1
00:00:02,550 --> 00:00:09,340
Hey, to everyone joining from around the world, I see we have people from Japan, from California, from Switzerland,

2
00:00:09,340 --> 00:00:17,380
lots of time zones here, Seattle and even from West Berkshire in the U.K. and, of course, from Oxford.

3
00:00:17,380 --> 00:00:21,940
Good day to all of you all. I hope you keeping well and safe in these difficult times.

4
00:00:21,940 --> 00:00:28,960
No doubt you've heard the good news from Rocksmith that we have a vaccine that's going to be affordable and globally distributed.

5
00:00:28,960 --> 00:00:35,800
So this is a significant day for us in Oxford and made even more significant by the fact that we have a

6
00:00:35,800 --> 00:00:45,130
presentation from Gene UNEF and Gina is going to give a talk that she was originally scheduled to give in March.

7
00:00:45,130 --> 00:00:50,260
And then circumstances intervened and then we had covered 19.

8
00:00:50,260 --> 00:00:56,980
But she's here and we Apsley delighted that she could join us.

9
00:00:56,980 --> 00:01:00,910
I'm in Golden. I was the founding director of the Oxford Martin School.

10
00:01:00,910 --> 00:01:08,920
And since 2006, I have been directing the programme on technological and economic change at the school.

11
00:01:08,920 --> 00:01:18,520
Gina is the professor of technology and society at the Oxford Internet Institute, was actually with founding groups and the Oxford Martin School.

12
00:01:18,520 --> 00:01:22,810
She's also a professor in the Department of Sociology at Oxford.

13
00:01:22,810 --> 00:01:26,560
She works on the future of work and data rich environments.

14
00:01:26,560 --> 00:01:35,530
She's published three books and over a dozen research articles in many of the magazines and journals that we'll be reading.

15
00:01:35,530 --> 00:01:43,510
She's a pioneer in creating the area of human centred data science and leading a new project on the

16
00:01:43,510 --> 00:01:49,600
organisational challenges that companies face using artificial intelligence for decision making.

17
00:01:49,600 --> 00:01:56,350
She's had fellowships at the British Academy, the Institute for Advanced Study at Princeton and elsewhere.

18
00:01:56,350 --> 00:02:02,590
And she's an adviser to many companies, organisations and others on this agenda.

19
00:02:02,590 --> 00:02:08,950
She's also an advisor on Álex, the Women's Forum for Economy and Society.

20
00:02:08,950 --> 00:02:17,520
Gina. Delighted to have you. I'm here to talk about data work. This is the new project that my team and I have been working on.

21
00:02:17,520 --> 00:02:20,980
And it's going to I'm going to talk about three distinct projects.

22
00:02:20,980 --> 00:02:33,550
Now, the stakes at hand are that we are looking at a transformation of the basic everyday infrastructure of how we do organising.

23
00:02:33,550 --> 00:02:41,860
I like this word infrastructure because it reminds us from science and technology studies that once technology is going to put into place,

24
00:02:41,860 --> 00:02:46,110
they're often invisible or everyday.

25
00:02:46,110 --> 00:02:55,140
We don't so much think about our roads and the structure when we're trying to get it from point A to B in the west.

26
00:02:55,140 --> 00:03:01,500
In the north. The Northern Hemisphere, we don't think necessarily, and fortunately,

27
00:03:01,500 --> 00:03:10,140
we don't have to think about things like our water infrastructure and we think about our connectivity infrastructure because it's helped us allow.

28
00:03:10,140 --> 00:03:19,890
Allowed us to do. Allowed us to do things that we haven't we wouldn't be able to do otherwise during this pandemic.

29
00:03:19,890 --> 00:03:21,060
That said,

30
00:03:21,060 --> 00:03:29,820
infrastructure helps us think about how small technical decisions being made today end up having enormous everyday implications for all of us.

31
00:03:29,820 --> 00:03:36,240
And that's what I'd like to talk about. There's three myths of everyday A.I. that we need to address.

32
00:03:36,240 --> 00:03:45,930
The first is that A.I. is not. It's not a province anymore of large technology companies.

33
00:03:45,930 --> 00:03:54,720
That is that we are seeing the integration of new modes of organising new kinds of technologies into everyday decisions,

34
00:03:54,720 --> 00:04:07,590
from legal decisions to marketing decisions to civics decisions, decisions about who gets credit or who gets put in jail.

35
00:04:07,590 --> 00:04:14,070
And in this way, what the conversation so far about ethics has targeted those who make A.I.,

36
00:04:14,070 --> 00:04:18,640
those who those who the large technology firms who are driving the innovation.

37
00:04:18,640 --> 00:04:27,030
And so I would say that the context for us to think about in terms of data work is not that we're targeting large tech firms,

38
00:04:27,030 --> 00:04:32,280
but rather the places where many people on this call would work.

39
00:04:32,280 --> 00:04:41,430
Smaller firms, firms outside of technology firms, organisations of all sorts that are starting to integrate, purchase,

40
00:04:41,430 --> 00:04:49,260
acquire new kinds of data, processing, new kinds of technologies that make decisions and automate processes.

41
00:04:49,260 --> 00:04:58,920
The second kind of key challenge for us and thinking about data work is that we've talked about artificial intelligence as as a suite of technologies,

42
00:04:58,920 --> 00:05:07,590
as a smart tech, but fuelling and driving the expansion of artificial intelligence into more parts of our everyday lives.

43
00:05:07,590 --> 00:05:14,730
It's really this question of A.I. as a data so that, you know, if the first is A.I. is everywhere, it's in smaller firms.

44
00:05:14,730 --> 00:05:17,190
We need to we need to change our conversation.

45
00:05:17,190 --> 00:05:26,010
The second kind of key point that we need to think about is that we need to think of what is happening, not as just a technological phenomenon,

46
00:05:26,010 --> 00:05:35,790
although the processing power and cloud computing are allowing and enabling and driving new kinds of new ways of generating data.

47
00:05:35,790 --> 00:05:40,860
It is about a data question, right. And how data is being used in the Web.

48
00:05:40,860 --> 00:05:48,270
And that's going to be one of the key areas to look at tonight. The third is the idea that AI systems are automatic.

49
00:05:48,270 --> 00:05:53,820
Now, this seems almost like an oxymoron, but bear with me for a moment.

50
00:05:53,820 --> 00:06:04,350
When we talk about integrating artificial intelligence systems into everyday life, we might think of them as driving a cent of automation.

51
00:06:04,350 --> 00:06:15,000
And that's certainly the push behind many of these systems. But what is fuelling much of the expansion is a whole host of human work.

52
00:06:15,000 --> 00:06:22,990
And it's really in the intersection of these three points I of the every day and many different kinds of organisations.

53
00:06:22,990 --> 00:06:27,720
I as a as a data innovation and not just a tech innovation.

54
00:06:27,720 --> 00:06:33,000
And the notion that A.I. systems involve enormous amounts of human effort.

55
00:06:33,000 --> 00:06:38,480
That's really where we need to be thinking in terms of data work. And that's where our talk is tonight.

56
00:06:38,480 --> 00:06:46,880
So with my team. Maggie McGrath and Diana Prakash earlier this spring.

57
00:06:46,880 --> 00:06:50,180
Earlier this year, we were rereleased report called A at Work.

58
00:06:50,180 --> 00:07:01,730
And what we did was actually fairly simple. We surveyed global newspapers for accounts of how artificial intelligence was making the news.

59
00:07:01,730 --> 00:07:06,500
And first thing we did was set aside basically everything that looked like a company press release.

60
00:07:06,500 --> 00:07:11,810
Right. So those cases where someone was advertising how well their product worked,

61
00:07:11,810 --> 00:07:19,130
or we know from existing research that much of the journalism that is on artificial

62
00:07:19,130 --> 00:07:22,880
intelligence is coming from the sources of technology companies that make it.

63
00:07:22,880 --> 00:07:33,860
So we set those aside and instead we focussed on how news stories were covering, where I was working and organisations.

64
00:07:33,860 --> 00:07:41,280
And we came up with three common challenges that we saw in the about a thousand articles that met our criteria.

65
00:07:41,280 --> 00:07:49,100
And this over the course of the year. And those challenges were challenges of transparency, integration and reliance.

66
00:07:49,100 --> 00:07:51,500
That is what we began to see.

67
00:07:51,500 --> 00:08:03,380
If we if we take and and collate the stories of where A.I. is not quite living up yet to its potential as a tech, as a transformative technology.

68
00:08:03,380 --> 00:08:09,950
We see the stories, share these news stories, share three key challenges now in transparency.

69
00:08:09,950 --> 00:08:16,670
We see this lack of transparency between companies and customers about what I can do, how long it takes,

70
00:08:16,670 --> 00:08:23,630
how much work is involved in making it, and sort of closing the loop as as as they say.

71
00:08:23,630 --> 00:08:29,720
So companies on the one hand tout eyes this automating solution rather than being open about the amount

72
00:08:29,720 --> 00:08:36,200
of money or time or effort needed to produce and sustain systems that actually work in practise.

73
00:08:36,200 --> 00:08:41,570
And this led to a whole host of problems that that that arose in our reports,

74
00:08:41,570 --> 00:08:46,010
because we could see that that the people, the clients behind these systems felt deceived,

75
00:08:46,010 --> 00:08:55,730
they felt deceived by privacy, promises, etc. And companies felt that there was a whole host of unglamorous ways on the on the one hand,

76
00:08:55,730 --> 00:09:05,670
they had signed up for know, sparkling new systems. On the other hand, they were getting what looked like lots of back and data cleaning.

77
00:09:05,670 --> 00:09:13,410
Next, there was a huge gap in integration and we caught we use this to think about how what was the gap between

78
00:09:13,410 --> 00:09:20,910
the conditions under which I was trained in real life environments and how it's used in practise.

79
00:09:20,910 --> 00:09:30,870
And we we saw that in many systems, the systems that that that made the news that they somehow were were trained with one set of data.

80
00:09:30,870 --> 00:09:38,790
But unfortunately, the real life situations in which it was rolled out or integrated into were messier and less organised.

81
00:09:38,790 --> 00:09:49,260
And so it often meant that there were enormous amount of human labour used to train and manage the systems even when they were put in place.

82
00:09:49,260 --> 00:09:53,190
And that meant that companies often struggled to scale artificial intelligence

83
00:09:53,190 --> 00:09:59,090
system so they could work across a broad array of scenarios and problems.

84
00:09:59,090 --> 00:10:06,300
And that takes away or presents a problem for some of the notions of scale.

85
00:10:06,300 --> 00:10:12,960
And there was a notion of reliance, and we like to think of this as like companies were mine either too much on their eye

86
00:10:12,960 --> 00:10:19,650
system or the idea that an ISIS was in place and investing a lot of agency autonomy

87
00:10:19,650 --> 00:10:25,500
and authority into the systems rather than training their workforce staff or bringing

88
00:10:25,500 --> 00:10:31,740
them alongside and making those those decisions practical in the workplace.

89
00:10:31,740 --> 00:10:36,510
We also see that there's a did these challenges of both over and underlined.

90
00:10:36,510 --> 00:10:41,440
So let me take each one of these in turn. First is with transparency.

91
00:10:41,440 --> 00:10:52,960
Now, my colleague Mary Grey at Microsoft Research and her colleague said Hearths, I'm sorry, have just released a wonderful book,

92
00:10:52,960 --> 00:11:00,750
goes to work on the platform work of labouring labelling so that when we see a systems,

93
00:11:00,750 --> 00:11:09,390
we know that there's a lot of contract gig or on demand platform labour that goes in to building these systems.

94
00:11:09,390 --> 00:11:20,550
And often often that kind of transparency is is a problem that we don't see who's involved and the work that's involved in making these systems work.

95
00:11:20,550 --> 00:11:25,620
However, we in our report kind of pushed on a different idea of transparency,

96
00:11:25,620 --> 00:11:32,820
the notion that companies often had different kinds of labour in different parts

97
00:11:32,820 --> 00:11:36,810
of the organisation that were really important in making their systems happen.

98
00:11:36,810 --> 00:11:44,550
And so and so in this sense, the transparency was both. Is the system being transparent with with where work is happening?

99
00:11:44,550 --> 00:11:51,240
Who is doing the work also geographically? Is the work being done within the company or outsourced to a third?

100
00:11:51,240 --> 00:11:57,930
And again, several of the cases found where even internally within organisations,

101
00:11:57,930 --> 00:12:04,080
they would think that a process was happening inside their organisation when it was really happening outside.

102
00:12:04,080 --> 00:12:07,260
The next is this kind of thinking about integration. Right.

103
00:12:07,260 --> 00:12:14,160
And example that how systems actually integrate into existing workplaces really presented a lot of problems.

104
00:12:14,160 --> 00:12:23,210
So we saw several cases where there was problems scaling data across multiple hospitals where data would be collected in one hospital,

105
00:12:23,210 --> 00:12:30,420
a system and a system would be trained. But then bringing that to be able to make sense in another hospital or scaling

106
00:12:30,420 --> 00:12:34,830
it even across different departments of an organisation became a problem.

107
00:12:34,830 --> 00:12:40,470
So we see these challenges of integration really as when data moves from one context to another.

108
00:12:40,470 --> 00:12:48,570
And companies, they can't quite fit the way I they eye into their existing organisation and an existing strategy.

109
00:12:48,570 --> 00:12:54,870
And then third was this notion of reliance. Right. So when was the dependence on I just.

110
00:12:54,870 --> 00:12:57,900
Right. Neither sort of a Goldilocks moment.

111
00:12:57,900 --> 00:13:08,610
And there were several challenges where companies were were touting that their system was was doing one thing, but really.

112
00:13:08,610 --> 00:13:15,360
Well, it was a it was a shield. And what was being done was people in behind the system.

113
00:13:15,360 --> 00:13:23,340
So in countries like China and India, there were crowd sourced to labour actually providing the work.

114
00:13:23,340 --> 00:13:27,030
So it was weak data and privacy laws and cheap labour,

115
00:13:27,030 --> 00:13:34,950
this kind of intersection that really gave different places on the globe a competitive advantage in this kind of work.

116
00:13:34,950 --> 00:13:42,630
And and so, therefore, the the the reliance isn't so much on the technical system, but rather on a new form of outsourcing.

117
00:13:42,630 --> 00:13:44,490
So what does this mean for data work?

118
00:13:44,490 --> 00:13:53,070
Well, we define data work as the work that needs to happen in terms of helping the entire interpretation and contextualisation of data.

119
00:13:53,070 --> 00:14:02,700
In practise, that's first. Second data work is really the work that is involved in helping translate data for fairness and inclusion.

120
00:14:02,700 --> 00:14:10,300
So that, too, needs some kind of context around a system.

121
00:14:10,300 --> 00:14:16,980
And third, that we define data work is the communication that needs to happen with a whole host of stakeholders,

122
00:14:16,980 --> 00:14:21,360
often including conversations about privacy and ethics.

123
00:14:21,360 --> 00:14:26,730
And so this comes from a paper that we published earlier this spring.

124
00:14:26,730 --> 00:14:31,530
Who does the work of data and ACM interactions? And that's with my colleagues.

125
00:14:31,530 --> 00:14:40,570
Mueller, Boesen, Pine, Nielsen and myself. And and in this, we we we we ask three main questions.

126
00:14:40,570 --> 00:14:43,890
Right. Who if we're going to understand this hidden,

127
00:14:43,890 --> 00:14:53,580
invisible lead labour of data work that is happening within organisations preparing organisations for A.I. systems.

128
00:14:53,580 --> 00:14:57,940
We need to ask first who is ensuring that the data are meaningful?

129
00:14:57,940 --> 00:15:05,970
Who's doing the work that helps integrate data solutions into practise within the organisation?

130
00:15:05,970 --> 00:15:13,980
So who's taking the responses? So the sorting and making sure that systems are organised around it.

131
00:15:13,980 --> 00:15:20,370
Second, who's really doing the work of organising an infrastructure thing for making it even possible?

132
00:15:20,370 --> 00:15:24,870
And so much of what we've seen in terms of what Grey and sorry,

133
00:15:24,870 --> 00:15:32,310
called ghost work in terms of some of the platform labour really is that a lot of the organising an infrastructure in work is hidden labour.

134
00:15:32,310 --> 00:15:39,300
But even within organisations, we see that there's work around organising an infrastructure that needs to happen.

135
00:15:39,300 --> 00:15:47,730
And then finally, who is doing this work that is attending to questions of ethics, privacy and people's concerns about their data?

136
00:15:47,730 --> 00:15:52,290
Again, difficult to automate that particular part of work.

137
00:15:52,290 --> 00:15:56,130
And so we took these three questions and we looked at two sets of hospitals,

138
00:15:56,130 --> 00:16:06,990
a set of work in hospitals around creating Redding billing data in the United States for analysis.

139
00:16:06,990 --> 00:16:12,630
So as many people on the call will know, in the United States with privatised healthcare system,

140
00:16:12,630 --> 00:16:22,560
each procedure within a healthcare setting has a particular billing code, and those billing codes can be aggregated to provide real sense.

141
00:16:22,560 --> 00:16:30,240
How long did someone stay in hospital? What what kinds of treatments end up with?

142
00:16:30,240 --> 00:16:39,840
What kinds of outcomes? We'll see, we looked both at the United States and in Denmark, where a site where a set of workers,

143
00:16:39,840 --> 00:16:47,910
medical secretaries do the work of transcribing and making certain kinds of data meaningful within the health care setting.

144
00:16:47,910 --> 00:16:55,380
And so by doing those two and this is where we're really missing out on the slides, we come up with a data,

145
00:16:55,380 --> 00:17:03,540
will we come up with a way that we can model where intelligibility and transparency?

146
00:17:03,540 --> 00:17:11,370
Ongoing optimisation of resources and the work of context, information and metadata occur within the organisation.

147
00:17:11,370 --> 00:17:23,040
And so we think about these three really primary types of work, both as new kinds of work that's necessary to make A.I. systems even function,

148
00:17:23,040 --> 00:17:30,990
but also as recognising work that we might not have seen as crucial in the data driven revolution.

149
00:17:30,990 --> 00:17:40,140
So, for example, are building codes, experts and hospitals in Denmark and the United States were both really key and

150
00:17:40,140 --> 00:17:46,770
making sure that certain kinds of data were ready for data scientists to process.

151
00:17:46,770 --> 00:17:54,000
They were really important for making sure that the right codes got attached to the right procedures.

152
00:17:54,000 --> 00:18:01,350
And that was a kind of infrastructure work, a sort of care of data work that opens up new possibilities for analysis.

153
00:18:01,350 --> 00:18:06,510
And third, they were the ones who interfaced with people who had concerns about whether or

154
00:18:06,510 --> 00:18:13,400
not the codes about them were right or had concerns or questions about privacy.

155
00:18:13,400 --> 00:18:22,460
So before we open up for discussion, and I promise, since we're doing this without slide's, I'm going to I'm going to bring us to a close very soon.

156
00:18:22,460 --> 00:18:32,320
I will just say that the third project that my teams worked on in this scheme and really where I see both opportunity and

157
00:18:32,320 --> 00:18:41,360
a great opportunity for people to get involved is around the discussion of what is and is an artificial intelligence.

158
00:18:41,360 --> 00:18:51,770
So helping people make sense of the systems that we see and unpacking some of the myths about automated work and the future of work,

159
00:18:51,770 --> 00:19:03,770
I think is one of the things that we need to be working on. So with my team early in this in 2020, we released the A to Z of A I or the A to Z of a I.

160
00:19:03,770 --> 00:19:12,560
And if you Google that, you will find it. We partnered with Google to create an educational product that really helps people with a

161
00:19:12,560 --> 00:19:18,200
buy a simple bite sized explainers to help people understand what a is and how it works.

162
00:19:18,200 --> 00:19:25,190
And that and that project has now been rolled out in fifty nine different countries, in 13 languages.

163
00:19:25,190 --> 00:19:32,330
And it's part of an effort to kind of tease out some of the misperceptions that we have.

164
00:19:32,330 --> 00:19:40,370
So I have a series of practise and a series of policy research recommendations.

165
00:19:40,370 --> 00:19:43,220
Before we wrap up.

166
00:19:43,220 --> 00:19:51,350
And if this were slide enabled, if this weren't 20, 20 and the slide and they talk, we would leave these up as we have the conversation.

167
00:19:51,350 --> 00:19:59,080
But I just want to say that for a practise agenda, we really need to be thinking about organising for data saturated societies.

168
00:19:59,080 --> 00:20:07,090
These questions around artificial intelligence. And who does the work and making AI systems even possible.

169
00:20:07,090 --> 00:20:11,350
That raises questions of digital data and the public good,

170
00:20:11,350 --> 00:20:17,770
and it helps us think about projects that seek to understand how we can harness some kinds of data.

171
00:20:17,770 --> 00:20:21,490
Commercial data, perhaps for responsible reuse.

172
00:20:21,490 --> 00:20:31,660
It helps us think about how we want to intervene and think through those questions of who is advocating for the data of data subjects.

173
00:20:31,660 --> 00:20:36,580
Second, it reminds us that people's understanding of their privacy and their own data is

174
00:20:36,580 --> 00:20:42,030
not a task that we can just leave and assume that people will do it on their own,

175
00:20:42,030 --> 00:20:50,650
that there actually is quite a bit of organising and stakeholder engagement work that needs to be done both on the part of companies doing the work,

176
00:20:50,650 --> 00:21:00,700
but also for those of us who advocate for responsible use of data. We really need to be working in this intersection of helping to people to to to

177
00:21:00,700 --> 00:21:04,930
learn to interrogate the values and implications of data driven businesses.

178
00:21:04,930 --> 00:21:12,880
And then finally, third, I mean, this is part of this is we really need to empower citizens and upscale societies as part of this.

179
00:21:12,880 --> 00:21:19,840
So so the work of making AI systems is simply too important to be left to simply large technology

180
00:21:19,840 --> 00:21:25,870
companies who have an interest in as as as vendors of selling systems on to small companies.

181
00:21:25,870 --> 00:21:32,080
We really need to support this responsible utilisation of data, not just by thinking about who's designing systems,

182
00:21:32,080 --> 00:21:36,280
but helping people who are going to be using them on an ongoing, ongoing method.

183
00:21:36,280 --> 00:21:43,860
And that, of course, brings up the policy questions of how we can enable policymakers to up skill as well.

184
00:21:43,860 --> 00:21:53,740
And finally, in terms of in terms of research agenda, I think this applied everyday infrastructure is one way that we can begin to think about

185
00:21:53,740 --> 00:22:00,340
and move A.I. and ethics questions out of the realm of large scale technology makers,

186
00:22:00,340 --> 00:22:06,190
but really start to think about how we map, track and measure these changes and all of our lives.

187
00:22:06,190 --> 00:22:14,850
What's happening in the organisational settings where we work and how can we see that we're moving from

188
00:22:14,850 --> 00:22:21,670
these these questions of a more and more data gathered about us and managed about us in different ways?

189
00:22:21,670 --> 00:22:34,720
Can we identify ways that social, cultural, organisational and causal factor factors really shape who can intervene and hold accountable A.I. systems?

190
00:22:34,720 --> 00:22:44,400
So where can we do the social science that helps us ensure that systems are deployed and integrated in responsible ways?

191
00:22:44,400 --> 00:22:52,960
And then can we begin to think about the changing social norms and conventions that are happening around AI systems and organisations?

192
00:22:52,960 --> 00:23:00,430
When do we cede power to automated systems and when do we remember that behind the interface of

193
00:23:00,430 --> 00:23:05,590
many of these systems is a whole host of human labour that also needs to be held accountable.

194
00:23:05,590 --> 00:23:16,960
And so with that, I indulge. Thank you all for indulging me with my technical issues this evening and invite and come in and join our conversation.

195
00:23:16,960 --> 00:23:23,590
Thanks very much, Gina, for that admirably clear presentation.

196
00:23:23,590 --> 00:23:35,080
I love your organisation. Of Everything in 3D, which is incredibly helpful, is where every training is so that people remember threes.

197
00:23:35,080 --> 00:23:42,100
It helps when you have no slides to follow. However, it is impressive that you remembered all the 3D as well.

198
00:23:42,100 --> 00:23:50,330
Admirably clear and and really extremely urgently needed, because without knowing it,

199
00:23:50,330 --> 00:23:56,680
we are all walking into this maze and sometimes catastrophically so,

200
00:23:56,680 --> 00:24:05,500
not least in the UK government's use of Excel spreadsheets it and understand for track and trace.

201
00:24:05,500 --> 00:24:10,600
I have a number of questions, but I am conscious of the time and we did start a little bit late.

202
00:24:10,600 --> 00:24:18,920
So let me let me just impose a few. And having been in some of these data factories in Kenya, the Samasource and others,

203
00:24:18,920 --> 00:24:25,480
and and also admire the work that your colleague Mark Graham stop and thinking about the rights of people,

204
00:24:25,480 --> 00:24:31,660
which is very allied to the work you're doing. No doubt this is maybe just beginning.

205
00:24:31,660 --> 00:24:40,560
We've got your book that argues that I am at the bottom, but some of what you were speaking about didn't sound too much like a guide.

206
00:24:40,560 --> 00:24:51,190
To me, it sounds more like filling in Excel spreadsheets. Is is that a slippery slope or, you know, without reading a book?

207
00:24:51,190 --> 00:24:59,020
Can you give us a quick definition of of a. But I do think all the participants to get the book obviously answered to look at the deeper explanation.

208
00:24:59,020 --> 00:25:05,440
Thank you. The. So is it is it is.

209
00:25:05,440 --> 00:25:11,770
It just excels. There's a there's a joke that says when it's a as sales or consultant,

210
00:25:11,770 --> 00:25:18,430
a as a spreadsheet, when it's someone who's a data scientist, it's machine learning.

211
00:25:18,430 --> 00:25:24,850
That's absolutely true here. When we look at how companies are talking about what it is that they're doing.

212
00:25:24,850 --> 00:25:40,720
They they they they they put they put a gloss that allows more computation, that infers much more computational power than they're actually doing.

213
00:25:40,720 --> 00:25:44,860
So with the Women's Forum, you mentioned the beginning in the introduction.

214
00:25:44,860 --> 00:25:53,680
I've been doing this autumn a series of focus groups with chief data scientists in in companies in Europe and the US.

215
00:25:53,680 --> 00:25:55,930
And these are incredible company leaders. Right.

216
00:25:55,930 --> 00:26:06,310
These are Fortune 500 companies, banks, large manufacturing firms, large consumers, consumer services firms.

217
00:26:06,310 --> 00:26:13,030
Named the sector we've had we've had an interview with someone working in their data science team.

218
00:26:13,030 --> 00:26:19,510
And we ask the primary question, what are you doing about responsible A.I. ethics?

219
00:26:19,510 --> 00:26:30,520
What are you doing for responsibility in your data systems? And each each data scientist knows they have a huge responsibility to do.

220
00:26:30,520 --> 00:26:37,660
And none of them can articulate yet what it is they should be doing on a practical level.

221
00:26:37,660 --> 00:26:43,450
So there's an enormous amount of Catch-Up that our corporate leaders are doing in

222
00:26:43,450 --> 00:26:48,220
terms of figuring out how to put into practise something that actually works.

223
00:26:48,220 --> 00:26:56,230
So, sure, some of the some of the challenges that we're looking at are challenges of any kinds of large scale centralisation,

224
00:26:56,230 --> 00:27:00,190
large scale control over globalised systems and supply chains.

225
00:27:00,190 --> 00:27:09,460
But the challenges here are that we risk building infrastructures that once put into place as technical data.

226
00:27:09,460 --> 00:27:16,240
Infrastructures become difficult to untangle, difficult to hold accountable and difficult to intervene in.

227
00:27:16,240 --> 00:27:22,530
And so that's why we are suggesting that there's an urgency at this moment for really kind of.

228
00:27:22,530 --> 00:27:29,180
Guy getting getting getting the conversation more involved in holding these systems accountable.

229
00:27:29,180 --> 00:27:37,580
Yes. And I think you you very clearly articulated the urgency of this and the need.

230
00:27:37,580 --> 00:27:42,710
I have many questions, but let's go to some of the questions, the questions that have been posed by participants.

231
00:27:42,710 --> 00:27:47,120
You are able to vote for these questions, so to vote.

232
00:27:47,120 --> 00:27:53,330
If you if you're keen on a particular question and I see we have eight.

233
00:27:53,330 --> 00:27:58,070
Let me take the first, which is from Ali Steadman, who I happen to know.

234
00:27:58,070 --> 00:28:05,300
Hi, Ali. Good to see that you are participating in this. What does the equity of opportunity in tech look like?

235
00:28:05,300 --> 00:28:12,140
Where do we aim and how do we know when we've reached it? That is a fabulous question, because on the one hand,

236
00:28:12,140 --> 00:28:19,970
we want to see the expansion of the types of people involved in designing and building A.I.S systems for the world.

237
00:28:19,970 --> 00:28:23,960
Right now, we see a concentration of that effort in the global north.

238
00:28:23,960 --> 00:28:31,940
We see it in the US and in Europe. We see that only 18 percent of people working in Azi are women.

239
00:28:31,940 --> 00:28:41,660
So we have we have enormous challenges of racial, ethnic, global gender diversity in building the system.

240
00:28:41,660 --> 00:28:49,490
On top of that, we need to start building capacity. And I would suggest that there's a couple of challenges.

241
00:28:49,490 --> 00:28:57,260
One is that the the equity of opportunity is necessary for building better tech, but it's not sufficient.

242
00:28:57,260 --> 00:29:00,350
It's not the only thing that's going to get us there.

243
00:29:00,350 --> 00:29:10,310
The second would be that we really need to increase the capacity in the global south in order to ensure that systems are properly localised.

244
00:29:10,310 --> 00:29:13,850
So if we are just relying on systems that are built in one place,

245
00:29:13,850 --> 00:29:22,750
trained on data from one place and then integrated into into the systems around the world, that's that's a recipe for disaster.

246
00:29:22,750 --> 00:29:29,920
Yes, integrating bias and a massive concerns, not least Shakir Mohammed.

247
00:29:29,920 --> 00:29:40,720
Deep minds been writing very admirably about the colonialism of data and data algorithms for votes for doxa rulers.

248
00:29:40,720 --> 00:29:44,940
Question. Can you comment on technological determinism? Is I.

249
00:29:44,940 --> 00:29:48,490
Trends such as those discussed. Yeah.

250
00:29:48,490 --> 00:29:55,660
Susceptible to the fallacy of technological determinism. That's a fantastic question.

251
00:29:55,660 --> 00:30:06,280
We certainly hear that determinism in how industry leaders talk about and support and and push the inevitability of A.I.

252
00:30:06,280 --> 00:30:18,270
So if we look at this kind of data work question, we where we see people in these large and these large hospital settings,

253
00:30:18,270 --> 00:30:26,560
the people who are involved in the day to day operations of getting data systems ready for data analysis.

254
00:30:26,560 --> 00:30:30,850
There's a whole host of new kinds of jobs that are being involved.

255
00:30:30,850 --> 00:30:37,930
It's not inevitable that that it automates work or that it displaces all kinds of work.

256
00:30:37,930 --> 00:30:46,420
But instead it's creating these new moments with with with different kinds of opportunities for people to be involved.

257
00:30:46,420 --> 00:30:49,840
And so and so when I think about the technological determinism,

258
00:30:49,840 --> 00:30:56,950
I think the technology has its own drive and it will naturally kind of go in one way or another.

259
00:30:56,950 --> 00:31:05,050
Whereas whereas what we can what we can see from how from from from older industries like healthcare and construction

260
00:31:05,050 --> 00:31:13,170
to that I know very well in this case that the pathway that new technologies take is not predetermined at all.

261
00:31:13,170 --> 00:31:22,090
So so I think we have I think we as educators have work cut out for us where we are.

262
00:31:22,090 --> 00:31:31,380
We make sure that how the technology industry talks about the inevitability of their genius is held to account.

263
00:31:31,380 --> 00:31:37,990
Absolutely. So there's also people on YouTube participating and we've been since.

264
00:31:37,990 --> 00:31:43,280
A couple of questions that have come across from that.

265
00:31:43,280 --> 00:31:51,240
And the first one which seems to disappear is I was wondering, this is doesn't say who it's from.

266
00:31:51,240 --> 00:31:55,000
Andre Grisman. I was wondering how Prof.

267
00:31:55,000 --> 00:32:03,100
Mathaf thinks data saturation in a rich healthcare will impact medical education in the next five years.

268
00:32:03,100 --> 00:32:09,100
Brilliant question from a brilliant colleague in the United States.

269
00:32:09,100 --> 00:32:23,350
So we absolutely have to use these notions of data work to help people understand the context of that that results they're looking at.

270
00:32:23,350 --> 00:32:34,570
If we continue to think about the outputs of A.I.S systems as decontextualised outputs, we end up in a dangerous place, especially in medical systems.

271
00:32:34,570 --> 00:32:44,770
If we if we understand where the data come from, what kind of context they have and and who is advocating and who is responsible for ensuring that

272
00:32:44,770 --> 00:32:52,600
there is the organisational stitching together between the outputs and the the the the organisation,

273
00:32:52,600 --> 00:33:01,690
the practises within the organisation. We would have much more highly contextualised and much more relevant results.

274
00:33:01,690 --> 00:33:10,930
It's life and death consequences and health care. So where we are right now, today, in this moment where lot we are, we have a lot to be grateful for.

275
00:33:10,930 --> 00:33:17,290
Large scale computing power and the integration of global supply chains is going to help

276
00:33:17,290 --> 00:33:23,330
us end this global pandemic sooner than than than any pandemics ever ended before.

277
00:33:23,330 --> 00:33:32,680
Right. We we we should be cheering large scale computing power and the collaboration possible from teams in this particular way.

278
00:33:32,680 --> 00:33:45,760
But we're not going to get to those great societally beneficial outcomes unless we realise that these data systems are highly dependent,

279
00:33:45,760 --> 00:33:55,150
highly contextualised and sometimes highly fragile. Many of the cases that we looked at in our work report come from healthcare,

280
00:33:55,150 --> 00:34:03,040
where data from one context was simply tried to be brought into another context of data gathered in one particular

281
00:34:03,040 --> 00:34:11,560
hospital reflected choices that that that were very specific to that hospital and not applicable to to other hospitals.

282
00:34:11,560 --> 00:34:14,530
So how does this influence medical education?

283
00:34:14,530 --> 00:34:22,460
We have to train people in health care who are going to be using these systems on how to use them responsibly, how to be critical.

284
00:34:22,460 --> 00:34:28,320
Tumours of the data they're using. They are our front line defence for bad.

285
00:34:28,320 --> 00:34:33,470
And when a guy goes awry. In fact, that was from someone else.

286
00:34:33,470 --> 00:34:39,800
But Andre Guzman said they'd like to extend this this discussion from medical schools to business schools,

287
00:34:39,800 --> 00:34:48,980
etc., what should be the core of an education generally related to a Ayane data?

288
00:34:48,980 --> 00:34:56,510
Carl Broten, Bergström and Jevin West at the University of Washington have a wonderful new book out called Calling B.S.,

289
00:34:56,510 --> 00:35:02,570
where they developed a training course around critical human centred data science.

290
00:35:02,570 --> 00:35:09,980
And they basically are training people to say, wait a second, I am calling out and I won't use the profanity.

291
00:35:09,980 --> 00:35:16,940
I'm calling out the as wrong as patently wrong.

292
00:35:16,940 --> 00:35:20,960
Some of the data, some of the seemingly data driven evidence here,

293
00:35:20,960 --> 00:35:29,210
I think it's incredibly important in this particular moment when we see attacks and challenges on science and evidence,

294
00:35:29,210 --> 00:35:30,080
we need to stand up,

295
00:35:30,080 --> 00:35:39,050
on the one hand for science and evidence and help ensure that we continue to build public trust in good science and good evidence.

296
00:35:39,050 --> 00:35:42,860
But on the other hand, building systems,

297
00:35:42,860 --> 00:35:50,870
building fragile A.I. systems that are not robust or they don't deliver as promised isn't helping us get there.

298
00:35:50,870 --> 00:35:57,650
It's if we want to make sure that we're not simply selling new 21st century versions of snake oil.

299
00:35:57,650 --> 00:36:02,570
And so I think it does happen where I think one of the solutions,

300
00:36:02,570 --> 00:36:09,080
just like we would train medical professionals and knowing how to push back on and query A.I.S systems,

301
00:36:09,080 --> 00:36:13,970
they don't need to know enough to design them, but they need to know how to how to operate them.

302
00:36:13,970 --> 00:36:18,440
It's gonna be the same in business schools. We're not necessarily going to have that.

303
00:36:18,440 --> 00:36:23,420
The MBA is in the CEO's necessarily as people designing systems,

304
00:36:23,420 --> 00:36:31,070
but they're going to need to know enough to ask their team and to hold them account for how the systems get integrated into their existing practises.

305
00:36:31,070 --> 00:36:35,660
And that's the education I think, that we really need to be doing right now. Absolutely.

306
00:36:35,660 --> 00:36:41,570
Including in Oxford. There's a reporter asks with three votes.

307
00:36:41,570 --> 00:36:48,980
What specific regulatory mechanisms would you like to see you put in place, put in place to address these issues?

308
00:36:48,980 --> 00:36:52,460
We've already started to see questions around people's individual data. Right.

309
00:36:52,460 --> 00:37:01,580
How do we how do we advocate for ourselves, you know, with the European GDP are, for example, general data protection.

310
00:37:01,580 --> 00:37:06,500
How do we how do we get individuals to advocate for this?

311
00:37:06,500 --> 00:37:15,710
We some of my esteemed colleagues have called for a kind of algorithm regulation.

312
00:37:15,710 --> 00:37:21,850
Right. That we that we that we regulate particular kinds of data systems and data structures.

313
00:37:21,850 --> 00:37:31,760
I don't necessarily think that's the path because we have in place enough outcomes regulation.

314
00:37:31,760 --> 00:37:34,700
So, for example, in the education system,

315
00:37:34,700 --> 00:37:43,520
we want to make sure that data systems are not failing our most disadvantaged students and that they're treating people equitably.

316
00:37:43,520 --> 00:37:47,180
We have we have ways of monitoring that.

317
00:37:47,180 --> 00:37:54,110
When we met with company leaders this autumn with the Women's Forum asking data scientists what they're doing about responsible tect,

318
00:37:54,110 --> 00:38:04,020
we found, surprisingly, that some of the most advanced conversations around responsible tech were happening in banking.

319
00:38:04,020 --> 00:38:08,510
Why would that be? Why would it be in banking and not in technology, for example, on banking?

320
00:38:08,510 --> 00:38:15,620
They already are under such a highly regulated, regulated system around how they make their decisions and choices that they

321
00:38:15,620 --> 00:38:21,470
had to be really sure when they integrated new systems into their analysis,

322
00:38:21,470 --> 00:38:27,800
that they could explain them, that they could explain them to the customers, that they could explain them to them themselves,

323
00:38:27,800 --> 00:38:34,570
and that they could be assured that these systems were not causing new forms of discrimination.

324
00:38:34,570 --> 00:38:42,890
It wasn't because that there was a special banking regulation in place, but it's because we have certain regulations around our financial data.

325
00:38:42,890 --> 00:38:49,340
And so and so I think that that's the the the way it's a framework, as it were,

326
00:38:49,340 --> 00:38:53,150
for how policymakers need to be thinking about regulation in this space.

327
00:38:53,150 --> 00:39:00,710
We need to be thinking about what existing frameworks in each realm of our our our lives as citizens.

328
00:39:00,710 --> 00:39:06,290
Do we need to be adjusting to think through how these systems are going to integrate into that?

329
00:39:06,290 --> 00:39:13,210
Because that's the infrastructure. So we're building. Right.

330
00:39:13,210 --> 00:39:22,890
There are three votes for this question from Gwen in your research, have you spoken to employees interacting with data driven system?

331
00:39:22,890 --> 00:39:26,800
If so, what are their main concerns and challenges in their work?

332
00:39:26,800 --> 00:39:32,880
And how do these differ by race, gender or age?

333
00:39:32,880 --> 00:39:38,820
In the data work article? In the article, who does the work of data that we published over the summer?

334
00:39:38,820 --> 00:39:44,310
We spoke overwhelmingly to women and in the United States it was women of colour who were

335
00:39:44,310 --> 00:39:53,160
having who had very stable bureaucratic jobs and large hospitals doing the state of work.

336
00:39:53,160 --> 00:39:58,230
They are the champions of the Azi era, large scale data.

337
00:39:58,230 --> 00:40:06,960
And those hot data analysis in those hospitals did not get done without these women who worked in literally the back offices.

338
00:40:06,960 --> 00:40:16,980
The challenges with our medical secretaries in in in Denmark was that the hospitals where the we interviewed,

339
00:40:16,980 --> 00:40:23,490
they were actually looking at automating their work because some of the some of the work that they had done up transcription was being automated.

340
00:40:23,490 --> 00:40:32,280
And so their concerns and challenges really were, how can I can how can I keep doing this work that I conceive of as care,

341
00:40:32,280 --> 00:40:35,850
as caring for our patients data, as caring for others,

342
00:40:35,850 --> 00:40:44,370
as making sure that the record stands properly while not getting recognised for doing that work.

343
00:40:44,370 --> 00:40:52,200
So the challenges really were how how can how can we support this invisible labour?

344
00:40:52,200 --> 00:40:56,460
It's not being called a I work. It's not being called data science.

345
00:40:56,460 --> 00:41:01,590
And yet neither new systems nor data science can happen without it.

346
00:41:01,590 --> 00:41:08,640
So I think the main concern and challenge is making sure that we're supporting good organise.

347
00:41:08,640 --> 00:41:16,290
We're supporting the work that needs to happen within organisations to make sure this happens.

348
00:41:16,290 --> 00:41:20,460
Is a cold. What's the organisation that Mark Bram's involved in as well?

349
00:41:20,460 --> 00:41:28,620
That's right. So Mark Gramme, my calling Mark Ram is working on Fair Work and Fair is is looking at the platform economy,

350
00:41:28,620 --> 00:41:34,260
labour rights, so, so, so much of this work. Mark, the work of Mark Graham, the work of Mary Grey.

351
00:41:34,260 --> 00:41:41,370
Much of this is looking at platform based labour. We're not seeing so much platform based labour in banks and hospitals.

352
00:41:41,370 --> 00:41:50,780
And yet and yet that the work that's being done in those organisations is absolutely about the invisible work that's fuelling.

353
00:41:50,780 --> 00:42:02,490
Yeah. So another question from all of this has two votes, which is one another vote, which is what I'm going to give him a second question.

354
00:42:02,490 --> 00:42:08,790
How do you regard human level performance as a metric to beat in machine learning applications,

355
00:42:08,790 --> 00:42:13,770
especially in manufacturing, checking for defects and so on?

356
00:42:13,770 --> 00:42:19,660
And he says Andrew Migg and I wrote recently about its limitations.

357
00:42:19,660 --> 00:42:28,380
Right. So, listen, we know that large scale computing power is going to help humans solve really big problems.

358
00:42:28,380 --> 00:42:34,290
And there are really big problems we cannot solve without large scale human computing power.

359
00:42:34,290 --> 00:42:41,790
I think if we frame artificial intelligence as the core, Larry, or the competitor to human intelligence,

360
00:42:41,790 --> 00:42:46,410
we we we get to these questions, will, which is which is better.

361
00:42:46,410 --> 00:42:52,650
We don't actually ask the question which is better, my calculator or my set of hand calculations.

362
00:42:52,650 --> 00:42:57,060
OK, sure. My calculator can answer some things faster.

363
00:42:57,060 --> 00:43:03,360
I can I can do some other things easier. That's going to continue to be the case with artificial intelligence system.

364
00:43:03,360 --> 00:43:10,770
So if we think about if we think about using the spray human level performance,

365
00:43:10,770 --> 00:43:21,820
we again are pushing A.I. systems to act and think like people and not act and think like bits of technology that we use or.

366
00:43:21,820 --> 00:43:32,290
Bits of infrastructure that are going to be part of much bigger organisations and data ecologies and flows of data that that already exist.

367
00:43:32,290 --> 00:43:50,710
So so to personify in some ways is to hand over a powerful sensemaking tool that takes the power away from how we can intervene in these systems.

368
00:43:50,710 --> 00:43:55,580
Thank you, ask this question, but it's not a question of Sujata.

369
00:43:55,580 --> 00:44:03,500
Can you comment on deep sake's or the more concerning the general ábhar issues you've discussed so far?

370
00:44:03,500 --> 00:44:10,640
Listen. You know, I'm concerned about deep fakes almost as much as everybody else.

371
00:44:10,640 --> 00:44:18,680
We have at the moment big challenges to add to the notions of information.

372
00:44:18,680 --> 00:44:23,600
Is our information secure or do we how do we know what we know?

373
00:44:23,600 --> 00:44:27,800
And this is a this is a question about it. It's a pistol ecology.

374
00:44:27,800 --> 00:44:37,340
This is a question about how we know what is real. That is very much a function of where we are in the early 21st century.

375
00:44:37,340 --> 00:44:50,180
But that said, I am more concerned about ensuring that we have a healthy, robust new system, healthy, robust democracies,

376
00:44:50,180 --> 00:44:58,570
healthy, robust organisations than I am about a particular technology of having lifelike human pictures.

377
00:44:58,570 --> 00:45:11,900
Right. So so we know, for example, that the amount of misinformation circulating about elections on Facebook varies by country.

378
00:45:11,900 --> 00:45:20,300
And we know that's a function, not a Facebook, not of the people in one country being smarter or easier to do,

379
00:45:20,300 --> 00:45:26,210
but a function of the regulatory environment and the social environment in which those elections happen.

380
00:45:26,210 --> 00:45:30,170
And so and so on the one hand, we can worry about deep fakes.

381
00:45:30,170 --> 00:45:36,560
On the other hand, we really need to be thinking about how do we shoulder the responsibility and bolster

382
00:45:36,560 --> 00:45:42,990
our social institutions and organisations to make sure that we're supporting a society.

383
00:45:42,990 --> 00:45:50,450
That's a great point to end on, and I'm afraid we've come back to our which is raced by.

384
00:45:50,450 --> 00:45:54,260
Sorry to all the participants for the technical glitches at the beginning.

385
00:45:54,260 --> 00:45:59,780
That's part of this transition to a digital world.

386
00:45:59,780 --> 00:46:09,290
It's great to see that so many of you have joined this, some of you extremely late at night or early in the morning.

387
00:46:09,290 --> 00:46:20,330
And thanks so much to Gina for enlightening us on what is an immensely important topic in a way that I found exceptionally clear.

388
00:46:20,330 --> 00:46:26,720
There's a lot talked about in this area, and I found your presentation to really cut through a lot of it.

389
00:46:26,720 --> 00:46:30,440
You can follow Gina on at Gina soon.

390
00:46:30,440 --> 00:46:34,550
You can follow me on an underscore golden Geraldine.

391
00:46:34,550 --> 00:46:39,590
I n do look at the Oxford Martin school events page.

392
00:46:39,590 --> 00:46:46,580
The next event, which is really a must for anyone that's interested in the Internet.

393
00:46:46,580 --> 00:46:57,320
We wouldn't be here without this man. Interconnect is Sir Tim Berners Lee in conversation with some.

394
00:46:57,320 --> 00:47:03,650
Who was another founder, together with some of much of the computing private, Sir Tim,

395
00:47:03,650 --> 00:47:14,810
who is credited really as the person behind the World Wide Web, will be at the Oxford Martin School giving a talk at five o'clock on Thursday.

396
00:47:14,810 --> 00:47:21,740
So do register for that. If you haven't already and look forward to your engagement in the future.

397
00:47:21,740 --> 00:47:29,330
Thanks for all of your participation and thanks particularly to Gina for her great presentation.

398
00:47:29,330 --> 00:47:31,631
Stay safe and good luck to you.