8
00:01:04,980 --> 00:01:11,100
vignettes around the range and scope of what's going on around the challenges of

9
00:01:11,100 --> 00:01:15,600
ethics in many of the advanced technology settings that we were almost engaged in.

10
00:01:15,600 --> 00:01:18,930
And also to explain where this all came from.

11
00:01:18,930 --> 00:01:30,430
So we know that the Schwartzmann Institute itself, the Schwartzmann centre itself, is a much bigger endeavour involving the humanities in general.

12
00:01:30,430 --> 00:01:35,070
And one thing to be very clear about, when people heard the announcement of 150 million pounds.

13
00:01:35,070 --> 00:01:38,690
Some people mistakenly imagined that was for the Institute for Ethics in.

14
00:01:38,690 --> 00:01:43,800
Now, I know it's a it's a part of that gift.

15
00:01:43,800 --> 00:01:49,110
And it's the part that is particularly exciting because it's there to support human talent,

16
00:01:49,110 --> 00:01:54,930
to support appointments in the area, in the faculty of philosophy in particular.

17
00:01:54,930 --> 00:02:02,640
And I'll talk about those. So it's a significant but a very minority part of that gift.

18
00:02:02,640 --> 00:02:08,550
But it's a significant and the whole challenge of this is, of course,

19
00:02:08,550 --> 00:02:14,820
premised on the fact that we know that computer science and particularly artificial intelligence present huge challenges.

20
00:02:14,820 --> 00:02:24,480
This is one of my favourite quotes. This is from cybernetics. Father Norbert Viña in 1948 already foresaw that, as he called it,

21
00:02:24,480 --> 00:02:31,770
the ultra rapid computing machine was an ideal central nervous system for an apparatus for automatic control 1948.

22
00:02:31,770 --> 00:02:38,840
He foresaw the opportunities for. Unheard of amounts of good and evil.

23
00:02:38,840 --> 00:02:40,780
And boy, was he right there.

24
00:02:40,780 --> 00:02:49,640
And of course, you're the forcing function that we recognise in that much technology accelerates under the impulse of of conflict, sadly.

25
00:02:49,640 --> 00:02:59,360
So whether it's Bletchley Park, where, of course, so much of the early work on automated computing happened, kept much of it classified,

26
00:02:59,360 --> 00:03:06,020
sharing himself, of course, working there or it was the Manhattan Project where can computation was required.

27
00:03:06,020 --> 00:03:15,860
This forcing function accelerate both the demands, the technology, but also very quickly led to issues around the fundamental ethical questions.

28
00:03:15,860 --> 00:03:23,340
The other thing I like to present is the fact that modern A.I. is understood very often to simply be all about machine learning.

29
00:03:23,340 --> 00:03:33,590
And as somebody who has been working the field since the late 70s, I've seen a number of cycles of A.I. come and go.

30
00:03:33,590 --> 00:03:37,130
And in that time, different methodologies have held sway.

31
00:03:37,130 --> 00:03:42,860
Back in the day was rule based systems theory improving a whole range of methods to do search.

32
00:03:42,860 --> 00:03:49,640
And each time round the cycle of enthusiasm, particular human capabilities fell to the dominance of the machines.

33
00:03:49,640 --> 00:03:57,440
There was always a question around the ethical deployment of this, even back in the 80s when people were talking about building expert systems.

34
00:03:57,440 --> 00:04:03,050
What about the experts the systems would displace? Could we trust the actual diagnosis?

35
00:04:03,050 --> 00:04:09,200
The systems would give us worthy accounts, explicable and understandable and so on.

36
00:04:09,200 --> 00:04:15,320
So there is no there is no shift in some of the fundamental challenges facing us over many decades.

37
00:04:15,320 --> 00:04:21,530
But what has happened clearly in recent times is the emergence of a particular class of computing power,

38
00:04:21,530 --> 00:04:26,930
a method that has led to some breakthrough moments. These so-called deep neural networks,

39
00:04:26,930 --> 00:04:34,460
various forms of machine learning have led both to the try the triumph of machines over particular areas of human expertise.

40
00:04:34,460 --> 00:04:41,480
This, of course, is Alpha Go's triumph over Lisa Abdulle and most recently achieving extraordinary results

41
00:04:41,480 --> 00:04:47,000
in what was thought again to be quite difficult to master areas of problem solving.

42
00:04:47,000 --> 00:04:57,560
This is a multi-strategy king called Starcraft. It's it's the latest achievement, again, of the deep mind science research labs.

43
00:04:57,560 --> 00:05:04,550
Of course, within all of that, there's plenty of good things going on. But again, sometimes with issues around the ethics,

44
00:05:04,550 --> 00:05:15,590
this is the Google DMI at work on diagnosing various diseases from the retinal scans that were available,

45
00:05:15,590 --> 00:05:21,560
whether it's various forms of in this case, this is diabetic retinopathy.

46
00:05:21,560 --> 00:05:27,530
These systems are very good at detecting patterns and coming up with all sorts of really rather impressive classifications.

47
00:05:27,530 --> 00:05:33,140
But even there, there was an issue ultimately that fell into the whole area of was the data used

48
00:05:33,140 --> 00:05:37,310
to train these systems acquired with the appropriate level of informed consent.

49
00:05:37,310 --> 00:05:45,240
It is almost impossible to think of an AIM deployment mode in which ethical questions don't abound.

50
00:05:45,240 --> 00:05:52,100
Oxford's own working let net wonderful work to actually get to the level of human expertise in reading.

51
00:05:52,100 --> 00:05:57,530
You can imagine various forms of more intriguing deployment of these kinds of technologies.

52
00:05:57,530 --> 00:06:07,430
In fact, it was the central thesis of 2001. How, if you remember, read the lips of the astronauts in the pod, or indeed facial recognition,

53
00:06:07,430 --> 00:06:14,960
which is already causing concerns about its sensibilities and sensitivities around particular forms of bias.

54
00:06:14,960 --> 00:06:20,420
Given certain sorts of training regime and given a particular deployment notes.

55
00:06:20,420 --> 00:06:27,830
The all more obvious ones, of course, range from what you do with autonomous vehicles to how you should restrict

56
00:06:27,830 --> 00:06:35,480
and control the weaponization of of platforms such as A.I. included in drone.

57
00:06:35,480 --> 00:06:41,390
More extensively, we worry about the use of extensive amounts of data in domains such as predictive

58
00:06:41,390 --> 00:06:47,570
policing or indeed at a national scale when deployed by states such as China.

59
00:06:47,570 --> 00:06:55,340
In that social credit system, we worry about the flows of data from our mobiles, from our devices.

60
00:06:55,340 --> 00:07:01,490
In fact, my group in Oxford has spent a lot of time trying to understand just what that ecosystem looks like.

61
00:07:01,490 --> 00:07:05,360
And it's clear that the flows of data are extensive. They're numerous.

62
00:07:05,360 --> 00:07:11,210
This is just one particular flow of data from one app surrounded by a few of us on a phone that we were able to

63
00:07:11,210 --> 00:07:20,630
track and understand just how extraordinary the economy of data exchange is and where is the control of that.

64
00:07:20,630 --> 00:07:30,280
And whereas the oversight and insight in that often your ethical challenges are cheek by jowl with questions around governance and regulation.

65
00:07:30,280 --> 00:07:35,830
And within all of this, we have concerns about the emergence of dominant platforms both in.

66
00:07:35,830 --> 00:07:45,200
East and west, who seem to have so much control in this emerging world of data enabled A.I. algorithms

67
00:07:45,200 --> 00:07:49,760
and data able to aggregate algorithms that they say aren't just about machine learning,

68
00:07:49,760 --> 00:07:56,330
they can just as much be about expert system reasoning, about various forms of model based reasoning.

69
00:07:56,330 --> 00:08:04,850
There are a very large set of methods now available to the computer scientist and the engineer, the various people looking to exploit the methods,

70
00:08:04,850 --> 00:08:12,230
even if it is in some cases, statistics read, presented various forms of linear regression or whatever.

71
00:08:12,230 --> 00:08:19,040
There will still often be ethical challenges at the end of that deployment issue.

72
00:08:19,040 --> 00:08:24,710
So the question for us isn't just so to think about this in the narrow confines of what we might think

73
00:08:24,710 --> 00:08:34,340
of as a robotic controlled drone or a particular use of a neural network in a biometric system.

74
00:08:34,340 --> 00:08:42,760
It can be how data and Arab rhythms are used very broadly on the Web, how they're used in targeting and surveilling.

75
00:08:42,760 --> 00:08:46,200
So in the broadest sense, this was always the ambition.

76
00:08:46,200 --> 00:08:57,140
And so when I was asked to put together a proposal for the Schwartzmann initiative and this was very much driven out of his own concerns,

77
00:08:57,140 --> 00:09:06,230
Swartz was unconcerned that he was seeing a world emerge where his particular worry was around the concern of how these systems might be deployed.

78
00:09:06,230 --> 00:09:12,980
It was natural to turn to oxfords extraordinary, extraordinary heritage in this area.

79
00:09:12,980 --> 00:09:21,680
It's not hard to make the case for ethics, of course. But Oxford, you have the most extraordinary cast of characters here.

80
00:09:21,680 --> 00:09:28,850
Looking through a history that includes everybody from had to parfit to to to warn Ockham Murdoch.

81
00:09:28,850 --> 00:09:35,460
These are these these are extraordinary figures who have shaped our thinking in the ethical and moral space.

82
00:09:35,460 --> 00:09:43,640
And in some respects, of course, Mary Warnock's work also held an appeal to many because this was a person who

83
00:09:43,640 --> 00:09:50,200
managed to convene an entire regulatory framework and lead on that thinking

84
00:09:50,200 --> 00:10:00,590
and the deployment of a technology which at the time concerned people the whole way in which human fertilisation and embryology science was moving,

85
00:10:00,590 --> 00:10:03,050
reproductive science was moving.

86
00:10:03,050 --> 00:10:11,070
And that multistakeholder conversation, which had serious ethical underpinnings, was one of the reasons that we found the way forward to nuance.

87
00:10:11,070 --> 00:10:19,580
That interesting debate. And one of the things I think we will see as the landscape develops here in Oxford is

88
00:10:19,580 --> 00:10:26,900
this very interesting appeal to lessons learnt from fields such as medical ethics.

89
00:10:26,900 --> 00:10:30,770
Many of the problems addressed have near analogies.

90
00:10:30,770 --> 00:10:36,410
Some will turn out to be different, whether it is the use of human subjects or what is being done to them or the question

91
00:10:36,410 --> 00:10:41,900
of informed consent or how access is granted to two parts of the population.

92
00:10:41,900 --> 00:10:45,950
Many of these will have their equivalents in what we do.

93
00:10:45,950 --> 00:10:56,660
So when we looked at the proposal as we assembled the proposal, we were surrounded really by a surfeit of riches.

94
00:10:56,660 --> 00:11:01,550
This diagram just begins to reflect. This is we call this a heliocentric bottle.

95
00:11:01,550 --> 00:11:10,040
This is not to imply, by the way, that the you are all circling this quite not quite existent existing ethics in a institute.

96
00:11:10,040 --> 00:11:16,430
This is meant to represent the extraordinary level of import and cognate activities that

97
00:11:16,430 --> 00:11:20,570
we can look at from computer science and our involvement in the Alan Turing Institute,

98
00:11:20,570 --> 00:11:24,780
on the one hand, to the work that goes on in the Oxford Internet Institute.

99
00:11:24,780 --> 00:11:28,640
Hear about some of that today and indeed the work and the ahero Centre for Practical

100
00:11:28,640 --> 00:11:34,070
Ethics Work in the Human Future of Humanises Institute work in a governance,

101
00:11:34,070 --> 00:11:36,810
working diplomatic school on policy,

102
00:11:36,810 --> 00:11:48,170
a whole set of interests emerging in law around how we might think about legal restraints and applications in technology in medicine.

103
00:11:48,170 --> 00:11:54,200
Again, the Welcome Centre for Ethics, the Oxford Martin School, Oxford Foundry for Innovation,

104
00:11:54,200 --> 00:12:00,170
the Big Data Institute up in the Medical Science Division and Information and Engineering,

105
00:12:00,170 --> 00:12:05,000
where much of the really interesting robotics development goes on and other things too.

106
00:12:05,000 --> 00:12:09,550
So it's a very rich surround. This won't even be completely comprehensive.

107
00:12:09,550 --> 00:12:15,210
But what's compelling about that is that you set this effort.

108
00:12:15,210 --> 00:12:22,620
Within a nexus of really interesting research, and I think one of the things that Peter's looking to do is to is to is to help convene

109
00:12:22,620 --> 00:12:28,980
the plurality of those conversations around the interests of the ethics in a institute.

110
00:12:28,980 --> 00:12:35,640
Now, that, of course, is as yet to be staffed, to be to be established.

111
00:12:35,640 --> 00:12:44,400
But the interesting, absolute essential difference of this and some people said there are any number of ethics in a I or ethics initiatives,

112
00:12:44,400 --> 00:12:50,190
somebody counter some extraordinary number of ethics, a coach the other day and it was in the low hundreds.

113
00:12:50,190 --> 00:12:56,430
Believe it or not, lots of them are close similarities to one another.

114
00:12:56,430 --> 00:13:02,730
And a lot of it. A lot of it. One suspects that various forms of virtue signalling lots of it.

115
00:13:02,730 --> 00:13:12,270
You might suspect some kind of copying to get into something that seems an idea whose idea of the moment.

116
00:13:12,270 --> 00:13:20,040
I think the difference in the effort here is to locate this whole enterprise in the deep research of

117
00:13:20,040 --> 00:13:27,870
the faculty of philosophy in a tradition where the philosophical questions are primary are paramount.

118
00:13:27,870 --> 00:13:36,180
Now, we can divide those up in any any variety of ways in whatever set of questions you raise here.

119
00:13:36,180 --> 00:13:45,900
There are more whether it's about when you use or don't use the technology, whether it's fair, who's responsible for it, who has access to it.

120
00:13:45,900 --> 00:13:49,650
Does it sufficiently explain itself? Does it varies?

121
00:13:49,650 --> 00:13:53,880
You go from one geography to another to one culture to another?

122
00:13:53,880 --> 00:13:58,560
What about its utilisation for public and private goods?

123
00:13:58,560 --> 00:14:08,340
And all of this just provides a sense of the depth and range of ethical questions that will arise in the last few minutes.

124
00:14:08,340 --> 00:14:13,940
Let just me say something about the shape and structure of where we are.

125
00:14:13,940 --> 00:14:20,780
So this all began with a group of individuals who gave of their time and energy are hugely grateful

126
00:14:20,780 --> 00:14:28,080
to them in a steering committee that was put together back in actually almost a year ago.

127
00:14:28,080 --> 00:14:33,530
It seems extraordinary, just I think I was approached about a year ago and and then indeed,

128
00:14:33,530 --> 00:14:39,380
in January of this year, we began to think about shaping the proposal.

129
00:14:39,380 --> 00:14:48,230
So myself and Chris Timson, head of philosophy, Dan Grimly, Phil Howard from the ally, Mike Wooldridge, computer science.

130
00:14:48,230 --> 00:14:53,120
Aseel Fabbri, a philosopher. Mike Parker and Alison Noble.

131
00:14:53,120 --> 00:14:59,210
Mike Parker from the Welcome Institute. And Allison from Information Engineering.

132
00:14:59,210 --> 00:15:01,460
Actually, Alison, replace Angela McClain,

133
00:15:01,460 --> 00:15:11,720
who was the original originally on the committee as Angela took up the chief scientific adviser role in one of the Ministry of Defence.

134
00:15:11,720 --> 00:15:17,740
So that group had gotten together and tried to frame and shape.

135
00:15:17,740 --> 00:15:24,880
The proposal in a way, that was that was going to work for for the university.

136
00:15:24,880 --> 00:15:31,690
The plan ultimately is that there'll be a management committee of the institute to be confirmed.

137
00:15:31,690 --> 00:15:39,700
And it will be Cross Divisional. Probably will mirror the kinds of things we see within things like the Bullivant Neek School.

138
00:15:39,700 --> 00:15:43,690
The advisory board will comprise internal and external members.

139
00:15:43,690 --> 00:15:46,050
They will be advisory. They will not set the agenda.

140
00:15:46,050 --> 00:15:53,440
It's important to say, and we are at this moment about to advertise for a director for the institute.

141
00:15:53,440 --> 00:15:56,530
My role is essentially to try and steward this into existence.

142
00:15:56,530 --> 00:16:01,360
I think the important thing to say is of all of the activities associated with the guest.

143
00:16:01,360 --> 00:16:06,460
This is one of the first ones to to begin its work because the building,

144
00:16:06,460 --> 00:16:14,800
which will be placed on the Rateliff Observatory site, will be a number of years in development.

145
00:16:14,800 --> 00:16:23,920
I mean, it could be four to five years before that is actually opened. In the meantime, we would like to be undertaking this exciting research agenda.

146
00:16:23,920 --> 00:16:29,500
So a director and initially two associate professors are being will be being advertised,

147
00:16:29,500 --> 00:16:34,480
one of these in philosophy and one in philosophy and computer science.

148
00:16:34,480 --> 00:16:37,420
There are five associate professorships ultimately to be appointed, two.

149
00:16:37,420 --> 00:16:44,110
So this will be a substantial centre of gravity in terms of the ability to research and teach.

150
00:16:44,110 --> 00:16:51,880
And the other thing to say is that these will be at least the first two advertize will be college associations as well.

151
00:16:51,880 --> 00:16:57,820
So we're trying to wrap the Institute Stot into the Collegiate University.

152
00:16:57,820 --> 00:17:03,550
There will be a number of post-doctoral research fellows or jail appointments.

153
00:17:03,550 --> 00:17:06,610
There will be a significant number of Theophile students.

154
00:17:06,610 --> 00:17:11,500
We are looking to use the model of academic secondment to move people into the institute for periods

155
00:17:11,500 --> 00:17:18,790
of time where their research would help and be valuable and a significant vesting fellows programme.

156
00:17:18,790 --> 00:17:31,780
So in total, we might imagine that in four or so years time there would be 20 or so individuals at the core of the of the institute and ultimately,

157
00:17:31,780 --> 00:17:43,370
of course, looking to develop and expand with joint programmes of work to even more significant, more significant size.

158
00:17:43,370 --> 00:17:50,080
The other thing is probably worth mentioning is that there was always the ambition that we would develop.

159
00:17:50,080 --> 00:17:55,120
Various forms of curriculum, content to inform our teaching.

160
00:17:55,120 --> 00:18:00,650
And the interesting model here is to think about the evolution of our various.

161
00:18:00,650 --> 00:18:04,550
Courses where philosophy has been at their heart.

162
00:18:04,550 --> 00:18:08,990
Whether it is the original literary Amarna raise with the Greeks, as it was called, the modern Greeks,

163
00:18:08,990 --> 00:18:18,320
which reflected the fact that economic elites have become a subject that required required study, political science and philosophy.

164
00:18:18,320 --> 00:18:24,470
The scientific Greeks that we saw evolve philosophy, psychology, linguistics in the 20th century,

165
00:18:24,470 --> 00:18:28,730
physics and philosophy, maths and philosophy, computer science, philosophy.

166
00:18:28,730 --> 00:18:33,860
One of the interesting things about these is the extent to which computational thinking,

167
00:18:33,860 --> 00:18:40,640
the intrusion of new ways of understanding our world has been very much at the centre.

168
00:18:40,640 --> 00:18:51,380
So the question will be how we can develop content that is capable of being inserted into even Masters and on the great as undergraduate modules.

169
00:18:51,380 --> 00:19:02,300
That would certainly be an ambition. And I'm certainly aware that in computer science we have a significant need for appropriate ethics courses.

170
00:19:02,300 --> 00:19:12,350
It's not in and of itself the only thing, of course, going on. There is an ongoing cultural programme that will be being launched.

171
00:19:12,350 --> 00:19:15,950
In fact, there isn't events, A.I. and creativity, I think.

172
00:19:15,950 --> 00:19:25,100
On November the 20th, Wednesday, that's actually also being held that's being held at the maths institute.

173
00:19:25,100 --> 00:19:28,790
And that's certainly something to look out for. One of the kind of features, again,

174
00:19:28,790 --> 00:19:37,670
is that we look to really motivate the presence of this institute within the context of being human in the 21st century.

175
00:19:37,670 --> 00:19:43,400
How, despite all the concerns we have, or perhaps exactly because of those concerns,

176
00:19:43,400 --> 00:19:51,620
we can understand an appropriate set of balanced interventions with this technology.

177
00:19:51,620 --> 00:19:57,290
As I mentioned, that the site itself is not yet built.

178
00:19:57,290 --> 00:20:02,180
And indeed, the whole process for selecting architects is in train as we speak.

179
00:20:02,180 --> 00:20:14,380
To get that process underway. And this finally, just to say, the actually, I guess apart from a talk, I gave it up.

180
00:20:14,380 --> 00:20:17,590
So this is the single earliest manifestation of actual activity,

181
00:20:17,590 --> 00:20:25,060
which is all of you in this room listening to a variety of presentations as to why ethics has

182
00:20:25,060 --> 00:20:33,580
a place in the consideration deliberations of of Oxford's extraordinary intellectual history.

183
00:20:33,580 --> 00:20:35,890
And yeah, that really is my introduction.

184
00:20:35,890 --> 00:20:41,320
This is one of the first in fact, the first two of these are very much like community meetings, town hall meetings,

185
00:20:41,320 --> 00:20:47,560
where we're convening the interested parties together to understand what we're trying to achieve.

186
00:20:47,560 --> 00:21:02,250
So that's my introduction. Thank you. Thanks very much indeed, Nigel.

187
00:21:02,250 --> 00:21:08,610
When these seminars were conceived, it was not clear what form they should take.

188
00:21:08,610 --> 00:21:19,110
But as Nigel's explained, one of the functions of a very major function of the Institute for Ethics in A.I. is to bring together Oxford academics and

189
00:21:19,110 --> 00:21:28,920
researchers and students from from across a very wide range of centres and other institutions who often may not know each other,

190
00:21:28,920 --> 00:21:33,660
even though they're working in closely related areas. Just half a mile apart.

191
00:21:33,660 --> 00:21:42,330
So what we're planning to do in the first of these few seminars is not to go deeply into particular topics of interest,

192
00:21:42,330 --> 00:21:47,970
but rather to show a variety of work that's going on in different parts of the university.

193
00:21:47,970 --> 00:21:53,990
So those who attend can get to understand something of that ecosystem.

194
00:21:53,990 --> 00:21:59,760
We're combining this event, as you know, with nibbles and drinks afterwards.

195
00:21:59,760 --> 00:22:04,800
That's not just an added extra. That's an integral part of the conception.

196
00:22:04,800 --> 00:22:09,930
That way, you if you hear speakers saying things that particularly interest you,

197
00:22:09,930 --> 00:22:13,770
that people you might like to work with, call them afterwards, have a chat.

198
00:22:13,770 --> 00:22:22,010
We've got plenty of time for that. So the way the the rest of today's session until then is going.

199
00:22:22,010 --> 00:22:28,430
We've got three groups of speakers. The first two are from the ahero centre.

200
00:22:28,430 --> 00:22:38,600
The next three from the Oxford Internet Institute. And then we began to round off with two speakers who focussed particularly on A.I. in ethics.

201
00:22:38,600 --> 00:22:48,440
Sorry. Hey, I need medicine. I beg your pardon. So we we've got a mix here of some talks that are more theoretical.

202
00:22:48,440 --> 00:22:51,920
How is it that philosophy fits into this mix?

203
00:22:51,920 --> 00:22:57,050
How does philosophy influence practical ethics?

204
00:22:57,050 --> 00:23:03,830
From Oxford Internet Institute, the focus is going to be partly on law, legal regulation.

205
00:23:03,830 --> 00:23:10,100
And then when we get onto medicine, of course, we really are at the sharp end in a literal sense.

206
00:23:10,100 --> 00:23:20,530
And we will hear there about what it's like actually working with a AI in medicine and regulation within the health service.

207
00:23:20,530 --> 00:23:25,330
So without further ado, I'm going to ask Tom to kick off.

208
00:23:25,330 --> 00:23:33,440
Tom is from the Oxford Internet Institute. Sorry, I'm not doing very well, am I?

209
00:23:33,440 --> 00:23:38,980
I apologise. From the Here Heroes Centre, he trained Tom Douglas.

210
00:23:38,980 --> 00:23:46,590
It trained as a medical doctor and philosopher and his senior research fellow and director of research and development at the Occiput,

211
00:23:46,590 --> 00:23:52,170
Your Hero Centre for Practical Ethics. And he's also a Hugh Price fellow at Jesus College.

212
00:23:52,170 --> 00:24:02,220
His research currently focuses on the ethics of predicting and modulating behaviour, a topic on which will shortly take up an ELC consolidator award.

213
00:24:02,220 --> 00:24:06,290
So, Tom, over to you.

214
00:24:06,290 --> 00:24:17,270
So do you have a centre is based within the faculty of philosophy and is devoted to to doing research in practical or applied ethics.

215
00:24:17,270 --> 00:24:22,320
And as the name suggests, it's directed by at during zamili school.

216
00:24:22,320 --> 00:24:31,790
He holds the chair in practical ethics and currently includes 24 other academic staff members, as well as five doctoral students.

217
00:24:31,790 --> 00:24:43,530
Now we have a very broad remit, which is basically just to bring philosophical analysis to bear on practically important ethical questions,

218
00:24:43,530 --> 00:24:47,820
questions about how well you want to live, how society ought to be arranged and so on.

219
00:24:47,820 --> 00:24:56,160
And as you might suspect, given that broad agreement, we work on a fairly diverse range of topics, some of which are listed on the site.

220
00:24:56,160 --> 00:25:01,870
Historically, you've had quite a strong focus on ethical issues raised by medicine and the life sciences.

221
00:25:01,870 --> 00:25:14,440
But we've also worked in other areas like criminal justice, criminal justice, ethics, Internet ethics and military ethics.

222
00:25:14,440 --> 00:25:21,120
Like I said, so several of our projects bring philosophy to bear on questions in ethics.

223
00:25:21,120 --> 00:25:26,490
And what I want to do in that in the few minutes that I have left is just to briefly mentioned some of these.

224
00:25:26,490 --> 00:25:31,830
I went out to cover all of them just by way of illustrating some of the ways in which rethink some

225
00:25:31,830 --> 00:25:38,070
of the areas in which we think philosophical ethics might have something to say about AI ethics.

226
00:25:38,070 --> 00:25:47,760
And then Caressa is going to go into a bit more detail about exactly how philosophy can contribute to debates about AI ethics.

227
00:25:47,760 --> 00:25:54,780
OK, so the first strand of research that I wanted to mention is some work that Julian Cervalis Group and Guy Keohane are doing,

228
00:25:54,780 --> 00:26:05,980
along with Chris Kinjo from Melbourne University on the role of public preferences in informing the development of ethical algorithms.

229
00:26:05,980 --> 00:26:12,040
So as you're probably all aware, one of the challenges posed by, for example, autonomous vehicles is,

230
00:26:12,040 --> 00:26:16,970
is that a specifying how these vehicles should respond when posed with what looked like ethical dilemmas.

231
00:26:16,970 --> 00:26:23,230
So, say a choice between sacrificing an occupant of the vehicle and sacrificing a

232
00:26:23,230 --> 00:26:28,060
pedestrian or the choice between running over an adult or running over a child.

233
00:26:28,060 --> 00:26:33,430
So one natural response to this problem has been to kind of go out into the world and collect lots of data

234
00:26:33,430 --> 00:26:39,400
about what the public think about how autonomous vehicles should respond to these kinds of situations.

235
00:26:39,400 --> 00:26:44,230
And actually, quite a lot of that work has already been done.

236
00:26:44,230 --> 00:26:49,090
But what's what's not at all obvious is what we should be doing with the data that we're getting out of this

237
00:26:49,090 --> 00:26:55,780
empirical research that the social scientific data about what public the public preferences actually are.

238
00:26:55,780 --> 00:27:01,810
Because it certainly doesn't seem to be the case that we should just kind of unthinkingly implement the public will in this kind of area.

239
00:27:01,810 --> 00:27:10,150
So I suppose in a particular society that most people thought that autonomous vehicles you prioritise saving the lives of white people over others.

240
00:27:10,150 --> 00:27:15,610
Clearly, I wouldn't follow that. That's what we should programme autonomous vehicles to do.

241
00:27:15,610 --> 00:27:23,970
But on the other hand, it does seem plausible that at least some public attitudes should play some role in informing the design of these algorithms.

242
00:27:23,970 --> 00:27:31,060
As I what GM Guy and Chris are thinking about is exactly what that role should be.

243
00:27:31,060 --> 00:27:34,980
And just to kind of cut straight to one of their conclusions. One of their proposals,

244
00:27:34,980 --> 00:27:40,230
probably the most controversial proposal has been that public preferences should be

245
00:27:40,230 --> 00:27:45,580
put through a kind of philosophical filter before being built into the algorithm.

246
00:27:45,580 --> 00:27:51,510
So we should first cheque with these preferences are actually consistent with at least some plausible and

247
00:27:51,510 --> 00:27:57,990
widely held philosophical ethical theories and only preferences that make it through this kind of filter.

248
00:27:57,990 --> 00:28:04,170
What they call laundered preferences should be fed into algorithm design.

249
00:28:04,170 --> 00:28:10,100
OK, so the second strand of our research that I wanted to mention is a programme that I'm leading on

250
00:28:10,100 --> 00:28:16,340
the ethics of predicting and influencing behaviour supported by the welcome trust in the U.S.

251
00:28:16,340 --> 00:28:20,000
So this work has intersected with A.I. ethics and a few places.

252
00:28:20,000 --> 00:28:23,600
One of these is on the topic of crime prediction,

253
00:28:23,600 --> 00:28:30,170
where we've been doing some work with colleagues in the Department of Psychiatry to try to make some, I guess,

254
00:28:30,170 --> 00:28:35,270
practical suggestions about how we could improve the kinds of crime prediction algorithms

255
00:28:35,270 --> 00:28:40,880
that are increasingly being used both in criminal justice and in forensic psychiatrist.

256
00:28:40,880 --> 00:28:51,710
And by improve, I mean here making more accurate, but also mitigating some of the ethical concerns about bias and unfairness.

257
00:28:51,710 --> 00:28:55,820
But actually, the strand of our research that I wanted to go into a little bit more detail

258
00:28:55,820 --> 00:29:00,050
about is some work that we're doing on the idea of a right to mental integrity,

259
00:29:00,050 --> 00:29:07,730
because I think this is an interesting case of an area where existing thinking and medical ethics might have something of relevance to contribute,

260
00:29:07,730 --> 00:29:16,910
to add to ethics. So it's very widely accepted, especially in medical ethics, but also elsewhere,

261
00:29:16,910 --> 00:29:21,800
that we all possess something like a right to bodily integrity, a right against interference with our bodies.

262
00:29:21,800 --> 00:29:27,220
So this right would be infringed if someone physically assaults you, but it would also be infringed.

263
00:29:27,220 --> 00:29:33,420
For example, if a medical professional performed a medical procedure on you without your consent.

264
00:29:33,420 --> 00:29:38,250
So the question that we're interested in is whether there might be an analogies right to mental integrity.

265
00:29:38,250 --> 00:29:43,950
That would be a right against interference with your mind rather than your body.

266
00:29:43,950 --> 00:29:48,570
And I mean, this is a question that hasn't been much discussed either in philosophy or law.

267
00:29:48,570 --> 00:29:53,160
But we think it's going to be very relevant to medical ethics because quite a few medical

268
00:29:53,160 --> 00:29:58,740
interventions look like they might infringe something like a right to mental integrity.

269
00:29:58,740 --> 00:30:01,890
Perhaps the most obvious example here would be the use of compulsory psychiatric

270
00:30:01,890 --> 00:30:06,930
interventions on patients who have been sectioned under the Mental Health Act.

271
00:30:06,930 --> 00:30:10,980
But we think this question about mental integrity is also going to be relevant to

272
00:30:10,980 --> 00:30:17,430
nonmedical interventions and interventions that are not at all physically invasive.

273
00:30:17,430 --> 00:30:25,980
And I think one of the most interesting and important examples here would be what we might call A.I.S assisted manipulation.

274
00:30:25,980 --> 00:30:34,410
So supplies and online platforms that something like Facebook develops and deploys an artificial intelligence that can identify the psychological

275
00:30:34,410 --> 00:30:40,830
weaknesses of all of its users and almost perfectly target them with with

276
00:30:40,830 --> 00:30:45,480
content that will maximise the length of time that they spend on the platform.

277
00:30:45,480 --> 00:30:51,530
And as sort of maximally strengthen the desire to keep habitually checking that platform.

278
00:30:51,530 --> 00:30:57,480
It seems reasonable to ask whether we could think of this technology as infringing a rights to mental integrity,

279
00:30:57,480 --> 00:31:01,760
and we think that the answer to this question is going to depend on exactly how you understand that.

280
00:31:01,760 --> 00:31:05,780
Right. What kind of philosophical basis you think that it has.

281
00:31:05,780 --> 00:31:12,110
But at least on some plausible accounts of the right to mental integrity that we're considering.

282
00:31:12,110 --> 00:31:18,650
It seems that that A.I. based manipulation could actually infringe the right to mental integrity in just the same way as,

283
00:31:18,650 --> 00:31:21,590
for example, compulsory psychiatric interventions,

284
00:31:21,590 --> 00:31:30,440
which might suggest that we should be regulating a basement manipulation and other similar forms of manipulation much more stringently,

285
00:31:30,440 --> 00:31:36,620
perhaps a bit more closely to how we currently regulate psychiatry.

286
00:31:36,620 --> 00:31:41,590
The third and final example that I wanted to say something about is the work of Hannah Maslova

287
00:31:41,590 --> 00:31:48,650
and Steven Rainey on ethical issues raised by the use of neuro prosthetics for decoding speech.

288
00:31:48,650 --> 00:31:53,870
So this is this is the philosophical part of a large multidisciplinary project called brain coding,

289
00:31:53,870 --> 00:32:03,200
which is which is seeking to develop neuro prostheses or brain computer interfaces that could allow individuals who have lost the ability to speak,

290
00:32:03,200 --> 00:32:05,210
to communicate.

291
00:32:05,210 --> 00:32:14,660
So these devices would work by detecting brain signals, converting them into synthesised speech with the mediation of an eye language model.

292
00:32:14,660 --> 00:32:19,370
And that language model would do a significant amount of predicting and rephrasing

293
00:32:19,370 --> 00:32:23,840
in order to allow the speaker to speak at a kind of ordinary conversational pace.

294
00:32:23,840 --> 00:32:32,570
And with more or less ordinary fluency. So one of the ethical issues here concerns the extent to which we could hold people

295
00:32:32,570 --> 00:32:37,870
responsible for the utterances that they might make via a device like this.

296
00:32:37,870 --> 00:32:40,920
So ordinarily, we do hold people responsible for their speech acts.

297
00:32:40,920 --> 00:32:45,670
If someone says something racist or offensive, we tend to think that they can be blamed for that.

298
00:32:45,670 --> 00:32:49,740
And that seems to presuppose that they're responsible for what they've said.

299
00:32:49,740 --> 00:32:54,030
If I promise to do something, you'll probably think that I'm bound by that promise.

300
00:32:54,030 --> 00:33:00,210
And again, that seems to presuppose that I'm responsible for what I said when I made the promise.

301
00:33:00,210 --> 00:33:09,270
But one question is, to what extent would those kinds of responsibility assignments carry over to cases involving neuro prostheses of this sort?

302
00:33:09,270 --> 00:33:14,790
Given that in these cases, there might be a much more significant gap between the kind of mental act of intending

303
00:33:14,790 --> 00:33:19,860
or attempting to say something and the actual utterance that gets produced.

304
00:33:19,860 --> 00:33:28,470
So how does the mediation of an artificial intelligence in this kind of system affect the responsibility of the speaker for the for the utterance?

305
00:33:28,470 --> 00:33:36,060
How might the nature of A.I. make a difference to Assignment's response? But these are the kinds of questions that Stephen and Henare are addressing.

306
00:33:36,060 --> 00:33:41,220
And again, to kind of briefly cut to one of the conclusions, they're arguing that many of these cases,

307
00:33:41,220 --> 00:33:47,160
the mediation of artificial intelligence, could significantly diminish the responsibility of the speaker.

308
00:33:47,160 --> 00:33:56,220
And in a way that might require us this sort of significantly, I think, some of our ethical norms regarding speech and conversation.

309
00:33:56,220 --> 00:34:00,960
So those are some of the relevant areas on which the hero centre is working.

310
00:34:00,960 --> 00:34:05,490
There are others as well. Mike Robillard, a military ethicist in the centre,

311
00:34:05,490 --> 00:34:11,630
has been working on critiquing some of the existing debate about autonomous weapon systems or killer robots.

312
00:34:11,630 --> 00:34:20,340
And one of the DEHL students appreciate Mishra is working on the extent to which we might need to revise some of our concepts,

313
00:34:20,340 --> 00:34:28,620
like the standard of care and medical negligence in a world in which healthcare professionals are relying heavily on machine learning classifiers.

314
00:34:28,620 --> 00:34:31,020
But I'm going to stop there and hand over to accuracy.

315
00:34:31,020 --> 00:34:36,780
He's going to go into a bit more detail about some of the visual ways in which philosophy can kind of make

316
00:34:36,780 --> 00:34:49,270
progress on the types of topics that I've been talking about and also to tell you about some of her research.

317
00:34:49,270 --> 00:34:53,860
Thanks very much indeed, Tom. As you see, lots going on, and that's just the new hero.

318
00:34:53,860 --> 00:35:03,040
Senator Kerissa is a research fellow at the you're hearing senator and also at the Welcome Centre for Ethics and Humanities.

319
00:35:03,040 --> 00:35:08,470
So yet another senator we have in Oxford and is a member of Christchurch.

320
00:35:08,470 --> 00:35:12,880
He works on digital ethics and moral and political philosophy more generally.

321
00:35:12,880 --> 00:35:18,730
And she's the editor of the forthcoming Oxford Handbook of Digital Ethics, Overdue Gracen.

322
00:35:18,730 --> 00:35:23,650
Thank you, Peter. It's great to be here. Thank you all for coming.

323
00:35:23,650 --> 00:35:31,860
I'm going to talk a little bit about the place of philosophy in the ethics of a I have philosophy can contribute to this field.

324
00:35:31,860 --> 00:35:37,260
Practical ethics was developed in the 1970s, in the 1940s and 50s.

325
00:35:37,260 --> 00:35:44,020
It was understood that ethics didn't have much to do except analyse moral language,

326
00:35:44,020 --> 00:35:49,050
analyse the meaning of terms like good and rights and obligatory things like that,

327
00:35:49,050 --> 00:35:56,550
and maybe explored the tooth conditions for which a particular utterance was true or false.

328
00:35:56,550 --> 00:36:03,720
And then in the 1970s, I learnt social movements happened that put pressure on the discipline both externally and internally,

329
00:36:03,720 --> 00:36:08,250
and that made philosophy engage more directly with political issues.

330
00:36:08,250 --> 00:36:13,150
The Vietnam War was an important thing that happened.

331
00:36:13,150 --> 00:36:25,770
But philosophers started to think about feminism, questions about abortion and sexual orientation and discriminations on.

332
00:36:25,770 --> 00:36:29,880
This really changes discipline first, a lot of students became very,

333
00:36:29,880 --> 00:36:36,780
very much involved in philosophy and sign up to philosophy courses at a time when philosophy was kind of losing steam.

334
00:36:36,780 --> 00:36:45,360
And whenever I. I read about these moments in Lost for History, I always had a bit of jealousy in part.

335
00:36:45,360 --> 00:36:48,270
I just came to philosophy 30 or 40 years later.

336
00:36:48,270 --> 00:36:54,930
You know, how interesting and how important the change discipline in a way that can contribute positively to the world.

337
00:36:54,930 --> 00:37:00,810
And one moment that was particularly important was in 1972,

338
00:37:00,810 --> 00:37:08,410
The New York Times came out with a story about the Tuskegee experiment and the Tuskegee scandal.

339
00:37:08,410 --> 00:37:17,010
And this had been an experiment that carried out on for 40 years in which people who had syphilis were observed and they weren't treated.

340
00:37:17,010 --> 00:37:21,210
Even though treatment was available, there was a huge scandal.

341
00:37:21,210 --> 00:37:26,810
We pushed off the discipline of medical ethics forward.

342
00:37:26,810 --> 00:37:31,770
And I think in many ways we are in a similar situation when in the 1970s,

343
00:37:31,770 --> 00:37:36,660
we have new technologies that are facing us with new problems that we haven't faced before.

344
00:37:36,660 --> 00:37:42,840
And that computer scientists are not particularly well-trained to think about these things.

345
00:37:42,840 --> 00:37:51,570
And we have new scandals on the camera and a legal scandal, amongst others, that we make tangible that the need for for ethics.

346
00:37:51,570 --> 00:38:01,110
So practical ethics more generally. Tries to come up with ways to apply theory into practise.

347
00:38:01,110 --> 00:38:04,650
And so there's a question of, well. OK. So that's a practical philosophy.

348
00:38:04,650 --> 00:38:08,580
Does more, more or less. But how can philosophy contribute to ethics in particular?

349
00:38:08,580 --> 00:38:16,070
What is special about the philosopher and what is special about the philosopher in the context of the ethics of a I?

350
00:38:16,070 --> 00:38:20,970
And this is just the latest version of a question that has been bothering

351
00:38:20,970 --> 00:38:25,710
philosophers for a long time and many people have given different answers to.

352
00:38:25,710 --> 00:38:35,500
This is just a small sample. And although there are many controversies within within this debate, maybe the most important of which is.

353
00:38:35,500 --> 00:38:38,920
Are there more authorities and what do we mean by more authorities?

354
00:38:38,920 --> 00:38:47,350
There's much more consensus and disagreement within the debate and perhaps the most important points of consensus is that philosophy can

355
00:38:47,350 --> 00:38:55,990
offer a conceptual analysis in the hope of leading to better decisions and also better justifying decisions that have already been made,

356
00:38:55,990 --> 00:39:04,740
especially to those who lose out in that decision and with the hope that consumption rises, can make debates sharper.

357
00:39:04,740 --> 00:39:10,660
It can mitigate shorter. And it can make it less income to suit.

358
00:39:10,660 --> 00:39:16,990
So what is conceptual analysis and what kinds of things does it impute? It includes things like clarifying concepts.

359
00:39:16,990 --> 00:39:22,090
Sometimes people are fighting about something and they're not even talking about the same thing.

360
00:39:22,090 --> 00:39:26,770
On occasion, making sure that people are talking about the same thing even leads to solving problems.

361
00:39:26,770 --> 00:39:32,320
This is something that we decided defended. Of course, that's not always the case.

362
00:39:32,320 --> 00:39:37,160
Sometimes philosophy can provide nuance like any other discipline.

363
00:39:37,160 --> 00:39:44,200
Ethics have developed a very technical language that can be much more nuanced than just ordinary moral language.

364
00:39:44,200 --> 00:39:48,490
So it's not only about right and wrong, it's about what's permissible, what's impermissible,

365
00:39:48,490 --> 00:39:54,470
what's obligatory, what might be Suburgatory above and beyond duty and so on.

366
00:39:54,470 --> 00:40:02,030
It's also about working out the implications of use, some views might be might feel very attractive or be very attractive in a first glance,

367
00:40:02,030 --> 00:40:05,690
and then you start working out what the implications,

368
00:40:05,690 --> 00:40:11,360
either practical implications or theoretical implications and certainly doesn't seem that attractive anymore.

369
00:40:11,360 --> 00:40:16,090
A good example is how your personal data has been treated.

370
00:40:16,090 --> 00:40:20,860
Some some people think that we should treat personal data as property. And that sounds quite intuitive,

371
00:40:20,860 --> 00:40:25,040
except when you start looking at the implications and how poverty differs from

372
00:40:25,040 --> 00:40:29,490
personal data and suddenly it doesn't seem like such a good idea anymore.

373
00:40:29,490 --> 00:40:38,490
Pointing out contradictions of this cause is full of contradictions and fallacies both in the media but also in parliament and everywhere in between.

374
00:40:38,490 --> 00:40:44,730
And flusters can point out those distinguishing questions of fact and questions of value.

375
00:40:44,730 --> 00:40:51,840
It's not always obvious and it's not always easy. For instance, in the 1960s, we thought that death was just a biological question,

376
00:40:51,840 --> 00:40:56,040
a medical question, whether somebody was dead or not, just for the doctor to decide.

377
00:40:56,040 --> 00:41:00,390
Suddenly, with mechanical ventilators, we realised, well, you know,

378
00:41:00,390 --> 00:41:06,190
we have these bodies that are warm, their hearts are beating, but their brains seem to be destroyed.

379
00:41:06,190 --> 00:41:12,500
And we can harvest their organs either alive or they're dead. And suddenly the philosopher comes and says, well, what do you mean by that?

380
00:41:12,500 --> 00:41:16,080
Is it the death of the body, the death of consciousness, the death of the person,

381
00:41:16,080 --> 00:41:23,300
the death of the interests and rights that typically attached to people?

382
00:41:23,300 --> 00:41:28,320
I'm finding providing theory and in practical ethics.

383
00:41:28,320 --> 00:41:33,510
Of course, there's a lot of theory that comes both from normative ethics, metaphysics and so on.

384
00:41:33,510 --> 00:41:41,280
And in the property, in the course of application, many times we realise the limits and possible mistakes of the theory itself.

385
00:41:41,280 --> 00:41:48,410
So that. Practical cases and also empirical facts inform the theory and change it and polish it.

386
00:41:48,410 --> 00:41:53,520
And this is an interesting process because, you know, philosophy has a bad reputation for a lot of disagreements.

387
00:41:53,520 --> 00:41:55,750
There are not a lot of progress in consensus.

388
00:41:55,750 --> 00:42:03,490
But in fact, when you study in detail of the the history of philosophy and theories get very much polished.

389
00:42:03,490 --> 00:42:08,350
First, there are consensus on some things that they used to that there was disagreement in the past.

390
00:42:08,350 --> 00:42:17,260
But even when there is disagreement, for example, consequentialist today is a much more sophisticated theory than it was in its origins.

391
00:42:17,260 --> 00:42:24,850
And partly theories get polished through bumping up with reality and taking a look at practical cases.

392
00:42:24,850 --> 00:42:28,510
Secondly, it's really important for philosophy to identify moral problems.

393
00:42:28,510 --> 00:42:31,780
And then again, some of these are not as obvious as you might think.

394
00:42:31,780 --> 00:42:39,400
So before bioethics came along, doctors engaged in all sorts of problematic practises that one, seen as a problem, weren't seen as problematic.

395
00:42:39,400 --> 00:42:47,980
So, for example, is not informing patients on a diagnoses randomising patients to treatment or placebo without informing them that they were

396
00:42:47,980 --> 00:42:59,160
part of a research or even conducting very invasive examinations like rectifies examination in patients who were unconscious.

397
00:42:59,160 --> 00:43:06,170
And this one seen as, you know, doctors having bad intentions or anything like it was just an.

398
00:43:06,170 --> 00:43:13,930
They love. Their philosophy can inspire moral thought.

399
00:43:13,930 --> 00:43:26,020
Through arguments on thought, experiments and analogies, we can raise moral thought about prejudice and invite people to consider certain situations.

400
00:43:26,020 --> 00:43:33,260
And in so doing, stimulate their their moral thinking and also challenge moral intuitions.

401
00:43:33,260 --> 00:43:40,870
And here there's a very important role for public philosophy and for public and engaging with the public in general.

402
00:43:40,870 --> 00:43:49,300
And then finally, philosophy can provide experience. Good ethicists have extensive experience tackling difficult issues.

403
00:43:49,300 --> 00:43:58,900
And it would be a waste to not involve such so much knowledge in a time when there's so much at stake.

404
00:43:58,900 --> 00:44:06,010
Of course, a person doesn't have to have a pay student athletes or have published in the best ethical journals to have good ethical insight.

405
00:44:06,010 --> 00:44:12,640
But spending also your hours of most of your days thinking about ethical questions, trying different methodologies,

406
00:44:12,640 --> 00:44:19,540
learning about past these pitfalls does provide some kind of experience that can that can be of use.

407
00:44:19,540 --> 00:44:24,430
And in this sense, private life is his most experience with medical ethics, of course.

408
00:44:24,430 --> 00:44:31,300
And I think there's a lot to learn from this analogy that still hasn't been worked out,

409
00:44:31,300 --> 00:44:36,100
both from the similarities in these two fields and the differences.

410
00:44:36,100 --> 00:44:43,900
I think digital ethics is much more political and medical ethics just cite one difference and that makes it is going to make it very, very different.

411
00:44:43,900 --> 00:44:48,340
And also, we have a lot to learn from successes, but also from failures.

412
00:44:48,340 --> 00:44:55,210
I mean, one of the biggest failures, in my view of medical ethics is how it hasn't been able to regulate Big Pharma property.

413
00:44:55,210 --> 00:45:02,470
So Big Pharma today gets away with, for instance, carrying out 100 experiments to prove that a drug works.

414
00:45:02,470 --> 00:45:08,560
99 of them show that the drug doesn't work. One shows that the drug might work and that one gets published.

415
00:45:08,560 --> 00:45:15,190
The 99 don't get published. And sometimes researchers can't even talk about it.

416
00:45:15,190 --> 00:45:22,630
This is a huge failure. And it sort of signals a challenge that we have with regulating industry because most are a good part

417
00:45:22,630 --> 00:45:27,310
of the research carried on an A.I. right now is carried out by industry and not in universities.

418
00:45:27,310 --> 00:45:34,000
And so this is a huge challenge for digital ethics to detect. So just to finish.

419
00:45:34,000 --> 00:45:40,270
I'm involved in the following research projects. I'm thinking about what digital effects can learn from medical ethics.

420
00:45:40,270 --> 00:45:45,320
I'm finishing a book about the privacy. I'm working on the ethics of prediction.

421
00:45:45,320 --> 00:45:48,750
Human beings have been using predictions since the Oracle of Delphi.

422
00:45:48,750 --> 00:45:54,830
And strangely enough, we haven't thought much about the ethics of prediction and what makes a prediction ethical.

423
00:45:54,830 --> 00:46:00,710
And most importantly, the thing that I'm most excited about is editing the Oxford Handbook of Digital Ethics.

424
00:46:00,710 --> 00:46:04,160
Many philosophers in the room are going to participate in that. Which is great.

425
00:46:04,160 --> 00:46:10,750
And it's going to cover all sorts of things from sex and friendship in the digital age to democracy.

426
00:46:10,750 --> 00:46:18,350
Does fear the use of killer robots, surveillance, privacy and so much more.

427
00:46:18,350 --> 00:46:25,100
So it's very exciting. This is a very exciting time. And I'm no longer jealous of the practical emphasis of the 1970s.

428
00:46:25,100 --> 00:46:32,480
I think this is so much better. And one day in 10, 15 years, when the institute has had enormous impact on the ethics of AI,

429
00:46:32,480 --> 00:46:46,060
each of us will be able to say that we were here the first day when we said.

430
00:46:46,060 --> 00:46:50,530
Hearing you talk about the old timers that they're interested in ethics,

431
00:46:50,530 --> 00:47:02,860
I noticed that two of those papers that you were highlighting were actually in a volume that I coedited back in 1992.

432
00:47:02,860 --> 00:47:06,700
Thank you very much. Both Tom Charissa, your hero, senator.

433
00:47:06,700 --> 00:47:11,140
Doing lots of interesting work. You'll get a chance to catch up later.

434
00:47:11,140 --> 00:47:22,090
Over over nibbles and drinks. One innovation that we're bringing into these seminars, we won't get to have use it very much this time.

435
00:47:22,090 --> 00:47:29,920
But our intention is, is to make it a general feature is to enable audience feedback through questions.

436
00:47:29,920 --> 00:47:35,600
Now, we're planning at the very end this time to have a session on questions.

437
00:47:35,600 --> 00:47:41,710
If you want to suggest a question that you'd like to have discussed by the speakers.

438
00:47:41,710 --> 00:47:51,420
If you go to Sligo dot com, that's s l i d o Sligo dot com.

439
00:47:51,420 --> 00:48:01,640
And. Type in A.I. ethics, one word as this session to log into, and if you submit a question there,

440
00:48:01,640 --> 00:48:05,690
then that question will magically appear on my device here.

441
00:48:05,690 --> 00:48:11,270
And you in the audience can vote for the question or questions that you'd most like to have answered.

442
00:48:11,270 --> 00:48:19,250
So at the end, once the speakers that have all spoken, the intention is to have a short question and answer session.

443
00:48:19,250 --> 00:48:25,650
Then I hope that will be interesting and it is at least an interesting experiment.

444
00:48:25,650 --> 00:48:33,180
Okay, now we're moving from the hero centre to the Oxford Internet Institute, and Victoria Nash is our first speaker.

445
00:48:33,180 --> 00:48:38,460
She's deputy director, associate professor and senior policy fellow there.

446
00:48:38,460 --> 00:48:46,530
Her research focuses on normative and practical challenges of regulating online behaviour with a particular focus on child protection,

447
00:48:46,530 --> 00:48:48,960
content, moderation and human rights.

448
00:48:48,960 --> 00:49:02,000
And Vicki is just going to give a quick introduction to the Internet Institute before handing over to Sandra and Brian.

449
00:49:02,000 --> 00:49:06,410
Coming up, speak. I am reminded of when my rule never I should never offer slides if I speak for

450
00:49:06,410 --> 00:49:14,480
less than five minutes because I would probably spend five minutes finding it.

451
00:49:14,480 --> 00:49:17,840
Thank you very much for the introduction. Thank you to the previous speakers.

452
00:49:17,840 --> 00:49:22,730
As a political theorist who did her defo here many years ago, she would have loved to have your presentation.

453
00:49:22,730 --> 00:49:24,460
That right beginning of my page to experts,

454
00:49:24,460 --> 00:49:30,650
to the whole three years to figure out how to explain to her why last week mattered, why my work is so crucially important.

455
00:49:30,650 --> 00:49:36,860
So I those pieces fit within just a couple of minutes of explanation about what the oil is.

456
00:49:36,860 --> 00:49:43,870
And then I'll hand over to my colleagues. We will take down a great depth to give an indication of what type of research we do.

457
00:49:43,870 --> 00:49:47,600
For those of you that don't know, we're relatively new. I love it.

458
00:49:47,600 --> 00:49:51,980
Every time we do have a new centre, it's too creative because I feel like an old hand.

459
00:49:51,980 --> 00:49:58,820
We've been around since 2001. We were set up to be a multidisciplinary centre based within social sciences,

460
00:49:58,820 --> 00:50:05,330
but multidisciplinary focussing on the societal implications of digital technologies.

461
00:50:05,330 --> 00:50:11,610
It's funny looking back, actually. I mean, the question that we are asking now are many in many ways the same questions that you're asking in 2001,

462
00:50:11,610 --> 00:50:15,710
but perhaps applied to new and innovative technologies.

463
00:50:15,710 --> 00:50:20,810
But I think that, you know, so much of what the new institute is going to do is is hugely exciting to us.

464
00:50:20,810 --> 00:50:23,300
And I'm frankly very delighted to see, you know,

465
00:50:23,300 --> 00:50:29,270
significant investment in this field and new opportunities for us to collaborate both with the humanities, but actually across the way.

466
00:50:29,270 --> 00:50:36,380
University is the chart that Niger presented showed us. So, yes, congratulations on those that helped make this happen.

467
00:50:36,380 --> 00:50:41,570
I work in this area. I would say a sort of maybe threefold.

468
00:50:41,570 --> 00:50:44,900
The first is, if you like, in the very broadest sense.

469
00:50:44,900 --> 00:50:48,710
So what you want an Nigel covered at the very beginning in the sort of, you know,

470
00:50:48,710 --> 00:50:52,580
the array of different topics and areas is just fundamental to what we do.

471
00:50:52,580 --> 00:50:55,070
So, you know, big, broad questions about.

472
00:50:55,070 --> 00:51:05,480
About innovation, about the development of new tools, new ways of using artificial intelligence to develop new products, new forms and functions.

473
00:51:05,480 --> 00:51:11,440
Are those in everyday life questions about what this means for how we regulate and govern these technologies?

474
00:51:11,440 --> 00:51:18,360
This this is just our everyday business and it's something that that all of my 90 50 faculty are frankly engaged in.

475
00:51:18,360 --> 00:51:24,990
However, there's a much narrower sentence in which actually quite the questions of ethics and I arise in a

476
00:51:24,990 --> 00:51:30,330
number of very specific sort of research has portfolios to which you're going to hear about tonight.

477
00:51:30,330 --> 00:51:36,090
Since 2014, I'm pleased to note that we've had in-house philosophies.

478
00:51:36,090 --> 00:51:38,580
We've had I've in the very early days of the.

479
00:51:38,580 --> 00:51:47,700
Let me John, if you're ready with his work around information, ethics, addressing questions like sort of moral, artificial,

480
00:51:47,700 --> 00:51:54,470
Egil, evil or the morality of artificial agents that that's been with us is safe runabouts of five, six years.

481
00:51:54,470 --> 00:51:57,940
But actually, we don't just have lots of both of these questions now.

482
00:51:57,940 --> 00:52:02,810
So as we've brought in several philosophers on top of each on it, what you're going to hear from shortly.

483
00:52:02,810 --> 00:52:06,240
But we also now have lawyers like Sandra who's going to speak to you.

484
00:52:06,240 --> 00:52:07,740
We have political scientists.

485
00:52:07,740 --> 00:52:15,350
We have data scientists again, though, whose work is fundamentally concerned with questions of ethics and I of writ large.

486
00:52:15,350 --> 00:52:23,960
And in those cases, it covers issues such as things like what constitutes fairness or unfairness, discrimination, if you like,

487
00:52:23,960 --> 00:52:30,150
in the uses of I brought political questions about once you once you maybe identified principles of fairness,

488
00:52:30,150 --> 00:52:35,320
how you might go about regulating for those or holding companies to account.

489
00:52:35,320 --> 00:52:37,930
We cover applied questions such as, again,

490
00:52:37,930 --> 00:52:45,580
if you if you identify unfair practises or sort of sources of lack of diversity in data collection, what does that mean for innovation?

491
00:52:45,580 --> 00:52:48,940
What does that mean for data collection? What does it mean for privacy?

492
00:52:48,940 --> 00:52:54,910
So if you like the fact that a multi disciplinary means that we aren't just approaching this, certainly from a philosophical angle,

493
00:52:54,910 --> 00:53:02,050
but we are taking some of these sort of core philosophical questions and playing them out across different disciplines and different topics,

494
00:53:02,050 --> 00:53:07,750
we hope, all with the aim of improving not just research outputs, research,

495
00:53:07,750 --> 00:53:13,820
understanding these issues, but also, if you like, societal understanding, societal practises.

496
00:53:13,820 --> 00:53:19,610
So if we have a sort of a broad focus, this very narrow focus in specific projects, again, which we'll hear about shortly.

497
00:53:19,610 --> 00:53:23,390
I just want to flag up. But we also have a focus on this in our teaching.

498
00:53:23,390 --> 00:53:27,430
We have full graduate degrees, two masters. One is social data science.

499
00:53:27,430 --> 00:53:31,570
One, social science, the Internet and two corresponding defo programmes.

500
00:53:31,570 --> 00:53:36,560
And in each of those, again, these broad questions around ethics and I arrive in different places.

501
00:53:36,560 --> 00:53:41,930
We actually have got a pure philosophy paper, for example, in one of our master's degrees.

502
00:53:41,930 --> 00:53:49,070
In another, we we embrace the core questions about what things like fairness and transparency would look like in the practise of social data science.

503
00:53:49,070 --> 00:53:56,570
So again, I really applaud, I think, the desire to ensure that the work of the new institute will also include content,

504
00:53:56,570 --> 00:54:03,080
fair fit for new courses, new option papers. Certainly we find it immensely satisfying experience to deliver those to our students.

505
00:54:03,080 --> 00:54:06,920
I'm happy to take questions later on on what the eyes, the eyes doing in this area.

506
00:54:06,920 --> 00:54:18,480
But it's my great pleasure to hand over to my. Such as?

507
00:54:18,480 --> 00:54:25,940
So next up is Sandra Baktir, who's associate professor and senior research fellow in law and ethics of A.I.,

508
00:54:25,940 --> 00:54:31,170
Big Data and robotics, as well as Internet regulation at the AIS.

509
00:54:31,170 --> 00:54:37,770
We've heard her current research focuses on the legal and ethical implications of a high profile,

510
00:54:37,770 --> 00:54:43,590
including profiling, explaining lendable A.I. and algorithmic bias.

511
00:54:43,590 --> 00:54:45,940
Thank you so much for this introduction.

512
00:54:45,940 --> 00:54:54,870
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.

513
00:54:54,870 --> 00:55:02,210
And as Bikila mentioned, I am a lawyer, so I usually look at things from a legal perspective and.

514
00:55:02,210 --> 00:55:07,820
But what I want to do is actually talk about free aspects where I think that the eyes disrupting

515
00:55:07,820 --> 00:55:14,420
legal concepts and I think the only way to fix that problem is actually via this mega approach.

516
00:55:14,420 --> 00:55:18,050
So I'm actually arguing that the law might not be fit for purpose,

517
00:55:18,050 --> 00:55:25,940
which is why we need ethicists and mulches to work to get it to actually give new strength to the law that we currently have.

518
00:55:25,940 --> 00:55:30,620
And yeah, I want to give you three examples where I think we need to focus a bit more.

519
00:55:30,620 --> 00:55:36,470
And one of the areas that I'm focussing on as well. One is accountability, one to Spanish and one is privacy.

520
00:55:36,470 --> 00:55:44,960
So I start with accountability. I've chosen the example of loan applications and how we have done this previously in the past.

521
00:55:44,960 --> 00:55:53,150
Usually when you go into a human human setting, if you go to a bank and you apply for a loan, the loan officer will ask you a couple of questions.

522
00:55:53,150 --> 00:56:00,410
They might ask you what your income is. If you have any savings, your employment history, all this questions make sense.

523
00:56:00,410 --> 00:56:06,930
There is an intuitive link where a person ask you about your financial status in order to find out if you've gone before the loan or not.

524
00:56:06,930 --> 00:56:14,180
There's an intuitive connexion there. What we see now is they've been moving away from those traditional data sources to make decisions.

525
00:56:14,180 --> 00:56:16,250
We use our traditional data sources.

526
00:56:16,250 --> 00:56:24,140
So a couple of examples here where credit institutions are now using Facebook profiles to decide if somebody should get at them.

527
00:56:24,140 --> 00:56:29,780
So they look at you, profile pictures, the friends that you have on Facebook, the groups that you join,

528
00:56:29,780 --> 00:56:36,840
the things that you post, the things that you like, and they infer if you're reliable borrower, not.

529
00:56:36,840 --> 00:56:43,770
Similar things happen in the insurance space where also, for example, profile social profile,

530
00:56:43,770 --> 00:56:49,920
that network information is being used to define if somebody should get insurance and what premiums should get.

531
00:56:49,920 --> 00:56:58,320
So but that's not the only thing. In general, whenever we make decisions, we start to use very untraditional data.

532
00:56:58,320 --> 00:57:05,940
That includes clicking behaviour, typing behaviour, geo location, ice and ice tracking,

533
00:57:05,940 --> 00:57:10,170
all of that to infer certain things about you, things that can be very privacy,

534
00:57:10,170 --> 00:57:17,400
invasive and unintentional and very counter-intuitive, where I don't really understand all of my data says above me,

535
00:57:17,400 --> 00:57:22,460
I have no idea how my browsing behaviour will affect my credit rating in the future.

536
00:57:22,460 --> 00:57:28,490
And this information can be and is rare that replicates it and and shared with a lot of people and not just in.

537
00:57:28,490 --> 00:57:34,070
Financial services, also important decisions, has to go to jail. Who can go to university?

538
00:57:34,070 --> 00:57:38,480
If you get hired, fired, promoted. All of this information of big data is being used for that.

539
00:57:38,480 --> 00:57:41,240
And that pose questions for accountability.

540
00:57:41,240 --> 00:57:47,000
Because if in the future, if I don't get a loan or don't get a job, the first thing that comes to mind is to ask why?

541
00:57:47,000 --> 00:57:50,930
Why did I not get a job? What happened? I want to have an explanation.

542
00:57:50,930 --> 00:57:55,370
And this is why I started thinking about this topic and look at it at first from a legal

543
00:57:55,370 --> 00:58:00,440
perspective to figure out if we do have a right that Ogwyn physicians have been explained to you,

544
00:58:00,440 --> 00:58:04,280
because that would make a lot of sense. And just looking at from the legal perspective,

545
00:58:04,280 --> 00:58:09,950
what I found is that I don't think we actually have a legally binding right that things have been explained to you.

546
00:58:09,950 --> 00:58:15,240
I was not very happy with the outcome of my paper. It's very hard to say from my academic perspective to actually admit to that.

547
00:58:15,240 --> 00:58:17,810
But it wasn't very happy with the outcome of the paper.

548
00:58:17,810 --> 00:58:22,820
But this was a starting point where I started to think about I don't actually care so much about what the law says.

549
00:58:22,820 --> 00:58:27,170
Just because something is not legally required doesn't mean it's not ethically desired.

550
00:58:27,170 --> 00:58:32,990
Which was the reason why Grant and Chris Russum and I got together and we now have

551
00:58:32,990 --> 00:58:36,950
a research programme at which is called The Governance of Emerging Technologies.

552
00:58:36,950 --> 00:58:44,240
So bretonneux that we'll be talking a couple of minutes is an ethicist and Chris Russell is a machine learning expert.

553
00:58:44,240 --> 00:58:52,790
So, yes, what basically starts like a joke in the sense that a lawyer, a philosopher and a scientist walk into a bar, which we did,

554
00:58:52,790 --> 00:58:59,480
and we spent three hours screaming at each other because we wanted to figure out what a good explanation would actually look like.

555
00:58:59,480 --> 00:59:02,730
What do we think of good explanations? Because we set the legal question of why.

556
00:59:02,730 --> 00:59:08,500
I wanted to know what is it that you want to know? So Brand was very much interested in the trust side of things.

557
00:59:08,500 --> 00:59:15,530
For him, it was very important to figure out what is justified, to believe what a good argument is, what make arguments valid.

558
00:59:15,530 --> 00:59:21,470
And I didn't care about that at all. I went to a justice and accountability. I wanted to be able to contest the decision.

559
00:59:21,470 --> 00:59:25,760
I'm not happy with it. And Curtis, a computer scientist, didn't care about each of those things.

560
00:59:25,760 --> 00:59:32,780
Just want to buck his coat and said, what's going on in the black box? So even though we were all very, very passionate about explanations,

561
00:59:32,780 --> 00:59:40,070
we started to understand that we think very different things about that and see the explanations in very different ways.

562
00:59:40,070 --> 00:59:44,300
But we finally found the solution and actually wrote a paper together and which

563
00:59:44,300 --> 00:59:50,990
has come come to factual explanations to try to reconcile those approaches.

564
00:59:50,990 --> 00:59:55,010
And our method actually allows you to understand a little bit of what's going on.

565
00:59:55,010 --> 01:00:01,670
The black box also gives you more trust in a system and gives you accountability because you would be able to,

566
01:00:01,670 --> 01:00:09,190
for example, contested decision, not happy with it. So the exciting part of that is that it actually got traction.

567
01:00:09,190 --> 01:00:11,900
And Google has implemented our our method.

568
01:00:11,900 --> 01:00:18,290
Now, last year in intensive LOSO nine can actually play with logarithms and understand what's going on in a black box.

569
01:00:18,290 --> 01:00:26,700
And Google is not the only one who is very fond of our idea because IBM, Microsoft and FLOC also implemented our our idea.

570
01:00:26,700 --> 01:00:32,310
So what that means is if you scream a lot to each other, it actually pays off at some point.

571
01:00:32,310 --> 01:00:40,850
But we also learnt that explanations are only one facet of true accountability, because explanations are not justifications.

572
01:00:40,850 --> 01:00:49,080
I can tell you, for example, I'm not giving you the loan because I don't like you face explanations.

573
01:00:49,080 --> 01:00:58,040
It doesn't mean it's justified. So actually, what you also want to look at is at the inferences to predictions and the opinions that algorithms out

574
01:00:58,040 --> 01:01:03,770
about you write all that big data is being collected about you and very sensitive things can be inferred.

575
01:01:03,770 --> 01:01:08,870
For example, if you're pregnant, if you're alive of worry, or if you could work, if you should get promotion,

576
01:01:08,870 --> 01:01:14,450
if you have undiagnosed disorders, all of that information can be very progress, invasive.

577
01:01:14,450 --> 01:01:20,930
And that is the actual thing that you're concerned with. So, again, I look at it from my eye, from a legal perspective,

578
01:01:20,930 --> 01:01:25,670
and try to look at data protection law because, you know, this is very close to your private life.

579
01:01:25,670 --> 01:01:30,530
So having friends is one example of personal data. You would get a lot of data protection rights.

580
01:01:30,530 --> 01:01:39,920
And again, what I found is that I actually disrupts the law in a very and traditional way where we have to think creatively about those problems.

581
01:01:39,920 --> 01:01:48,250
So, Brandon, I wrote a paper together which is called A Rights to Reasonable Inferences, Rethinking their protection law and Age Data.

582
01:01:48,250 --> 01:01:56,000
Any are actually calling for new standards because we found that the law, as it currently stands, is not good enough to protect us.

583
01:01:56,000 --> 01:02:00,720
As you can see, it's 130 pages trying to make the point that law is not good enough.

584
01:02:00,720 --> 01:02:06,890
It was actually planned to be a hundred and sixty, but read threatened if they were to work with me again if I don't stop writing.

585
01:02:06,890 --> 01:02:13,880
So it's just 30 pages. But it's very important to point out that we found a lot of loopholes in the law,

586
01:02:13,880 --> 01:02:18,500
some of which are that we don't even know if inferences of personal data.

587
01:02:18,500 --> 01:02:22,050
So all the assumptions that algorithms are making are being made above.

588
01:02:22,050 --> 01:02:27,060
It's not clear if it actually wasn't the PR, even if they do fall on the data protection law.

589
01:02:27,060 --> 01:02:32,400
What the purpose of data protection law is, is not to regulate how we make decisions about you.

590
01:02:32,400 --> 01:02:36,540
So if you have a problem over how you're being evaluated, how you've been seen,

591
01:02:36,540 --> 01:02:40,490
how you assess, you would need to find another law that helps you do that.

592
01:02:40,490 --> 01:02:46,500
Data protection law doesn't give you a remedy over how are you being seen by algorithms.

593
01:02:46,500 --> 01:02:52,860
And that is quite problematic because as we've already mentioned very often, we don't have standards for accurate,

594
01:02:52,860 --> 01:03:00,630
reasonable decision making because in essence, a lot of times we don't have a rights to get a job or insurance or don't go to university.

595
01:03:00,630 --> 01:03:06,570
So if you don't have any loss and regulate how those decisions have to be made and data protection doesn't give you doubts,

596
01:03:06,570 --> 01:03:10,800
maybe it's the ethical thing to think about what would be reasonable.

597
01:03:10,800 --> 01:03:14,400
And that leads me to the current research project that we've just started,

598
01:03:14,400 --> 01:03:19,780
which is a right to reasonable inferences in online advertising and in financial services.

599
01:03:19,780 --> 01:03:23,040
And we just had to postdocs starting with us.

600
01:03:23,040 --> 01:03:26,570
One is a philosopher, the other one a legal scholar.

601
01:03:26,570 --> 01:03:33,630
And we are trying to figure out what would be a reasonable standard for inferentially takes in those two sectors in the future.

602
01:03:33,630 --> 01:03:41,310
We're going to look at other sectors as well. That's will include, for example, health, employment and criminal justice.

603
01:03:41,310 --> 01:03:45,570
But this is the current focus of our project.

604
01:03:45,570 --> 01:03:52,650
The last area where I see that A.I. is actually disrupting the law has to do with non-discrimination law.

605
01:03:52,650 --> 01:03:59,400
Again, that's something that we are all aware of, that whenever we talk about the IBE, you also have to think about bias and discrimination.

606
01:03:59,400 --> 01:04:05,040
So, again, I did look at non-discrimination law and tried to figure out if the law is actually good

607
01:04:05,040 --> 01:04:09,990
enough and it's good enough to protect us against those unintended and undesirable risks.

608
01:04:09,990 --> 01:04:13,800
And again, I have to say, it doesn't and it doesn't.

609
01:04:13,800 --> 01:04:18,630
It actually is not really the laws fold or technology fold. It's just something that happened.

610
01:04:18,630 --> 01:04:24,420
If an Oscar nomination law is an answer to undesired behaviour of people, right.

611
01:04:24,420 --> 01:04:30,240
That means that very often we want to regulate or prohibit that people are

612
01:04:30,240 --> 01:04:34,740
unethical to each other or being racist or sexist or homophobic to each other.

613
01:04:34,740 --> 01:04:39,420
But the way that algorithms do that is very, very different than humans.

614
01:04:39,420 --> 01:04:44,970
So, for example, if you think about press discrimination,

615
01:04:44,970 --> 01:04:54,030
suppress discrimination means that the same product is offered to different people at different prices in the offline world.

616
01:04:54,030 --> 01:04:56,250
You will be there. It would be very easy for you to figure it out.

617
01:04:56,250 --> 01:05:05,010
You can just go to different stores and compare fruits and prices and pick the products that you think is fair or appropriate.

618
01:05:05,010 --> 01:05:08,550
For example, Tesco doesn't let you into the store anymore.

619
01:05:08,550 --> 01:05:13,740
You know that you have been excluded from the market. How do you know that in an online world?

620
01:05:13,740 --> 01:05:18,210
Right. How do you know if you're actually being offered the best price? Do you know what the prices are?

621
01:05:18,210 --> 01:05:22,480
People get offered. And do you know the advertising is that you don't see.

622
01:05:22,480 --> 01:05:26,580
Right. So that kind of discrimination, you don't feel it anymore.

623
01:05:26,580 --> 01:05:33,130
Whereas in the off line, Wolf, there will be often D ability of the complainant to see that something's off.

624
01:05:33,130 --> 01:05:40,560
Something feels unfair, unjust, but that sends a feeling of unfairness might not be as easy to grasp anymore.

625
01:05:40,560 --> 01:05:50,910
And similarly here with the law, if you look at the the classes that we want to protect, that relates to sex and gender,

626
01:05:50,910 --> 01:05:58,850
ethnicity, religious beliefs, because we had historical experiences where people have used that to discriminate against people.

627
01:05:58,850 --> 01:06:04,110
Right. So this is how we constructed the world, how we group people.

628
01:06:04,110 --> 01:06:07,410
But algorithms might move people completely differently, right?

629
01:06:07,410 --> 01:06:15,330
It could be the diagrammed start to discriminate against people who are born on a Tuesday, have brown socks and like dogs.

630
01:06:15,330 --> 01:06:24,900
But this class of people doesn't find information non-discrimination law because we never had it or similar here with video gamers,

631
01:06:24,900 --> 01:06:30,000
which, for example, in China is something that could cause your credit score to drop.

632
01:06:30,000 --> 01:06:37,270
Obviously, that class of people doesn't find any protection non-discrimination law, but maybe it shouldn't.

633
01:06:37,270 --> 01:06:44,430
Maybe. So this is D.M. the project that we are also working on, trying to figure out if the law doesn't give you enough protection.

634
01:06:44,430 --> 01:06:50,820
What would be the ethical thing to do? How can we increase and protect fairness and then algorithmic world?

635
01:06:50,820 --> 01:06:56,730
So, yeah, basically does the programme that we're currently working on.

636
01:06:56,730 --> 01:07:02,220
And I actually think that the one of the reasons why I'm so excited about all this, what's happening,

637
01:07:02,220 --> 01:07:12,300
because I think it finally is clear to everyone that if we want to use a AI for a good year, we need to think of at least three perspectives.

638
01:07:12,300 --> 01:07:16,020
Need to think about is the law actually helping you? Is a law good enough?

639
01:07:16,020 --> 01:07:20,040
And if it's not, what would be desirable? What would the ethical thing to do?

640
01:07:20,040 --> 01:07:24,850
And then actually map it again. Technology and figure out if it's actually feasible and if you do that,

641
01:07:24,850 --> 01:07:29,590
you can harness the full potential of A.I. but make sure that you protect human rights.

642
01:07:29,590 --> 01:07:41,560
Thank you. Thanks very much indeed, Sandra.

643
01:07:41,560 --> 01:07:50,740
And now Brent, who is Sandra's colleague. Brent Mitchell shot, is a philosopher and postdoctoral research fellow in data ethics at the Oh I.

644
01:07:50,740 --> 01:08:00,500
He specialises in fairness, interprete ability, group privacy and other aspects of the ethical governance of A.I. and complex algorithmic systems.

645
01:08:00,500 --> 01:08:07,580
Over you, Brent. Yes, so thank you so much for inviting me into this.

646
01:08:07,580 --> 01:08:12,160
And thank you, Senator, for being great introduction of our research programme.

647
01:08:12,160 --> 01:08:17,110
So what I'm going to talk about a bit is I'm going to talk about one of my most recent pieces of work,

648
01:08:17,110 --> 01:08:27,490
which I think speaks to the question of the relationship between law and ethics in the space of a I and particularly talks about the the frameworks,

649
01:08:27,490 --> 01:08:35,350
the initiatives that we now have, over 100 of them that are in some way trying to define the right sort of high level

650
01:08:35,350 --> 01:08:41,200
concepts or principles or values or tenets to in some way guide the development of A.I.,

651
01:08:41,200 --> 01:08:49,480
the government's way of use. I salute the talk is going to be based on this paper, which just came out about a week ago,

652
01:08:49,480 --> 01:08:58,540
where I look at the role the principles can and perhaps should play in the governance of A.I. ethically.

653
01:08:58,540 --> 01:09:04,210
Now, my paper, despite having come out a week ago, is apparently already out of date.

654
01:09:04,210 --> 01:09:12,020
I thought we only had 84 initiatives from across the world, but we have over 100 now.

655
01:09:12,020 --> 01:09:19,530
And I hope this gets crossing the gravity of the situation.

656
01:09:19,530 --> 01:09:22,920
And I suppose my main concern is that if we have all these initiatives,

657
01:09:22,920 --> 01:09:29,140
youre essentially creating a market where developers can pick and choose the set of principles that works best for them.

658
01:09:29,140 --> 01:09:33,430
Chris Russell was joking the other day that he. Here are my principles.

659
01:09:33,430 --> 01:09:42,430
If you don't like them, I have a bunch of others. But to me, the risk there is it gives the impression that there's sort of one way to do ethics,

660
01:09:42,430 --> 01:09:49,390
that you have a high level set of principles that you were then going to specify into a set of practical requirements.

661
01:09:49,390 --> 01:09:57,670
And unsurprisingly, you should expect that if you have both lots of different sets of principles, but also leave it up to,

662
01:09:57,670 --> 01:10:03,280
say, individual development teams or individual companies to choose how to specify those principles on the ground.

663
01:10:03,280 --> 01:10:09,040
You're going to get requirements that don't match with each other requirements and may be contradictory.

664
01:10:09,040 --> 01:10:13,600
And essentially, you can end up in a situation of extreme more relativism.

665
01:10:13,600 --> 01:10:21,610
And to me, that that that thought, that sort of ethics or at least ethical frameworks are there as something to,

666
01:10:21,610 --> 01:10:29,320
let's see, replace regulation or at least self-regulation can be seen as a way to replace hard regulation.

667
01:10:29,320 --> 01:10:33,010
That's me sort of does a disservice to the actual value that ethics can have.

668
01:10:33,010 --> 01:10:39,430
And in particular, the value of the ethics and the law in working together in a good way can have.

669
01:10:39,430 --> 01:10:45,580
So what I'm interested in is basically how can we make these ethics frameworks actually work within a,

670
01:10:45,580 --> 01:10:49,630
you know, an approach where we're open to hard regulation?

671
01:10:49,630 --> 01:10:57,110
Now, I'm happy that this has come up already, but it's just to say there's a clear connexion between bioethics and medical ethics.

672
01:10:57,110 --> 01:11:02,150
So I don't have to make the case for that. So thank you. A number of people mentioned that, actually.

673
01:11:02,150 --> 01:11:08,350
But just to say that there was some work done recently, a couple of papers done recently that was trying to look for some sort of consensus

674
01:11:08,350 --> 01:11:12,730
across all these different different frameworks and initiatives that we had.

675
01:11:12,730 --> 01:11:17,290
And what in particular, this piece of work from the A.I. from People Project.

676
01:11:17,290 --> 01:11:21,490
And also, it was adopted by the High-Level Expert Group on Artificial Intelligence,

677
01:11:21,490 --> 01:11:28,540
is that there's a set of principles and very closely mimic the classic principles of medical ethics.

678
01:11:28,540 --> 01:11:36,580
And that's really interesting because I think actually looking at medical ethics and looking how a development compares to it can tell you a lot

679
01:11:36,580 --> 01:11:46,020
about whether we should expect this sort of principled form of self governance to work in a similar way in a development as it does in medicine.

680
01:11:46,020 --> 01:11:51,700
That's not to say that medical ethics is without its failures about a principled approach is going to solve everything.

681
01:11:51,700 --> 01:11:57,910
It's just to say it's definitely a case where the use of principles for ethical governance, for ethical decision making is very clear.

682
01:11:57,910 --> 01:12:02,170
It's very prominent and it has had impact in practise.

683
01:12:02,170 --> 01:12:08,530
So what we're going to do in the remainder of my few minutes is just to look at those two professions a little bit closer and see, well,

684
01:12:08,530 --> 01:12:17,360
how exactly does a development compare to this profession where we've seen a principled approach to ethics emerge and have some success?

685
01:12:17,360 --> 01:12:23,620
And so there's four different characteristics I want to look at here. The first is the existence of common aims.

686
01:12:23,620 --> 01:12:30,460
And what I mean by that is that medicine has doctors have fiduciary duties towards their patients.

687
01:12:30,460 --> 01:12:37,600
The practise is broadly guided by a common set of aims, which is to benefit the health and well-being of the patient.

688
01:12:37,600 --> 01:12:41,700
And of course, we will disagree about how to do that best in practise.

689
01:12:41,700 --> 01:12:48,230
At the very least, talking about, say, public public interests versus the interests of individual patient.

690
01:12:48,230 --> 01:12:54,430
But there is this sort of commonality. There is this common ground from which ethical decision making can actually proceed.

691
01:12:54,430 --> 01:13:01,990
And that is reflected in the fact that you have very strong fiduciary duty between professionals and patients.

692
01:13:01,990 --> 01:13:07,920
What that allows for is. Basically, co-operative approach to ethical decision making to specifying these principles.

693
01:13:07,920 --> 01:13:14,910
And I'm not sure that that level of cooperation can be taken for granted in the space of a development at the very least.

694
01:13:14,910 --> 01:13:22,410
If you're talking about privately developed Dahai, the initial fiduciary duties owned by the development team will be to the

695
01:13:22,410 --> 01:13:27,250
shareholders of the company rather than the users of people affected by the system.

696
01:13:27,250 --> 01:13:31,710
And so my concern is, if you don't have the sort of common grounding of the direction we're all travelling,

697
01:13:31,710 --> 01:13:36,930
then I'm not sure that you actually can have that that you're going to end up with with

698
01:13:36,930 --> 01:13:40,650
ethical decision making that's fundamentally competitive rather than cooperative.

699
01:13:40,650 --> 01:13:49,050
And clearly, that's that's not something that's conducive to finding an appropriate balance between different interests and practise.

700
01:13:49,050 --> 01:13:53,820
The second characteristic I want to look at is the role of professional history.

701
01:13:53,820 --> 01:13:58,770
And this is just to say that medicine has obviously an oppression for a very long time.

702
01:13:58,770 --> 01:14:03,810
We have codes of conduct. We have accounts of what it means to be a good doctor.

703
01:14:03,810 --> 01:14:08,820
These are captured on things like credit go through more recently, the Aimé code, medical ethics.

704
01:14:08,820 --> 01:14:14,190
These are documents that have been revised over time, the very longstanding, the very detailed.

705
01:14:14,190 --> 01:14:18,930
They give opinions on particular types of interventions.

706
01:14:18,930 --> 01:14:20,880
And the fact that they have been sought,

707
01:14:20,880 --> 01:14:28,710
tested over time and revised is one of the reasons that they they continue to actually be useful in day to day professional practise.

708
01:14:28,710 --> 01:14:37,080
Now, if you compare that to the state of a development, I'm going to use software engineering here as an analogue to to a development.

709
01:14:37,080 --> 01:14:43,080
We do have professional bodies. I see ACM in the eye triply, of course, being two biggest ones.

710
01:14:43,080 --> 01:14:50,430
And we do have codes of ethics. But in comparison to medicine there, the relatively short comparison, relatively abstract.

711
01:14:50,430 --> 01:14:57,120
The ACM one was revised recently, but still remains comparatively abstract when compared to, say, the main code of ethics.

712
01:14:57,120 --> 01:14:57,950
And there's research,

713
01:14:57,950 --> 01:15:05,940
a recent research that suggests that the codes of ethics actually have very little influence on day to day decision making of engineers,

714
01:15:05,940 --> 01:15:09,360
which is clearly a problem because if you had to have a code of ethics,

715
01:15:09,360 --> 01:15:19,400
but it isn't shown to be particularly effective in the sense of influencing the behaviour of engineers, then you have to question what is the actual.

716
01:15:19,400 --> 01:15:26,660
The third characteristic on a briefly touch on is just the methods that we have for translating principles into practise.

717
01:15:26,660 --> 01:15:30,880
The fundamental problem that we have here in the ethics frameworks is that they're

718
01:15:30,880 --> 01:15:36,020
they're based on or they rely on what can be called essentially contested concepts,

719
01:15:36,020 --> 01:15:40,610
essentially very abstract concepts that can have lots of different meanings in practise.

720
01:15:40,610 --> 01:15:46,820
If I asked you if I took a survey of this room, well, who thinks that a system should treat people fairly?

721
01:15:46,820 --> 01:15:48,890
I imagine pretty much everybody would raise their hand.

722
01:15:48,890 --> 01:15:54,620
And yet everybody may have a completely different sense of what fairness would actually mean in practise.

723
01:15:54,620 --> 01:15:59,300
And the point is that those different means of fairness can be rationally held.

724
01:15:59,300 --> 01:16:06,710
They can be genuine. They can be defended. We shouldn't expect a single correct meaning of any of these terms.

725
01:16:06,710 --> 01:16:12,490
The problem that you have is that we've reached consensus on what the correct, essentially contested concepts are.

726
01:16:12,490 --> 01:16:17,930
We've we've reached this high level consensus. But to me, that doesn't actually reflect any sort of true consensus.

727
01:16:17,930 --> 01:16:25,010
And more just is a way to mask sort of really important normative and political disagreements have a framework.

728
01:16:25,010 --> 01:16:30,040
We agreed in the framework. But we certainly don't agree to what it means in practise.

729
01:16:30,040 --> 01:16:36,780
And of course, the other problem is that those sorts of concepts don't translate automatically into practical requirements.

730
01:16:36,780 --> 01:16:43,950
That is a very difficult process. Medicine has things like professional societies and boards, ethics review committees,

731
01:16:43,950 --> 01:16:49,140
accreditation and licencing schemes, pure self governance standards.

732
01:16:49,140 --> 01:16:53,530
All these things help you actually do that. Translation and practise.

733
01:16:53,530 --> 01:17:02,440
Software engineering does have some of these things, but it's lacking it's lacking mechanisms that are of, say, similar stature, similar importance.

734
01:17:02,440 --> 01:17:08,500
I think one of the reasons for that is that the profession would not be legally

735
01:17:08,500 --> 01:17:14,050
recognised in a legally recognised as a profession in the same way that medicine is.

736
01:17:14,050 --> 01:17:20,320
What I mean by that is doctors require a licence to practise software engineers, in some cases in limited contexts,

737
01:17:20,320 --> 01:17:30,640
do need a licence to practise, but it's in no way of the same importance or the same sort of the same coverage as as medical licences.

738
01:17:30,640 --> 01:17:35,860
And so this lack of sort of legal recognition of the profession is a serious problem because ethics tends

739
01:17:35,860 --> 01:17:42,790
to professional ethics at least really has teeth when it actually has legal mechanisms to back it up.

740
01:17:42,790 --> 01:17:49,350
When doing something unethical in your day to day behaviour as a as a professional could actually lead to,

741
01:17:49,350 --> 01:17:55,100
say, losing the ability to practise your profession.

742
01:17:55,100 --> 01:18:02,020
And so just to conclude here, I think we have a number of sort of legal gaps within ethical way.

743
01:18:02,020 --> 01:18:08,050
One is really important ones that there is not this legal recognition of A.I. developments as professions.

744
01:18:08,050 --> 01:18:12,820
And a lot of the initiatives we have are based on human rights frameworks or other

745
01:18:12,820 --> 01:18:16,720
sorts of frameworks that are not directly legally binding in the same way as,

746
01:18:16,720 --> 01:18:22,300
say, to Judy Pyar would be. And so to move forward, I think we can do a couple of things.

747
01:18:22,300 --> 01:18:29,290
One is to start thinking about ethics more in the sense of a business or organisational ethics and less as a professional ethics.

748
01:18:29,290 --> 01:18:35,710
Both sides are important. But I think there's too much emphasis placed on individuals doing wrong things software engineer,

749
01:18:35,710 --> 01:18:40,990
individual developers doing wrong things and less about the same business model that they're

750
01:18:40,990 --> 01:18:46,270
working within or the organisational practises themselves being unethical in some way.

751
01:18:46,270 --> 01:18:54,450
We had a chance to do that with a High-Level Expert Group. There were red lines initially supposed to be draughted to basically send out types of AOG.

752
01:18:54,450 --> 01:18:58,120
It should not be developed in Europe. In the end, those red lines were taken out.

753
01:18:58,120 --> 01:19:04,380
To me, these would been a very strong signal of A.I. Ethics is also a business ethics.

754
01:19:04,380 --> 01:19:12,460
And then just finally, I think we may need to revisit the idea of licencing developers of high risk A.I. applications.

755
01:19:12,460 --> 01:19:17,550
And we I think more than anything, we just need to develop much stronger empirical evidence base.

756
01:19:17,550 --> 01:19:22,320
This based on case studies, on specific, specific ethical challenges.

757
01:19:22,320 --> 01:19:26,700
So we can really start to understand how we disagree and agree with each other

758
01:19:26,700 --> 01:19:30,900
in practise about what these different concepts and principles go on there.

759
01:19:30,900 --> 01:19:40,480
Thank you very much. The.

760
01:19:40,480 --> 01:19:48,300
What we're getting is something of a taste of the range of very difficult issues that will need to be addressed by the institute.

761
01:19:48,300 --> 01:19:56,760
We're going to end with a session on in medicine. Comparisons have been drawn between medicine and they are quite a few times already.

762
01:19:56,760 --> 01:20:00,930
But now we're going to hear what A.I. in medicine is really like.

763
01:20:00,930 --> 01:20:04,260
And our first speaker is Professor Gil McBain.

764
01:20:04,260 --> 01:20:13,950
He's professor of statistical genetics, director of the Big Data Institute and amongst various other honours, a fellow of the Royal Society.

765
01:20:13,950 --> 01:20:21,430
Five years ago, he co-founded the Company Genomics P.L.C. to use genomics to improve healthcare.

766
01:20:21,430 --> 01:20:29,160
You know, I think the sort of the kindest way of describing me in the context of this room is very much as a non expert,

767
01:20:29,160 --> 01:20:35,340
at least when it comes to ethics. I am a statistician and a geneticist,

768
01:20:35,340 --> 01:20:43,050
and I've worked in the field of using genetics to understand human disease and to identify

769
01:20:43,050 --> 01:20:50,400
opportunities for new therapies or to better sort of predict where people are on that trajectory.

770
01:20:50,400 --> 01:20:55,620
I've done that for many years. And over the last few years,

771
01:20:55,620 --> 01:21:02,250
my role has shifted a bit from purely being someone who was doing research to someone

772
01:21:02,250 --> 01:21:07,170
who has been thinking about the kinds of infrastructure and kinds of communities that

773
01:21:07,170 --> 01:21:12,660
you need to build in order to make sure that this growing part of biomedical research

774
01:21:12,660 --> 01:21:17,460
is sort of done well at the university level and translated well and ethically,

775
01:21:17,460 --> 01:21:24,990
of course, into practise. And so behind that is my very much my role as the director of the Big Data Institute,

776
01:21:24,990 --> 01:21:35,070
which is one of these recent now not quite the latest new centres to have appeared within University of Oxford.

777
01:21:35,070 --> 01:21:38,550
For those of you who don't know us, where we are, a physical thing,

778
01:21:38,550 --> 01:21:46,850
where a new institute and new building up the hill in the old growth campus, and this is somewhere between 350 and 500.

779
01:21:46,850 --> 01:21:58,500
And then you can how you can people within this institute but who are united by this desire to use a data

780
01:21:58,500 --> 01:22:05,160
driven approach to understanding the causes of human disease and identifying groups to intervention.

781
01:22:05,160 --> 01:22:10,280
So we're entirely dry land. We're entirely computational, as it were.

782
01:22:10,280 --> 01:22:15,960
And really what we are about is creating this fuel for a car.

783
01:22:15,960 --> 01:22:22,320
Now, we've heard an awful lot about this sort of the ethical issues of how you actually

784
01:22:22,320 --> 01:22:28,860
use A.I. in context and under the regulatory or the legal aspects around that.

785
01:22:28,860 --> 01:22:33,390
Perhaps what we haven't heard quite so much about is the process by which you can

786
01:22:33,390 --> 01:22:39,540
acquire what is the fundamental part that has to go into any are a are a process,

787
01:22:39,540 --> 01:22:50,980
which is very much the data itself. Within the institute, just very briefly, there are four types of things that we do.

788
01:22:50,980 --> 01:22:56,530
First is about how we measure things. This sort of measurement technologies.

789
01:22:56,530 --> 01:23:02,380
The second is about how we bring all those data together to create the research, ready the analysis,

790
01:23:02,380 --> 01:23:09,070
ready datasets that our researchers and others can come on in and peer into

791
01:23:09,070 --> 01:23:14,230
and try and identify the structure that ultimately leads to these insights.

792
01:23:14,230 --> 01:23:23,230
We have people from statistics, computer science, engineering, epidemiology, genomics, et cetera, developing methods.

793
01:23:23,230 --> 01:23:31,330
That is, if you like, the VII algorithms which are going to peer into this these kinds of rich data set.

794
01:23:31,330 --> 01:23:38,890
And then finally, and probably why in here is that the fourth pillar of what we do with this institute is to

795
01:23:38,890 --> 01:23:45,670
think hard about the much more the why does societal aspects of this data driven landscape.

796
01:23:45,670 --> 01:23:50,560
So issues around consent, issues around privacy, security issues.

797
01:23:50,560 --> 01:23:54,520
Right. Governance issues around sharing intellectual property.

798
01:23:54,520 --> 01:24:03,350
And so. And we made a decision right at the start of putting this institute together that this was something that we wanted to go on.

799
01:24:03,350 --> 01:24:09,320
Actually, in the building, it's such an integral part of doing biomedical research these days.

800
01:24:09,320 --> 01:24:18,740
And the issues that us that come out of this kind of research are so deep that if we don't train people in how you

801
01:24:18,740 --> 01:24:25,730
should think about conducting this kind of research and you don't build the right practises into how people do it,

802
01:24:25,730 --> 01:24:34,400
prosecuting the research programmes actually at the point of implementation, then you've kind of you're starting on the wrong foot.

803
01:24:34,400 --> 01:24:42,920
So we very much put that at the heart of the institute. A march of ethos is based within the Big Data Institute.

804
01:24:42,920 --> 01:24:52,670
And we very recently we got funding from EPA Sarsae to set up a new central doctoral training in health data science.

805
01:24:52,670 --> 01:25:01,670
One of the key pillars of that programme being that these data scientists and machine learners and so on would be trained

806
01:25:01,670 --> 01:25:10,580
very much alongside all the other skills in what they need in the skills to think about the problems from that standpoint.

807
01:25:10,580 --> 01:25:15,710
So it really, really is central to how we we think.

808
01:25:15,710 --> 01:25:24,230
And if anyone's interested in use cases coming up to Lausanne, talking about the types of problems that we're working on,

809
01:25:24,230 --> 01:25:29,030
the types of dilemmas that we're faced with, then please do get in touch.

810
01:25:29,030 --> 01:25:32,160
We'd be more than happy to talk.

811
01:25:32,160 --> 01:25:42,690
So I just want to say just a couple of things about at least my sort of personal perspectives on why the types of research that we're doing now,

812
01:25:42,690 --> 01:25:47,970
which is very much within the tradition of biomedical or medical research.

813
01:25:47,970 --> 01:25:55,930
Why those are a bit different and why they're raising new challenges from from the sort of the ethical perspective.

814
01:25:55,930 --> 01:26:01,080
And I think a really important point to start, which is perhaps not so well understood,

815
01:26:01,080 --> 01:26:08,130
is actually that the growth of AI and machine learning technologies within biomedical

816
01:26:08,130 --> 01:26:15,420
research has really led to something of a shift in how medical research itself is conducted.

817
01:26:15,420 --> 01:26:20,360
And this comes back to this question of data. How we how we get the data is.

818
01:26:20,360 --> 01:26:26,880
So it used to be that in medical science, you had a hypothesis, you.

819
01:26:26,880 --> 01:26:30,980
You decided I wanted to test some particular question.

820
01:26:30,980 --> 01:26:36,770
And off the back of that, you designed an experiment, not experiment, gave you some data.

821
01:26:36,770 --> 01:26:43,820
You analyse the data on the back of that and make some conclusions. And maybe you came up with a new hypothesis.

822
01:26:43,820 --> 01:26:48,540
Importantly, those data were collected specifically for that purpose.

823
01:26:48,540 --> 01:26:55,910
And you went out you explain to people why you were going to collect that data and what you hoped to learn come out of it.

824
01:26:55,910 --> 01:27:00,680
That's a very clean way of doing science. But clearly, it's not massively scalable.

825
01:27:00,680 --> 01:27:06,710
There's one question that you could ask of those data and essentially one.

826
01:27:06,710 --> 01:27:17,310
Now, about ten, fifteen years ago, biomedical research sort of took a side step, it changed direction a bit in how it collected data,

827
01:27:17,310 --> 01:27:27,210
and a lot of that came out of the world of genomics where people realised who had been studying how genes affect diseases.

828
01:27:27,210 --> 01:27:32,580
They'd been studying sort of their favourite gene and their favourite disease in a particular combination.

829
01:27:32,580 --> 01:27:40,810
And the literature was full of incredibly bad results that never heated and were massively underpowered.

830
01:27:40,810 --> 01:27:45,920
But 15 years ago, what happened was a change in technology.

831
01:27:45,920 --> 01:27:55,970
Changes in technology that start things. That legs to our us being able to experiments not just on one gene and a handful of individuals,

832
01:27:55,970 --> 01:28:01,030
but the entire genome in tens of thousands of individuals.

833
01:28:01,030 --> 01:28:11,140
And that led to this idea that rather than going in with your specific hypothesis, actually the most powerful thing is to go in without a hypothesis.

834
01:28:11,140 --> 01:28:16,630
You go in and you just collect data and you let the data tell you what the answer is.

835
01:28:16,630 --> 01:28:22,990
And that idea has very much percolated from just thinking about, well, let's study the whole genome.

836
01:28:22,990 --> 01:28:30,130
And one disease, the Genome Association studies was essentially that idea to the idea that you go in and you

837
01:28:30,130 --> 01:28:36,100
collect the genome and you collect everything that you possibly can about an individual's health,

838
01:28:36,100 --> 01:28:41,980
environment, lifestyle, finances. You just collect everything you can.

839
01:28:41,980 --> 01:28:47,730
And later on, you decide what research from. No.

840
01:28:47,730 --> 01:28:55,740
The success of this programme is sort of made real by something with the UK biobank, which many of you will probably know about,

841
01:28:55,740 --> 01:29:06,240
but about somewhere between one and two percent of adults within the UK have consented to have their entire medical data,

842
01:29:06,240 --> 01:29:09,030
their entire genome sequence.

843
01:29:09,030 --> 01:29:18,620
I'm huge Moundsville axillary information about them, lifestyle and cognition that their parents, sometimes the children.

844
01:29:18,620 --> 01:29:28,820
Huge amounts of information made available to people like me and people like you and people in companies and people in China and people in the U.S.,

845
01:29:28,820 --> 01:29:36,560
all you have to do is this to sign up to a very few sort of restrictions about what you're going to do with the data.

846
01:29:36,560 --> 01:29:43,730
You have to say roughly what you're going to do with the data. You have to say that you're not going to try and identify these people.

847
01:29:43,730 --> 01:29:50,730
But beyond that, it's really not very much that you have to say that you're going to do.

848
01:29:50,730 --> 01:29:56,880
And as a consequence of that, there are people all over the world probing the tiniest details,

849
01:29:56,880 --> 01:30:04,630
the most intimate information, about half a million people within the U.K., some probably indeed within this.

850
01:30:04,630 --> 01:30:11,710
So it's an example of how our way of doing research is really shifting.

851
01:30:11,710 --> 01:30:18,460
This shift is exactly what enables the whole A.I. revolution in medicine and health care.

852
01:30:18,460 --> 01:30:26,440
But it, of course, brings up all sorts of questions about what it means to be informed about research project, which has no end.

853
01:30:26,440 --> 01:30:34,780
What it means in terms of can you ever can you ever comprehend the sorts of things that I might

854
01:30:34,780 --> 01:30:40,270
learn about you if I bring together lots of sorts of information that you would never have had?

855
01:30:40,270 --> 01:30:49,720
And what what would you like to know if I can, for example, predict whether you're going to get out on this disease in the next 10, 20, 50 years?

856
01:30:49,720 --> 01:31:00,040
Huge amounts of new challenges arising from it, which are where only, I think, just begin to describe the topic that I shall shut out.

857
01:31:00,040 --> 01:31:08,250
Thanks very much. Thank you, Joe.

858
01:31:08,250 --> 01:31:15,180
It's interesting how some of the points that you raised there actually could be potentially revolutionary,

859
01:31:15,180 --> 01:31:19,800
revolutionary or other areas of science, too, and views about philosophy and science.

860
01:31:19,800 --> 01:31:25,530
The idea that you play data and then you form a hypothesis rather than the other way round.

861
01:31:25,530 --> 01:31:30,420
Our final speaker today is Jess Morley. She's policy lead at Oxford.

862
01:31:30,420 --> 01:31:39,790
Evidence based medicine data lab research assistant at the Oxford Internet Institute Digital Ethics Lab and a subject matter expert,

863
01:31:39,790 --> 01:31:47,850
the NHS NHS X, whose aim is to drive digital transformation within the NHS.

864
01:31:47,850 --> 01:31:52,620
Here she focuses on the ethics policies and regulations needed to enable the use of

865
01:31:52,620 --> 01:31:59,550
healthcare data for the improvement of outcomes whilst minimising the associated risks.

866
01:31:59,550 --> 01:32:09,390
Thank you. I also don't have any slices. Borlase causes too much, so I'm also very aware that I am kind of the stop gap between you and wine.

867
01:32:09,390 --> 01:32:15,270
I'm getting out of this very room, so I will try and not be very boring.

868
01:32:15,270 --> 01:32:21,420
I think what you have had so far is a really wonderful introduction to what I am trying to talk about,

869
01:32:21,420 --> 01:32:27,060
because I sit in a number of different intersections in that I am both an employee of

870
01:32:27,060 --> 01:32:31,310
the university and a student at the university in a different part of the university.

871
01:32:31,310 --> 01:32:39,690
And I also am working in academia and I am actually making policy today because of Puzder.

872
01:32:39,690 --> 01:32:49,170
I have to very much be talking from the Oxford perspective. But if you have questions about specific things afterwards.

873
01:32:49,170 --> 01:32:56,790
But really what I try and do and I think sonder faced it really, really well is look at what is legally required,

874
01:32:56,790 --> 01:33:06,570
which in the context of medicine and data use, tends to be consent, anonymization, then what is ethically desirable and what's the standard?

875
01:33:06,570 --> 01:33:12,120
And then we try and develop what we refer to as a principle proportionate approach that

876
01:33:12,120 --> 01:33:19,620
tries to move the what of a ethics into the how of ethics and to really operationalise it.

877
01:33:19,620 --> 01:33:24,000
Look at a number of different things. First of all, how can I actually help?

878
01:33:24,000 --> 01:33:27,170
I think this is something that often gets missed in ethical conversations.

879
01:33:27,170 --> 01:33:33,390
People always talk about the risks, but actually a very big risk, particularly from the perspective of the NHS,

880
01:33:33,390 --> 01:33:42,330
that has a duty to capitalise on all of the ways that you can enable healthcare is the opportunity cost of not letting it happen.

881
01:33:42,330 --> 01:33:48,630
So to try and look at why does it help, which tends to be largely in categories such as diagnostic screening.

882
01:33:48,630 --> 01:33:51,810
And Josh, that's where most research currently focuses,

883
01:33:51,810 --> 01:33:57,450
primarily because of that's where the data is more easily accessible and it's why it's standardised.

884
01:33:57,450 --> 01:34:06,090
Population health. This is time basically better versions and less bad versions of Google flu discovery,

885
01:34:06,090 --> 01:34:12,300
such as drug discovery and what's sometimes referred to as P4 medicine or precision medicine.

886
01:34:12,300 --> 01:34:17,120
And those are the kind of areas that we look at in terms of why we think it might help.

887
01:34:17,120 --> 01:34:23,040
And then the question is, how can it actually hurt? You've already heard many of the reasons why.

888
01:34:23,040 --> 01:34:29,580
And one of the things that I've co-authored is one of those many ethical principles that Brent had up on that on that slide,

889
01:34:29,580 --> 01:34:35,600
which is the Department of Health and Social Care has a code of conduct for the NHS.

890
01:34:35,600 --> 01:34:42,360
So it's called the NHS Code of Conduct for data driven health care technologies, which just really both of your tongue.

891
01:34:42,360 --> 01:34:46,080
The reason it is cool that is because most of the applications that we see

892
01:34:46,080 --> 01:34:50,040
currently in these within the NHS are within the healthcare sector at large.

893
01:34:50,040 --> 01:35:00,230
It's at least at the frontline do not fall into the category of being a I, yet they tend to be fastpass, simpler basic statistics.

894
01:35:00,230 --> 01:35:08,340
So we've tried to encompass it with data, didn't get it. But the things that try and look at is that that code of conduct and almost everything that

895
01:35:08,340 --> 01:35:13,740
we've heard about so far tends to focus on just the individual protecting the individual,

896
01:35:13,740 --> 01:35:19,890
which stems from this connexion with medical ethics being around protecting the person.

897
01:35:19,890 --> 01:35:24,650
And as we have seen, autonomy, beneficence, not maleficent justice and acceptability.

898
01:35:24,650 --> 01:35:32,040
And was that's really been important when you're operating and thinking of the perspective from a systems perspective.

899
01:35:32,040 --> 01:35:36,660
You also need to think about how ethics, how ethical risks can arise.

900
01:35:36,660 --> 01:35:39,690
A number of other different levels in the interpersonal level.

901
01:35:39,690 --> 01:35:46,470
So how much it changed the relationship between clinicians and patients, for example, and also between the patient and themselves.

902
01:35:46,470 --> 01:35:51,030
This is a very interesting dichotomy that you have in the healthcare sector with a guy is that

903
01:35:51,030 --> 01:35:56,910
you are dealing with simultaneously a very physical being and an entirely digital sphere.

904
01:35:56,910 --> 01:36:04,680
And sometimes those two things do not actually match. And that chemical is on some and need to be aware of the other.

905
01:36:04,680 --> 01:36:11,950
The other empathic you can see is group level impact, sectoral level impacts and set societal level.

906
01:36:11,950 --> 01:36:15,660
So a group and societal level impacts might be where, for example,

907
01:36:15,660 --> 01:36:23,430
that we have problems with the fact that if the NHS is grounded on this principle, that it is careful, all obvious.

908
01:36:23,430 --> 01:36:29,940
It's like making that less true by focussing on on the introduction of.

909
01:36:29,940 --> 01:36:33,570
When I should be used in a way that makes it more true.

910
01:36:33,570 --> 01:36:40,920
But we could, in fact, introduce an new and evolving inequality and exacerbate those already existing.

911
01:36:40,920 --> 01:36:46,860
And then the other thing I said to level and this is particularly important again, is taking a systems perspective,

912
01:36:46,860 --> 01:36:57,210
is the issue with trust on how might you lose trust in the provision of health care from the states because you have made mistakes too early on?

913
01:36:57,210 --> 01:37:01,320
Until there is all of the sort of levels that you might think about.

914
01:37:01,320 --> 01:37:06,830
And then the other area that we need to try and look at is where you might have.

915
01:37:06,830 --> 01:37:11,780
Different ethical issues arising at different levels of the machine learning pipeline.

916
01:37:11,780 --> 01:37:19,430
One of the things I often say is if you have written ethics in to the business case that coded it out by the time it got to the point,

917
01:37:19,430 --> 01:37:23,600
it's really clear example that we're currently tackling with us.

918
01:37:23,600 --> 01:37:31,730
A couple of things is one is because the question of liability is a little bit unsure.

919
01:37:31,730 --> 01:37:37,100
It tends to be interpreted that it land on the on the final clinician who's made the decision.

920
01:37:37,100 --> 01:37:45,670
But all of. Tools that are deployed at the front line or deployed in the app stores are things like skin latex, for example,

921
01:37:45,670 --> 01:37:51,640
which is an app that can take a photo of a mole and it will tell you your level of risk are skewed towards false

922
01:37:51,640 --> 01:37:57,750
positives because it's less risk to the company to say that you have a diagnosis when you don't want to do it.

923
01:37:57,750 --> 01:38:02,590
What's the issue that this is having is that whilst these technologies are

924
01:38:02,590 --> 01:38:08,020
supposed to be deployed in order to to enable people to take care of themselves.

925
01:38:08,020 --> 01:38:15,430
In fact, it is actually doing the opposite and driving people towards frontline care because people get these very high level ratings of risk.

926
01:38:15,430 --> 01:38:21,560
Similarly, we have an issue with. Like breast cancer screening algorithms, for example,

927
01:38:21,560 --> 01:38:28,250
that have drastically increase the capacity to recognise the potential of the pathology being that because can.

928
01:38:28,250 --> 01:38:33,140
A machine can be many scans at once. But a doctor can still only treat one person at once.

929
01:38:33,140 --> 01:38:39,500
So you effectively created a bottleneck, which is an ethical side from the perspective of that.

930
01:38:39,500 --> 01:38:45,590
Then they have the person who is living with the potentially anxiety inducing diagnosis.

931
01:38:45,590 --> 01:38:50,210
And that has very big implications for their medical integrity and that sort of mental integrity.

932
01:38:50,210 --> 01:38:56,960
But be waiting much longer for treatment. So those are the kinds of areas that we try and look at.

933
01:38:56,960 --> 01:39:03,350
I'm trying. Then develop standards and policies that build on the regulatory framework.

934
01:39:03,350 --> 01:39:11,510
There is work going on to assess the regulation and the various things like create regulatory sandboxes in order to test ideas.

935
01:39:11,510 --> 01:39:17,620
That takes a much longer than it takes to develop standards and policies and the implementation of them.

936
01:39:17,620 --> 01:39:22,700
I'm sure one of the things, for example, that we have looked at is the code of conduct.

937
01:39:22,700 --> 01:39:31,960
First of all, it is an abstraction of an abstraction. It's based on the ethical principles, but it is far more trying to operationalise them.

938
01:39:31,960 --> 01:39:38,960
So instead of saying things like autonomy, it says things like design for your specific user state,

939
01:39:38,960 --> 01:39:46,410
how the technology will impact them and how it will help and how it is better than what is already in place.

940
01:39:46,410 --> 01:39:50,870
And there's a number of operational things that we have done to make that a little

941
01:39:50,870 --> 01:39:55,160
bit more realistic because of the complexities that you're dealing with in medicine,

942
01:39:55,160 --> 01:39:57,230
such as safety and efficacy.

943
01:39:57,230 --> 01:40:05,750
For example, one of the things that we developed was the standards for evidence of efficacy in digital health technologies.

944
01:40:05,750 --> 01:40:11,960
And similarly, what is socially acceptable commercial models for health pages.

945
01:40:11,960 --> 01:40:19,100
And then to build a platform that shows how people can actually go in,

946
01:40:19,100 --> 01:40:24,230
put provide that evidence to the entities of how you are meeting those principles.

947
01:40:24,230 --> 01:40:29,960
What evidence have you provided and allow people to actually. This is all available in a big report.

948
01:40:29,960 --> 01:40:35,660
That's and it was published a couple of weeks ago. Anyone wants to see really that's what we're trying to do.

949
01:40:35,660 --> 01:40:43,790
Look at what are the issues to do with the fact that data is now circulating outside of the system to in many places,

950
01:40:43,790 --> 01:40:50,230
you now have private companies that may know more about your health than the healthcare system

951
01:40:50,230 --> 01:40:56,390
and that how does that breaking down traditional barriers that have been in good places?

952
01:40:56,390 --> 01:41:00,560
We had earlier things like the difference between medical research and medical practise

953
01:41:00,560 --> 01:41:05,480
when you're working in a cybernetic loop with data that starts to break down.

954
01:41:05,480 --> 01:41:10,910
Look at the risks that approach at different levels of society,

955
01:41:10,910 --> 01:41:18,290
different stages in the machine learning pipeline and how we actually operationalise those so that instead of just saying,

956
01:41:18,290 --> 01:41:32,560
isn't it nice that everyone is going to be fat? How will we actually know?

957
01:41:32,560 --> 01:41:37,420
Thank you very much, Jess. We're giving a little bit over the originally advertised time.

958
01:41:37,420 --> 01:41:47,050
So anyone who needs to go do feel free. But we've got three questions from the audience that have attracted nine votes.

959
01:41:47,050 --> 01:41:52,750
So I think it would be good to spend just a little bit of time addressing those.

960
01:41:52,750 --> 01:42:01,300
One of them said, what place does diversity and inclusion play in work done in the ethics of a I?

961
01:42:01,300 --> 01:42:06,610
What will the institute do to ensure a range of backgrounds and voices are involved?

962
01:42:06,610 --> 01:42:15,640
That's a biggie. And I don't think this is the time to be discussing the various policies that the institute will be implementing.

963
01:42:15,640 --> 01:42:24,760
But I'm going to take the opportunity to give a plug to the podcast Future Makers, where we have two excellent episodes last year.

964
01:42:24,760 --> 01:42:36,430
One of them discussing the bias of algorithms, in fact, where Sandra and Brent were present and another one does A.I. have a gender?

965
01:42:36,430 --> 01:42:44,650
And Charissa was a contributor to that. So in each case, you've got a discussion of 45, 50 minutes about these very difficult issues.

966
01:42:44,650 --> 01:42:47,470
Future makers, Oxford. And you'll find it.

967
01:42:47,470 --> 01:42:59,710
We had a Hulce series 10 podcasts on A.I. and ethical and other issues arising from A.I. But those those of you

968
01:42:59,710 --> 01:43:05,860
who aren't to the other questions that I'm going to mention that might want to say a little bit about that issue.

969
01:43:05,860 --> 01:43:16,250
So, Sandra, I think this is one aimed at you. What if I can make predictions that accurately predict predict risk?

970
01:43:16,250 --> 01:43:25,440
For example, about credit, health, criminal behaviour, but are, in effect discriminatory?

971
01:43:25,440 --> 01:43:33,790
Did you come to the front? Yeah. Many of our other speakers would like to do.

972
01:43:33,790 --> 01:43:44,180
And you seem to be the obvious. Yes. That is could be all a very wet period.

973
01:43:44,180 --> 01:43:49,170
It's for the record. Yeah, that's a very good question.

974
01:43:49,170 --> 01:43:53,730
And it is one that I don't yet have an answer to.

975
01:43:53,730 --> 01:44:00,240
And I think it is like the core question of what inspires my research in general,

976
01:44:00,240 --> 01:44:05,130
which is it's not just about the question what's technically possible.

977
01:44:05,130 --> 01:44:08,160
It's also about the question, is it ethical or desirable?

978
01:44:08,160 --> 01:44:15,150
So just because we get very good at certain things doesn't necessarily mean we have to do them.

979
01:44:15,150 --> 01:44:22,170
And I don't like the idea that the general idea of just because it's there, we should also use it.

980
01:44:22,170 --> 01:44:24,000
I think we should actually have a very,

981
01:44:24,000 --> 01:44:32,070
very thorough discussion of how we want to deploy those systems and if we actually want to welcome those in our society.

982
01:44:32,070 --> 01:44:33,960
So when people talk about bias,

983
01:44:33,960 --> 01:44:43,860
they often imagine that the problem is a system that's actually not very good at predicting things because it is biased by preconceptions.

984
01:44:43,860 --> 01:44:50,990
You're saying even if the system really predicts very well, but actually.

985
01:44:50,990 --> 01:44:55,910
Particular aspect to feeding into that decision that we'd rather didn't, right?

986
01:44:55,910 --> 01:45:00,260
We may want to hold back on implementation. I think it's two different things.

987
01:45:00,260 --> 01:45:05,450
I think there might be let's say there is a predictor that says the shape of your

988
01:45:05,450 --> 01:45:10,280
nose is a very good predictor of whether you're going to fail in law school or not.

989
01:45:10,280 --> 01:45:15,320
And I have statistical evidence that shows that. Right. And we know it's super accurate.

990
01:45:15,320 --> 01:45:18,590
We could still think. I don't. I think that's stupid.

991
01:45:18,590 --> 01:45:22,820
And I don't want my nose to have anything to do with whether or not I get admitted to Lorsque or not.

992
01:45:22,820 --> 01:45:24,980
Right. And that it's not a technical problem.

993
01:45:24,980 --> 01:45:34,520
That's a, you know, a question of what we think of what are desirable or socially acceptable criteria to make decisions based on rates.

994
01:45:34,520 --> 01:45:40,160
Grades may be reference Thatchers maybe, but noses are not, even if they are good predictors.

995
01:45:40,160 --> 01:45:46,390
The other thing would bias might be introduced is that we don't know what the shape of the nose actually correlates with.

996
01:45:46,390 --> 01:45:50,570
Right. It might be that, you know. I don't know.

997
01:45:50,570 --> 01:45:55,370
I women have particularly short noses and we don't know dad.

998
01:45:55,370 --> 01:45:58,220
And all the sudden moves start discriminating against people, women.

999
01:45:58,220 --> 01:46:04,190
And we don't know it because we don't have the link between gender and those as yet because we don't have enough research.

1000
01:46:04,190 --> 01:46:09,740
The only thing that we see is an interesting correlation between shape of noses and being successful in law school.

1001
01:46:09,740 --> 01:46:18,010
Right. So we have to be very, very careful when we choose does decision making tools and we don't know the actual causal link between those two,

1002
01:46:18,010 --> 01:46:23,900
which is related very much to your work on inscrutability of the machine, learning algorithms and all that's yapping.

1003
01:46:23,900 --> 01:46:30,980
You can't just take the results and use them without really understanding. Yes, I think that's that's the ethical change, the ethic of responsibility.

1004
01:46:30,980 --> 01:46:36,830
When we use big data for life changing decisions, then we should have a right to at least understand what's going on.

1005
01:46:36,830 --> 01:46:42,260
Respire, what's going on, or be honest about the fact we don't know what's going on and don't pretend to be,

1006
01:46:42,260 --> 01:46:47,720
you know, just trust him without actually questioning them. Thank you very much.

1007
01:46:47,720 --> 01:46:50,510
Another really big question here.

1008
01:46:50,510 --> 01:47:02,890
Do we foresee a time that A.I. will perform the work of ethics better than human intelligence based on our current definitions of AI and ethics?

1009
01:47:02,890 --> 01:47:12,810
Any volunteers may maybe your brain. Thank you.

1010
01:47:12,810 --> 01:47:20,220
That's an excellent question. I have a paper on that. I am very happy to to share, especially because I'm looking for feedback.

1011
01:47:20,220 --> 01:47:32,160
So if you show me any more, I will. But in a nutshell, I think that any entity that doesn't have the ability to feel pain and suffer and

1012
01:47:32,160 --> 01:47:36,780
feel pleasure and know what that is cannot possibly have any moral understanding.

1013
01:47:36,780 --> 01:47:41,580
Or so I argue. Right. So dot, dot, dot.

1014
01:47:41,580 --> 01:47:49,020
Read the map. Do any of our other speakers want to say anything about either of those big questions?

1015
01:47:49,020 --> 01:47:53,670
Can I say something about risk? Yeah. I mean, I find that question super interesting.

1016
01:47:53,670 --> 01:47:57,390
And it's one part of my interest in the ethics of prediction.

1017
01:47:57,390 --> 01:48:07,320
It seems to me at the moment there is a presumption that institutions have unlimited rights to assess risk and make decisions accordingly.

1018
01:48:07,320 --> 01:48:13,620
And I find that quite questionably questionable. So take insurance. Used to be the case that insurance was about pooling risk.

1019
01:48:13,620 --> 01:48:18,900
Right. So you got a big pool of people and you know that you have some notions of

1020
01:48:18,900 --> 01:48:23,640
statistics such that you know that some people are going to have health problems, other others won't.

1021
01:48:23,640 --> 01:48:30,180
And it balances itself out so that you can calculate how much you have to charge people to survive.

1022
01:48:30,180 --> 01:48:37,740
And now it seems like we can actually assess very, very accurately whether somebody is going to have a health problem or not.

1023
01:48:37,740 --> 01:48:41,670
And then if somebody is high risk, then you don't have to take them into insurance.

1024
01:48:41,670 --> 01:48:45,740
And it seems to me that there is this is a this is a process more in general, in society.

1025
01:48:45,740 --> 01:48:50,550
It's also in the job market and so on, in which it's not that risk is being diminished.

1026
01:48:50,550 --> 01:48:55,920
Overall, it's rather that it's being displaced. So the burden of risk is being pushed onto individuals.

1027
01:48:55,920 --> 01:49:04,380
And as individuals, we have much less resources to face some of these risks that maybe should be faced by society as a community.

1028
01:49:04,380 --> 01:49:10,280
So this is something that worries me. Thank you very much. Well.

1029
01:49:10,280 --> 01:49:15,030
I see, I see, I haven't done it yet. Very briefly, if you don't want to these question.

1030
01:49:15,030 --> 01:49:20,800
Yes, super short. Oh yeah. This would be a sliding the beginning of your.

1031
01:49:20,800 --> 01:49:27,450
What you can bring to this set of 30 exposed theological loop works of art of

1032
01:49:27,450 --> 01:49:34,350
those very much issues where machine learning and any I can tell at bones.

1033
01:49:34,350 --> 01:49:39,920
And how does that relate to your Arthur? Definitely, that's a very good point.

1034
01:49:39,920 --> 01:49:41,810
And you're right to push me on that.

1035
01:49:41,810 --> 01:49:49,310
And I think I can assist us in many things that can assist us in figuring out the empirical details, which may be very important for ethics.

1036
01:49:49,310 --> 01:49:58,400
But at the moment, as long as I is not sentenced, it's very questionable whether it can assess us in weighing that and making us decide.

1037
01:49:58,400 --> 01:50:06,530
In the end, how do how do we morally weigh these different things about bringing up these issues in matter?

1038
01:50:06,530 --> 01:50:17,400
This is my favourite philosopher. David Hume was fond of emphasising facts are one thing, but moral judgement arguably requires human sentiment.

1039
01:50:17,400 --> 01:50:30,780
On that note, let's thank all our speakers again.

1040
01:50:30,780 --> 01:50:37,650
This this has been quite an unusual seminar, but the main aim of it has been to let us all get to know each other,

1041
01:50:37,650 --> 01:50:41,910
to see some of the wonderful stuff that's going on around the university.

1042
01:50:41,910 --> 01:50:45,270
In subsequent seminars, we'll be digging more deeply.

1043
01:50:45,270 --> 01:50:53,040
It seems to me that this slide show technology is rather useful and in future will leave more time for questions and interaction.

1044
01:50:53,040 --> 01:50:57,180
Thank you very much for coming. Now onto the refreshments.

1045
01:50:57,180 --> 01:51:00,623
Thank you.