AI’s Diversity Problem

WEBVTT
1
00:00:00.720 --> 00:00:04.719
The Tech Policy Design Institute acknowledges and pays our respects
2
00:00:04.799 --> 00:00:09.759
to all First Nations people. We recognize and celebrate that,
3
00:00:09.880 --> 00:00:15.119
among many things, Indigenous people were Australia's first tech innovators.
4
00:00:19.399 --> 00:00:22.679
I've had this response of yes, well, those biases already
5
00:00:22.760 --> 00:00:26.679
exist in the world, so the AI system isn't making
6
00:00:26.719 --> 00:00:29.440
anything worse, but it will make some things better for
7
00:00:29.480 --> 00:00:32.799
other people. So that's an acceptable.
8
00:00:32.600 --> 00:00:33.320
Risk to take.
9
00:00:34.000 --> 00:00:36.719
It's often about just looking around you to see who's
10
00:00:36.759 --> 00:00:38.280
in the room and who's on the table.
11
00:00:38.520 --> 00:00:38.719
Right.
12
00:00:38.759 --> 00:00:41.520
It's often as simple as that, and it's usually the
13
00:00:41.560 --> 00:00:43.840
people who are not in the rooms and not on
14
00:00:43.920 --> 00:00:47.439
those tables, not having those conversations, are the people that
15
00:00:47.479 --> 00:00:49.439
are at the most risk of being left out.
16
00:00:49.640 --> 00:00:53.119
This is a great opportunity when we're having a technological
17
00:00:53.439 --> 00:00:57.759
revolution to embed more equality into the systems that are
18
00:00:57.759 --> 00:00:59.880
going to become so fundamental to our lives.
19
00:01:00.200 --> 00:01:03.600
And I think it's also inspiring women to recognize that
20
00:01:03.640 --> 00:01:06.280
this is actually an opportunity to smash the patriarchy if
21
00:01:06.319 --> 00:01:07.040
we get this right.
22
00:01:07.319 --> 00:01:20.120
I really believe that right.
23
00:01:25.159 --> 00:01:28.640
Hello and welcome to another episode of tech Mira. I'm
24
00:01:28.719 --> 00:01:31.719
Johanna Weaver, your host, and this is the podcast where
25
00:01:31.719 --> 00:01:36.159
we talk about how technology is shaping our world and
26
00:01:36.200 --> 00:01:39.599
how we the humans, or in the case of this podcast,
27
00:01:39.879 --> 00:01:45.359
women can help shape technology back. And particularly today, we're
28
00:01:45.359 --> 00:01:49.079
going to focus in on a field of artificial intelligence
29
00:01:49.120 --> 00:01:53.120
called AI safety. And actually, if you ask people what
30
00:01:53.159 --> 00:01:57.239
AI safety means, you'll probably get as many different answers
31
00:01:57.040 --> 00:01:59.200
as many people as you ask.
32
00:02:00.120 --> 00:02:02.519
For some people, AI safety.
33
00:02:02.319 --> 00:02:05.719
Is looking at things like existential risks, so you know,
34
00:02:05.799 --> 00:02:10.719
AI becomes so powerful that we as humans are unable to.
35
00:02:10.719 --> 00:02:11.879
Control it anymore.
36
00:02:12.560 --> 00:02:16.199
For other people, when they talk about AI safety, they're
37
00:02:16.240 --> 00:02:21.120
thinking about more immediate harms, things like bias and discrimination
38
00:02:21.400 --> 00:02:24.840
in the outputs of artificial intelligence. And what we're going
39
00:02:24.879 --> 00:02:28.560
to focus on today is what it means to be
40
00:02:28.840 --> 00:02:34.719
answering this question of what harms and importantly whose harms
41
00:02:35.159 --> 00:02:38.719
are important. Where they're having more diverse voices at the
42
00:02:38.759 --> 00:02:43.919
table when we're building, deploying, and using artificial intelligence, whether
43
00:02:44.000 --> 00:02:47.800
that would fundamentally change the outcomes that we get from
44
00:02:47.919 --> 00:02:52.080
artificial intelligence. And we have two incredible guests with us
45
00:02:52.120 --> 00:02:55.360
today to explore this topic. My first guest is doctor
46
00:02:55.400 --> 00:02:59.680
Jemma Killen, and Gemma is the executive director of the
47
00:02:59.719 --> 00:03:04.400
Working with Women Alliance. Her work self evidently focuses on
48
00:03:04.479 --> 00:03:08.080
gender equality, but she's done some recent research that is
49
00:03:08.159 --> 00:03:12.759
particularly looking at how AI can affect women's economic participation,
50
00:03:13.479 --> 00:03:17.719
health outcomes, and how do we get an equitable distribution
51
00:03:18.000 --> 00:03:19.680
of the benefits of AI.
52
00:03:20.400 --> 00:03:23.080
So, Gemma, welcome to the pod, thanks for having me.
53
00:03:23.479 --> 00:03:27.159
And our second guest today is Professor Rashena. Hotter. She's
54
00:03:27.199 --> 00:03:31.039
a software engineering researcher and the director of the Humanized
55
00:03:31.120 --> 00:03:36.319
Lab at Monash University. Her work focuses on the human
56
00:03:36.560 --> 00:03:41.240
AI collaboration, with a particular focus on AI bias and
57
00:03:41.280 --> 00:03:45.520
the sociotechnical factors that shape the development of safe and
58
00:03:45.560 --> 00:03:48.879
trustworthy AI. Rashina, it's wonderful to have you on the
59
00:03:48.879 --> 00:03:49.439
pod today.
60
00:03:49.759 --> 00:03:50.599
Thank you for having me.
61
00:03:50.680 --> 00:03:55.199
Johannah, So, when we say the words AI safety, it
62
00:03:55.240 --> 00:04:00.240
really does mean different things to different people. For the
63
00:04:00.280 --> 00:04:04.800
two of you, when you hear AI safety, what's the
64
00:04:04.840 --> 00:04:07.759
first risk that you think of? And maybe it's a
65
00:04:07.840 --> 00:04:12.759
risk that is underrecognized and perhaps something that others aren't
66
00:04:12.800 --> 00:04:14.800
as focused on. So Jimmy, why don't we go to
67
00:04:14.840 --> 00:04:17.399
you first and then Raschida will come to you next.
68
00:04:18.000 --> 00:04:19.199
Thanks for the question.
69
00:04:19.360 --> 00:04:19.720
Joanna.
70
00:04:19.959 --> 00:04:22.160
I think that one of the things that we tend
71
00:04:22.160 --> 00:04:26.519
to overlook when we're talking about AI safety is the
72
00:04:26.560 --> 00:04:31.160
harms that women face through the use of AI. And
73
00:04:31.279 --> 00:04:34.439
there's a big spectrum of those harms. There's the sort
74
00:04:34.480 --> 00:04:37.800
of bias that's embedded in the systems that we use
75
00:04:37.879 --> 00:04:41.639
now to assess bank loans, to assess rental applications, to
76
00:04:41.720 --> 00:04:45.839
assess job applications, that bake in the sort of discrimination
77
00:04:45.920 --> 00:04:49.319
that we've seen for decades against women. But there's also
78
00:04:49.959 --> 00:04:54.680
the nefarious use of AI that happens and targets women.
79
00:04:54.800 --> 00:04:58.759
So in one set, in one example, there's I could
80
00:04:58.759 --> 00:05:01.920
buy a chatbot for my website, program it to spit
81
00:05:02.000 --> 00:05:07.519
out disinformation about how harmful contraception is or how abortions
82
00:05:07.519 --> 00:05:11.399
are morally wrong, and no one would hold me accountable
83
00:05:11.720 --> 00:05:14.600
for that. I would be allowed to just let people
84
00:05:14.759 --> 00:05:17.439
assume that that information is correct through the use of
85
00:05:17.480 --> 00:05:21.399
the chatbot. But also we know that there's increasing use
86
00:05:21.600 --> 00:05:26.120
of AI tools by perpetrators of domestic and family violence
87
00:05:26.279 --> 00:05:30.439
in the family court, for example, to put forward vexatious claims,
88
00:05:30.600 --> 00:05:34.240
delay family core proceedings, keep people wrapped in a cycle
89
00:05:34.279 --> 00:05:37.240
of domestic and family violence. So those are the sorts
90
00:05:37.279 --> 00:05:39.240
of harms that I think about when I think about
91
00:05:39.279 --> 00:05:40.439
AI safety.
92
00:05:40.079 --> 00:05:43.079
And Rashida, does that echo with you and do you
93
00:05:43.120 --> 00:05:45.120
think about additional harms? No?
94
00:05:45.240 --> 00:05:48.920
Absolutely, I think what Jemma said really captures well some
95
00:05:49.000 --> 00:05:51.920
of the issues we're facing as women in this age
96
00:05:51.920 --> 00:05:56.120
of AI. I'd also like to add from the perspective
97
00:05:56.279 --> 00:06:00.639
of developers and designers of AI, that we know from
98
00:06:00.720 --> 00:06:06.600
recent reports that anywhere from twenty to thirty percent of developers, designers,
99
00:06:06.680 --> 00:06:10.000
and researchers in AI are women. So that's a very
100
00:06:10.079 --> 00:06:13.639
short small percentage of the total popilation. And what that
101
00:06:13.800 --> 00:06:17.759
means is the risk there is when we develop AI
102
00:06:18.439 --> 00:06:21.519
that we might get these blind spots into our products
103
00:06:21.519 --> 00:06:27.240
and services, the examples that Jammas shared for instance, and
104
00:06:27.279 --> 00:06:30.360
that can happen because there is not enough diversity in
105
00:06:30.439 --> 00:06:33.759
the team that is designing this AI systems to begin with.
106
00:06:34.040 --> 00:06:36.800
And then there's the upstream effect in terms of the
107
00:06:36.839 --> 00:06:41.000
foundational and frontier models, the AI models, the large language models,
108
00:06:41.240 --> 00:06:44.000
they're based on, these massive amounts of data that is
109
00:06:44.040 --> 00:06:48.360
already biased against women and other marginalized groups to begin with.
110
00:06:48.639 --> 00:06:49.839
So you have this whole.
111
00:06:49.639 --> 00:06:54.040
Cascading effect right from the foundational models into the products
112
00:06:54.079 --> 00:06:57.000
and services because there's not enough diversity in the teams
113
00:06:57.040 --> 00:07:00.120
and then onto the users. So it's a chain and
114
00:07:00.120 --> 00:07:03.120
it's a cascading effect, as I said. And the other
115
00:07:03.199 --> 00:07:06.199
aspect of it, in terms of the risk, which is
116
00:07:06.360 --> 00:07:10.199
now being pointed out by various people, is this idea
117
00:07:10.279 --> 00:07:15.680
of AI competency penalty, which is to say that when
118
00:07:15.720 --> 00:07:19.639
women use AI or try to upskill themselves using AI
119
00:07:19.759 --> 00:07:24.639
at work, they're just seen as proof of their incompetence. Oh,
120
00:07:24.720 --> 00:07:27.040
she's using AI because of course she couldn't have done
121
00:07:27.040 --> 00:07:30.160
it without AI, Whereas if the same thing is being
122
00:07:30.199 --> 00:07:34.879
done by men, it's seen as pragmatic use of new technology.
123
00:07:34.920 --> 00:07:37.800
They're on with it, they're really you know, moving with
124
00:07:37.920 --> 00:07:41.439
the times. So it's a very ironic situation. And it
125
00:07:41.560 --> 00:07:44.920
is the as I said, the AI competency penalty that
126
00:07:45.040 --> 00:07:48.600
women face exclusively. So I mean the way I put
127
00:07:48.639 --> 00:07:51.639
it is like damned if she does, damned if she doesn't.
128
00:07:51.759 --> 00:07:54.360
And I want to draw on a particular element that
129
00:07:54.399 --> 00:07:57.399
you're talking about the Reshena in terms of the people
130
00:07:57.439 --> 00:08:01.800
who are designing these technologies and in particular looking at
131
00:08:02.040 --> 00:08:04.560
I love the way you're talking there about the foundational models,
132
00:08:04.560 --> 00:08:07.519
and for those of our listeners who maybe don't live
133
00:08:07.560 --> 00:08:11.399
and breathe AI, the foundational models are called foundational or
134
00:08:11.439 --> 00:08:15.480
frontier models because they actually are then used and built
135
00:08:15.480 --> 00:08:19.240
into so many other types of artificial intelligence. So people
136
00:08:19.319 --> 00:08:21.759
take these foundation models and then they tweak them and
137
00:08:21.839 --> 00:08:25.519
modify them. And so if you have bias in those
138
00:08:25.680 --> 00:08:28.399
foundational models, which frankly there is because a lot of
139
00:08:28.480 --> 00:08:31.160
these models have been trained by learning off the internet,
140
00:08:31.160 --> 00:08:33.679
and we all know how glorious many parts of the
141
00:08:33.679 --> 00:08:37.120
internet are, then you are starting to see that reflected
142
00:08:37.399 --> 00:08:41.600
in the outcomes that these models are producing. We've spent
143
00:08:42.120 --> 00:08:45.240
a lot of time talking about bias, or I have
144
00:08:45.399 --> 00:08:48.960
over the last few years, and listening to you, it
145
00:08:49.000 --> 00:08:51.639
does make me wonder if we're actually asking the wrong question.
146
00:08:51.919 --> 00:08:55.960
Rather than focusing on the output, which is bias, whether
147
00:08:56.080 --> 00:08:59.080
or not we should be looking more directly at the input,
148
00:08:59.159 --> 00:09:02.919
which is how to actually get more diversity in terms
149
00:09:02.960 --> 00:09:05.480
of the people that are making this technology. And that's
150
00:09:05.679 --> 00:09:09.120
more diversity of women, but also more diversity full stop,
151
00:09:09.279 --> 00:09:12.320
of all types. So do you think that would make
152
00:09:12.600 --> 00:09:13.799
a big difference?
153
00:09:13.960 --> 00:09:15.279
And how do we do that?
154
00:09:15.360 --> 00:09:17.279
What are the practical steps that we would need to
155
00:09:17.320 --> 00:09:19.720
take to get more diversity in the people that are
156
00:09:20.120 --> 00:09:22.519
that are building and making these technologies.
157
00:09:22.799 --> 00:09:26.120
So that's a great question, and I think the diversity
158
00:09:26.120 --> 00:09:29.600
problem is very real. So one of the pieces of
159
00:09:29.639 --> 00:09:32.879
research we did last year with our collaborators at CSIRO
160
00:09:34.039 --> 00:09:37.159
was to look at these foundational models, as you said,
161
00:09:37.200 --> 00:09:41.200
the GPT and a core pilot at that time. We
162
00:09:41.240 --> 00:09:44.720
made a little scenario for them to rank candidates in
163
00:09:44.840 --> 00:09:49.480
order of preference for a particular job advertisement, right, and
164
00:09:49.559 --> 00:09:51.919
so this was all kind of, you know, an experimental setup.
165
00:09:52.200 --> 00:09:56.639
We made these profiles exactly same except for the gender, right,
166
00:09:56.720 --> 00:09:59.159
And you know, the details are available in the paper,
167
00:09:59.200 --> 00:10:01.080
but I'll cut to the chase. In terms of what
168
00:10:01.200 --> 00:10:04.840
we found, it was, you know, not surprising in some
169
00:10:04.879 --> 00:10:09.919
ways that chat GPT was hugely biased towards men and
170
00:10:10.000 --> 00:10:16.240
men from Caucasian backgrounds, global North settings and also interestingly
171
00:10:17.360 --> 00:10:22.200
athletic built. So there was there was, there was, there
172
00:10:22.320 --> 00:10:24.519
was a part of about it in the recruitment, but
173
00:10:24.559 --> 00:10:27.639
we also asked it to draw software engineers where you know,
174
00:10:27.720 --> 00:10:31.320
you there's lesser chance of actually hiding the biases. So
175
00:10:31.639 --> 00:10:35.879
in those drawings it was predominantly these Caucasian men who
176
00:10:35.879 --> 00:10:39.120
are super athletic kind of situation. There was not a
177
00:10:39.159 --> 00:10:44.360
single representation of anyone with the visible disability. And co
178
00:10:44.519 --> 00:10:48.039
Pilot was kind of funny because it was funny and
179
00:10:48.159 --> 00:10:50.519
biased in a funny way in that it would always
180
00:10:50.600 --> 00:10:53.879
just pick the first candidate in the list. It didn't
181
00:10:53.919 --> 00:10:57.639
care who the other ones, who the other ones were.
182
00:10:57.720 --> 00:10:59.639
Maybe it had a tough day and it just didn't
183
00:10:59.639 --> 00:11:01.519
want to spend the time doing it. I mean, I've
184
00:11:01.559 --> 00:11:04.039
been in recruitment processes where I've felt like doing that.
185
00:11:04.720 --> 00:11:07.360
It's hilarious because it's basically what it means is in
186
00:11:07.399 --> 00:11:10.080
a real world context. Imagine the first person who walks
187
00:11:10.080 --> 00:11:12.960
in for the interview gets the job. So, you know,
188
00:11:13.200 --> 00:11:15.639
it's funny because when we put these things out there,
189
00:11:15.720 --> 00:11:18.480
and it's this sense of oh, it's coming from a computer,
190
00:11:18.600 --> 00:11:21.360
therefore it must be objective and right, and it's not,
191
00:11:21.679 --> 00:11:24.519
and it's super biased. Coming back to your question, I
192
00:11:24.559 --> 00:11:29.879
think the fact that some areas are particularly challenging and
193
00:11:29.960 --> 00:11:33.799
not right for putting AI into, for example, recruitment, So
194
00:11:33.960 --> 00:11:36.480
and that is something that's looked at as being you know,
195
00:11:36.600 --> 00:11:38.639
a silver bullet. All right, we can fix this issue
196
00:11:38.679 --> 00:11:41.600
and just put AI into the filtering. Not a good idea,
197
00:11:41.759 --> 00:11:44.399
not with all of these biases. So that's one issue.
198
00:11:44.679 --> 00:11:46.519
You have to do a little bit of extra effort
199
00:11:46.519 --> 00:11:49.159
if you want diverse teams and the short list, to
200
00:11:49.240 --> 00:11:52.000
have a short list that is already diverse and not
201
00:11:52.159 --> 00:11:55.279
just your one token woman on the list. We know
202
00:11:55.360 --> 00:11:58.360
that doesn't work. And then right up to the stage
203
00:11:58.360 --> 00:12:00.399
where they're the shorter list for the indie views and
204
00:12:00.399 --> 00:12:04.080
so one has to keep getting representative. The panels need
205
00:12:04.120 --> 00:12:07.120
to be diverse, So the hiring panels themselves need to
206
00:12:07.120 --> 00:12:09.600
be diverse. And of course there's a whole pipeline issue
207
00:12:09.639 --> 00:12:13.080
right from how we educate our kids and how we
208
00:12:13.159 --> 00:12:17.320
talk about AI in terms of you know, or technology
209
00:12:17.360 --> 00:12:20.440
generally as being something for all genders, not just for boys.
210
00:12:20.759 --> 00:12:25.120
So Rashenea's just referred there to some examples of AI
211
00:12:25.360 --> 00:12:28.679
being used in employment. You have some really good examples
212
00:12:28.679 --> 00:12:31.600
from the work you've done looking at in areas of
213
00:12:31.720 --> 00:12:35.120
healthcare and some of the biases that are emerging with
214
00:12:35.320 --> 00:12:39.639
artificial intelligence and healthcare. Could you talk to talk us
215
00:12:39.639 --> 00:12:41.919
through some of the findings that you came out with.
216
00:12:42.399 --> 00:12:46.200
So the research that we looked at around women and
217
00:12:46.480 --> 00:12:49.879
AI and healthcare, so that women are more likely to
218
00:12:50.039 --> 00:12:53.200
use AI to fill in healthcare gaps. And we think,
219
00:12:53.279 --> 00:12:57.200
you know, we all know what happens with medical misogyny. Basically,
220
00:12:57.320 --> 00:13:02.000
women's health isn't taken seriously in the real world AI aside,
221
00:13:02.559 --> 00:13:05.440
but it means that women's health can be really expensive,
222
00:13:05.639 --> 00:13:09.240
take a really long time to access proper care, and
223
00:13:09.840 --> 00:13:13.279
to fill the gap, women are turning to general AI
224
00:13:13.399 --> 00:13:15.960
tools in order to get the care that they need,
225
00:13:16.039 --> 00:13:18.840
get the kind of diagnoses that they want. But there's
226
00:13:18.879 --> 00:13:23.960
high rates of misinformation and disinformation in out there and
227
00:13:24.080 --> 00:13:27.679
being scraped by AI tools. There was a piece of
228
00:13:27.720 --> 00:13:30.519
research that I read recently that said that people who
229
00:13:30.519 --> 00:13:33.519
are peddling peptides at the moment, for example, are filling
230
00:13:33.840 --> 00:13:38.200
Reddit and chatboards with information that is specifically written to
231
00:13:38.320 --> 00:13:43.279
appeal to AI scraping tools, so that when you ask
232
00:13:43.559 --> 00:13:47.360
a chatbot a question about peptides, the information that the
233
00:13:47.360 --> 00:13:50.240
peptide companies have put out there is scraped up and
234
00:13:50.279 --> 00:13:52.919
delivered to you as though it is a neutral summary.
235
00:13:53.480 --> 00:13:56.519
And the impacts on women are much more significant than
236
00:13:56.559 --> 00:14:02.240
they are for men, because our healthcare is already so deprioritized.
237
00:14:02.519 --> 00:14:06.840
And do you think that if we're talking about diversity,
238
00:14:07.120 --> 00:14:11.120
is it about improving the fairness in those outcomes or
239
00:14:11.200 --> 00:14:14.840
is it more fundamentally different. Does it change what we're
240
00:14:14.840 --> 00:14:16.759
looking at when we're thinking about risks.
241
00:14:16.840 --> 00:14:19.039
So Jemma Watter not come to you on that one.
242
00:14:19.200 --> 00:14:22.360
Yeah, I think which things you consider to be risks
243
00:14:22.480 --> 00:14:26.360
and where you look for problems depend on your life experience.
244
00:14:26.559 --> 00:14:30.240
And if we have only a certain group of white
245
00:14:30.279 --> 00:14:35.360
men developing these tools, then of course they won't think
246
00:14:35.440 --> 00:14:41.200
to check whether their tools encourage misinformation spread or are
247
00:14:41.240 --> 00:14:45.799
biased against women in hiring processes. And in the same
248
00:14:45.840 --> 00:14:48.519
way that when men are on hiring panels, they might
249
00:14:48.559 --> 00:14:52.279
think it's perfectly reasonable to put the cvs that have
250
00:14:52.799 --> 00:14:54.720
big time gaps in them to the bottom of the
251
00:14:54.759 --> 00:14:58.320
pile without considering the impacts of the care load on women.
252
00:14:58.759 --> 00:15:02.559
When they're developing systems, they might approach it in the
253
00:15:02.600 --> 00:15:04.799
same way. So I think it's really important to embed
254
00:15:04.919 --> 00:15:08.480
diversity in the development of tools from the very beginning.
255
00:15:09.159 --> 00:15:11.879
And Rishena, when we were on the panel, you had
256
00:15:11.919 --> 00:15:15.399
a great analogy about a toaster, which I think actually
257
00:15:15.440 --> 00:15:18.799
helps to emphasize this point that Jemma is making.
258
00:15:18.840 --> 00:15:19.840
Can you talk us through that.
259
00:15:20.360 --> 00:15:24.440
Sure the toaster analogy. So this was where now I
260
00:15:24.480 --> 00:15:27.679
was thinking about your questions, Johanna, about what comes to
261
00:15:27.759 --> 00:15:31.159
mind when you think of AI safety, right, and this
262
00:15:31.240 --> 00:15:33.559
is something most of us do in the morning, is
263
00:15:33.840 --> 00:15:36.519
toast a piece of bread or so, and I was
264
00:15:36.559 --> 00:15:41.799
just thinking about the number of regulations and standards a
265
00:15:42.240 --> 00:15:46.120
humble toaster has to actually go through and pass before
266
00:15:46.120 --> 00:15:48.919
it is allowed to be inside of our homes is
267
00:15:49.120 --> 00:15:50.919
quite a few, right, So it has to be checked
268
00:15:50.919 --> 00:15:55.720
for electrical safety, for thermal safety. In fact, apparently there's
269
00:15:55.720 --> 00:15:57.879
also something to do with a tipping point where it's
270
00:15:57.960 --> 00:16:01.399
not accidentally tipped over by ped and by kids. So
271
00:16:01.440 --> 00:16:04.200
there's an angle at which and then there's the auto
272
00:16:04.360 --> 00:16:07.399
electric shut off in case there's something going wrong, and
273
00:16:07.480 --> 00:16:10.360
of course then the things go on from there around
274
00:16:10.399 --> 00:16:13.519
fire hazards and whatnot. So this is a simple home
275
00:16:13.559 --> 00:16:17.399
appliance and an amount of risk assessment around a simple
276
00:16:17.440 --> 00:16:21.240
home appliance and safety standards that it has to actually pass.
277
00:16:21.559 --> 00:16:25.000
And then the fact that the users know for a
278
00:16:25.039 --> 00:16:27.720
fact that this has been quality checked and it has
279
00:16:27.799 --> 00:16:31.320
actually passed those standards before they are comfortable using it. Now,
280
00:16:31.360 --> 00:16:35.159
compare that to AI or using you know, Chatgypt or Gemini,
281
00:16:35.200 --> 00:16:38.440
what have you, all of these. Where are the standards,
282
00:16:39.159 --> 00:16:42.759
Where is the regulation, Where is the there's a little
283
00:16:42.759 --> 00:16:45.279
disclaimer that says, of course we can get it wrong,
284
00:16:45.440 --> 00:16:49.399
and don't get me started about hallucination. But seriously, I mean,
285
00:16:49.559 --> 00:16:53.159
compared to a humble toaster, there isn't enough really risk
286
00:16:53.240 --> 00:16:57.639
assessment for something as pervasive in our lives and as
287
00:16:57.759 --> 00:17:01.240
Ai and our kids are using it, our adults are
288
00:17:01.320 --> 00:17:05.480
using it, everybody is using it without any concern for
289
00:17:05.839 --> 00:17:08.400
all of the risks that it carries. So I just
290
00:17:08.440 --> 00:17:11.440
find it mind boggling. So that's my toaster analogy.
291
00:17:11.680 --> 00:17:13.799
And I think you also made the point when we
292
00:17:13.839 --> 00:17:16.160
spoke last about you know, and if the toaster was
293
00:17:16.240 --> 00:17:18.640
known to shock women three times more than men.
294
00:17:18.839 --> 00:17:21.799
Thank you for reminding me, Thank you for reminding me.
295
00:17:21.920 --> 00:17:25.079
So that's the other part of it, which is even
296
00:17:25.119 --> 00:17:28.519
more shocking, which is we cannot even think of a
297
00:17:28.559 --> 00:17:33.599
toaster that is actually known to shock say, young women
298
00:17:34.240 --> 00:17:38.759
ten times as more as men, or to shock disabled people,
299
00:17:39.400 --> 00:17:43.079
or not toast decide not to toast for people with
300
00:17:43.160 --> 00:17:48.000
disabilities every other Monday, like, we cannot even imagine such
301
00:17:48.039 --> 00:17:50.839
a product. And yet we see all these biases in
302
00:17:50.960 --> 00:17:54.920
AI where the same prompt, depending on you know, your
303
00:17:55.000 --> 00:17:57.880
profile or background, or whether it is a prompt about
304
00:17:58.160 --> 00:18:00.799
a different gender or a different city and so on,
305
00:18:01.200 --> 00:18:05.880
is going to give you bizarre and almost unexpected results. Right,
306
00:18:06.200 --> 00:18:08.720
So there's no regulation in that sense as well. And
307
00:18:08.799 --> 00:18:11.079
so it's a matter of safety and it's a matter
308
00:18:11.119 --> 00:18:14.079
of inclusion. And these are standard things that we expect
309
00:18:14.119 --> 00:18:17.720
from any product or service, even a whole humble toaster.
310
00:18:17.880 --> 00:18:20.319
So why not from something as powerful as AI?
311
00:18:20.880 --> 00:18:24.319
And I would argue actually that Australian law already requires.
312
00:18:24.319 --> 00:18:26.960
So you know, we have anti discrimination laws that prohibit
313
00:18:27.039 --> 00:18:29.839
this type of thing, right. The challenge is that they're
314
00:18:29.880 --> 00:18:34.640
not currently being applied or enforced against the AI companies,
315
00:18:34.680 --> 00:18:37.559
and so how do we do that? And that requires,
316
00:18:37.720 --> 00:18:40.440
you know, a big shift in terms of the regulator's capacity,
317
00:18:40.440 --> 00:18:43.759
but it also requires much more transparency from the companies
318
00:18:43.960 --> 00:18:44.599
as well.
319
00:18:45.240 --> 00:18:46.920
Now, Regina, in.
320
00:18:46.839 --> 00:18:49.759
Your previous response, you were talking about employment and how
321
00:18:49.839 --> 00:18:51.960
do we get more people to the table. I really
322
00:18:52.000 --> 00:18:53.880
do want to come back to that point, but before
323
00:18:53.880 --> 00:18:56.960
we do, I want to really double down again on
324
00:18:57.000 --> 00:19:00.759
this question of whose harms matter. So you know you've
325
00:19:00.799 --> 00:19:02.599
both given us some really good examples.
326
00:19:02.640 --> 00:19:03.279
They're both in.
327
00:19:03.319 --> 00:19:07.880
Healthcare and in employment. Why do you think it is
328
00:19:08.359 --> 00:19:12.720
that we allow or that society has allowed AI to
329
00:19:12.920 --> 00:19:15.680
roll out when we know that there are these biases,
330
00:19:16.039 --> 00:19:19.279
and when we know that it is having disproportionate harm
331
00:19:19.680 --> 00:19:21.960
on particular parts of the community.
332
00:19:22.519 --> 00:19:23.440
Why do you think it.
333
00:19:23.440 --> 00:19:28.240
Is that these groups are consistently having this type of
334
00:19:28.519 --> 00:19:32.640
disadvantage entrenched almost through artificial intelligence?
335
00:19:33.319 --> 00:19:35.480
And are there other groups.
336
00:19:35.079 --> 00:19:37.960
That maybe we haven't yet mentioned, that we think we
337
00:19:38.000 --> 00:19:40.920
should also be paying more attention to.
338
00:19:41.039 --> 00:19:41.960
So here I'm really.
339
00:19:41.759 --> 00:19:43.839
Looking at whose harms matter.
340
00:19:44.440 --> 00:19:45.839
Gemma, let's go to you first.
341
00:19:46.039 --> 00:19:47.720
I think one of the things that we found when
342
00:19:47.720 --> 00:19:51.319
we were looking at, for example, bias in rental application
343
00:19:51.839 --> 00:20:00.240
assessments was that aisystems routinely deprioritize applications from people have
344
00:20:00.279 --> 00:20:03.960
income support, for example. And we read a few tests
345
00:20:04.000 --> 00:20:07.319
where people had submitted an application with their income support
346
00:20:07.440 --> 00:20:10.279
and without their income support, and even though the one
347
00:20:10.400 --> 00:20:14.279
without the income support listed had a lower income, it
348
00:20:14.440 --> 00:20:15.599
was valued more.
349
00:20:15.960 --> 00:20:16.039
So.
350
00:20:16.119 --> 00:20:18.799
I think there is also that element of the way
351
00:20:18.880 --> 00:20:23.640
that people with less money, lower socioeconomic status are also
352
00:20:23.720 --> 00:20:27.119
discriminated against through AI systems and that's not something that
353
00:20:27.160 --> 00:20:30.400
we routinely talk about. And to your other point about
354
00:20:30.519 --> 00:20:34.319
why we're allowing this to happen, I think it's because
355
00:20:34.400 --> 00:20:37.960
there are certain discriminations or certain biases that we still
356
00:20:38.359 --> 00:20:42.279
consider socially acceptable. A lot of the conversations that I've
357
00:20:42.319 --> 00:20:46.480
had with policy makers, for example, about the biases in
358
00:20:46.640 --> 00:20:50.920
AI systems, I've had this response of, yes, well, those
359
00:20:50.960 --> 00:20:55.279
biases already exist in the world, So the AI system
360
00:20:55.519 --> 00:20:58.359
isn't making anything worse, but it will make some things
361
00:20:58.359 --> 00:21:02.680
better for other people. So that's an acceptable risk to take.
362
00:21:03.359 --> 00:21:06.319
So I think there is a sense that's just the
363
00:21:06.319 --> 00:21:09.839
way of things as it stands, So why would we
364
00:21:10.079 --> 00:21:12.920
address it now, even though this is a great opportunity
365
00:21:12.920 --> 00:21:18.720
when we're having a technological revolution to embed more equality
366
00:21:18.759 --> 00:21:21.279
into the systems that are going to become so fundamental
367
00:21:21.319 --> 00:21:21.960
to our lives.
368
00:21:22.240 --> 00:21:24.400
Yeah, and I think what you're articulating their gena is
369
00:21:24.440 --> 00:21:26.759
exactly what we get to in this podcast. Right, these
370
00:21:26.759 --> 00:21:29.799
systems are being made by people, and perhaps those people
371
00:21:30.359 --> 00:21:33.319
need to be more diverse, but it's also we get
372
00:21:33.359 --> 00:21:37.359
to shape these systems, and how do we have more
373
00:21:37.359 --> 00:21:39.680
of an active role in shaping what these systems are
374
00:21:39.720 --> 00:21:42.799
so that we're reducing the inequalities because there is the
375
00:21:42.839 --> 00:21:46.319
real potential AI to be used for good, but it
376
00:21:46.359 --> 00:21:51.200
won't just happen if we don't actually pragmatically intervene and
377
00:21:51.480 --> 00:21:52.119
take action.
378
00:21:52.880 --> 00:21:55.240
Reshina, would there be Are there any other.
379
00:21:55.160 --> 00:21:58.240
Groups that you are particularly focused on in terms of
380
00:21:58.359 --> 00:22:02.000
people's whose arms perhaps are not getting enough attention.
381
00:22:02.319 --> 00:22:05.480
Yeah, So this is something I often say, which is
382
00:22:05.680 --> 00:22:08.640
when we are in these conversations, say at the AI
383
00:22:08.720 --> 00:22:11.519
safety forum, when we were there, or when we are
384
00:22:11.559 --> 00:22:14.440
you know, having these conversations at our workplace, it's often
385
00:22:14.480 --> 00:22:17.519
about just looking around. You see who's in the room
386
00:22:17.599 --> 00:22:20.160
and who's on the table. Right, It's often as simple
387
00:22:20.160 --> 00:22:22.599
as that, and it's usually the people who are not
388
00:22:22.680 --> 00:22:26.000
in the rooms and not on those tables, not having
389
00:22:26.000 --> 00:22:28.559
those conversations are the people that are at the most
390
00:22:28.640 --> 00:22:32.000
risk of being left out. So for example, have people
391
00:22:32.160 --> 00:22:36.680
with visible or invisible disabilities you have oftentimes it's actually
392
00:22:36.759 --> 00:22:39.759
quite sadly our first nations people that are you know,
393
00:22:39.880 --> 00:22:42.799
left out of these conversations. You know, we'll have people
394
00:22:42.880 --> 00:22:46.319
who are from as Jemma said, low socioeconomic background. I'm
395
00:22:46.319 --> 00:22:50.079
also going to add people from particular age groups. So kids,
396
00:22:50.480 --> 00:22:56.079
young adults, older people oftentimes not considered as actively as
397
00:22:56.559 --> 00:22:58.920
you know we should. So it's almost like we are
398
00:22:58.960 --> 00:23:05.039
designing for the default average human, which does not exist, right,
399
00:23:05.440 --> 00:23:09.640
There is no average human and all of these variations,
400
00:23:09.640 --> 00:23:12.759
if you like, from that average default is the reality
401
00:23:12.799 --> 00:23:14.839
of who we are, and we have to consider that.
402
00:23:15.359 --> 00:23:18.839
So it's a range of people that get left behind
403
00:23:18.960 --> 00:23:21.759
is just quite shocking and quite sad. But I was
404
00:23:21.799 --> 00:23:24.119
also going to quickly add on a little bit about
405
00:23:24.160 --> 00:23:27.160
the people who are designing, as you hinted at, Johanna,
406
00:23:27.240 --> 00:23:30.279
and in particular, there's this acronym doing the round these
407
00:23:30.359 --> 00:23:36.359
days around mangoes without the E. So it's meta anthropic
408
00:23:36.839 --> 00:23:42.319
in video Google and open AI and SpaceX right, and
409
00:23:42.400 --> 00:23:46.000
if you picture just for a moment, they're CEOs is
410
00:23:46.039 --> 00:23:51.440
an extremely homogeneous group of leaders and they're literally deciding
411
00:23:51.559 --> 00:23:55.720
the fate of the world because this technology is just
412
00:23:55.799 --> 00:23:58.400
about in every product and service and it's going to,
413
00:23:58.559 --> 00:24:03.160
you know, keep getting more and more pervasive. And that's
414
00:24:03.240 --> 00:24:06.880
kind of the pinnacle of this AI gender gap, as
415
00:24:06.920 --> 00:24:09.839
I kind of refer to it, right. So from there
416
00:24:09.920 --> 00:24:12.799
then it cascades do the design groups and the governance
417
00:24:12.960 --> 00:24:16.759
teams and the developers and then into the how people
418
00:24:16.960 --> 00:24:19.440
use it. So it starts right at the top and
419
00:24:19.480 --> 00:24:20.720
it's cascades from there.
420
00:24:21.000 --> 00:24:23.319
Yeah, and I guess I mean they're all right at
421
00:24:23.319 --> 00:24:25.839
the top, but also you know, to flip it as well,
422
00:24:26.000 --> 00:24:28.680
right at the foundation, right, because those tools that you're
423
00:24:28.720 --> 00:24:30.880
talking about are all the tools that are being used
424
00:24:30.880 --> 00:24:33.440
to build all of the other systems on top of them.
425
00:24:34.160 --> 00:24:36.559
And I couldn't agree with you more about about the
426
00:24:37.519 --> 00:24:40.799
impact that it's having on different communities. I mean, it's
427
00:24:40.799 --> 00:24:44.160
a very small and not hugely significant, but you know,
428
00:24:44.160 --> 00:24:47.720
I'm chronically dyslexic, and I swear the spell check on
429
00:24:48.079 --> 00:24:50.200
a lot of products is getting worse and worse and
430
00:24:50.240 --> 00:24:53.000
worse because they're investing more and more energy into making
431
00:24:53.039 --> 00:24:56.400
their AI models be really efficient. And it drives me
432
00:24:56.440 --> 00:24:59.599
bonkers because I know that these models can do this
433
00:25:00.079 --> 00:25:02.119
and it's really quite noticeable.
434
00:25:02.359 --> 00:25:03.920
Or what type of.
435
00:25:03.880 --> 00:25:07.680
Harms perhaps are not being considered and whose perspectives may
436
00:25:07.680 --> 00:25:11.920
be are not present in the discussions. What I'd like
437
00:25:12.000 --> 00:25:13.720
to focus on now is what do.
438
00:25:13.680 --> 00:25:14.480
We do about that?
439
00:25:14.599 --> 00:25:18.720
What's the practical things that we can all focus on
440
00:25:18.799 --> 00:25:21.519
to change this? Right? Because this is about how do
441
00:25:21.599 --> 00:25:25.039
we help shape a better outcome because they're not powerless
442
00:25:25.079 --> 00:25:29.000
in this scenario. Rashina, you spoke a little bit earlier
443
00:25:29.319 --> 00:25:33.759
about from an employment perspective and getting having diversity in
444
00:25:33.759 --> 00:25:36.200
the panels, having consideration in the way that you're doing
445
00:25:36.240 --> 00:25:38.240
the advertisement, these types of things.
446
00:25:38.920 --> 00:25:42.720
Are there other practical steps that.
447
00:25:43.519 --> 00:25:47.400
Either people working in these companies can take or also
448
00:25:47.599 --> 00:25:51.240
listeners who maybe don't work for tech companies but are
449
00:25:51.359 --> 00:25:54.119
concerned about these impacts?
450
00:25:54.200 --> 00:25:55.599
What can we do about.
451
00:25:55.279 --> 00:26:00.799
This as regular users? I think there the answer is
452
00:26:00.839 --> 00:26:04.000
almost at our fingertips. This is not foolproof, but it's
453
00:26:04.000 --> 00:26:06.240
a place to start, which is you can actually have
454
00:26:06.279 --> 00:26:10.319
conversations with AI about what it's not good at, So
455
00:26:10.400 --> 00:26:12.880
you can start there. And so for example, if you
456
00:26:12.920 --> 00:26:16.759
were looking at using AI for a specific task like
457
00:26:16.960 --> 00:26:19.880
helping you with your home renovation or whatever it is,
458
00:26:20.279 --> 00:26:22.519
that you can also get it to critique and you
459
00:26:22.559 --> 00:26:26.079
can get other AIS to critique the outputs of another AI,
460
00:26:26.480 --> 00:26:29.200
and that way where you can get them to look
461
00:26:29.240 --> 00:26:31.200
at or what are the bits that I'm missing? Now,
462
00:26:31.240 --> 00:26:33.079
as I said, it's not fool proof, it's a place
463
00:26:33.160 --> 00:26:37.599
to start, but that's on an everyday basis. But I
464
00:26:37.640 --> 00:26:41.759
actually have more hope from our next generation. So that's
465
00:26:41.839 --> 00:26:45.240
where I'm So as a researcher, I get to research
466
00:26:45.400 --> 00:26:49.599
the unfortunate biases and so on. But then as an educator,
467
00:26:50.160 --> 00:26:54.200
I have this unique opportunity to educate the next generation.
468
00:26:54.720 --> 00:26:56.920
And even though I'm not in schools, I'm in tertiary
469
00:26:57.039 --> 00:27:00.519
education in schools. I think that is a huge obity
470
00:27:01.079 --> 00:27:05.759
to help kids understand and young adults understand this is
471
00:27:05.880 --> 00:27:10.119
not rocket science, okay. If anything, one of the strongest
472
00:27:10.839 --> 00:27:15.160
positives out of AI is to democratize technology and you
473
00:27:15.200 --> 00:27:18.440
can learn. You can learn using AI again with all
474
00:27:18.480 --> 00:27:21.839
of the disclaimers around you know, bias and hallucination, but
475
00:27:21.960 --> 00:27:26.000
you can actually learn and start to look inside the
476
00:27:26.039 --> 00:27:29.759
box of this, you know, technology development, and it's not
477
00:27:30.119 --> 00:27:34.000
for some complex rocket science NASA people over there, it's
478
00:27:34.079 --> 00:27:37.480
all of us. So the next generation of kids, I
479
00:27:37.519 --> 00:27:41.960
feel like, specially kids, especially girls, and especially kids from
480
00:27:42.279 --> 00:27:46.440
marginalized communities, they need to be empowered to own this
481
00:27:46.759 --> 00:27:50.359
and look at themselves as possible leaders in the space
482
00:27:50.759 --> 00:27:53.440
coming up, and not just as you know, as the
483
00:27:53.480 --> 00:27:56.839
people who are on the margins of these statistics that
484
00:27:56.920 --> 00:27:59.839
are always consistently being left behind. So the more, they're
485
00:28:00.079 --> 00:28:03.599
power to understand the technology and really understand what's behind
486
00:28:03.599 --> 00:28:06.839
this black box that they can take charge of that
487
00:28:06.920 --> 00:28:09.160
in the future and actually shape it in the right direction.
488
00:28:10.200 --> 00:28:11.279
Well what about for you, Jimma.
489
00:28:11.920 --> 00:28:14.559
I think one of the things that always freaks me
490
00:28:14.599 --> 00:28:17.799
out is that we have data that says in Australia,
491
00:28:18.519 --> 00:28:20.759
men are more likely to have their training and their
492
00:28:20.839 --> 00:28:23.799
education paid for by either their employer or their government,
493
00:28:24.319 --> 00:28:27.039
and for women that's not the case, and that creates
494
00:28:27.039 --> 00:28:29.680
a huge gap across the range of industries. But we
495
00:28:29.799 --> 00:28:32.880
know also the government has said that it's women's jobs
496
00:28:32.880 --> 00:28:37.079
that will be most displaced by AI, So there's a
497
00:28:37.079 --> 00:28:40.720
real opportunity here to invest in training women in how
498
00:28:40.720 --> 00:28:44.640
to use and understand AI systems, particularly so that they're
499
00:28:44.680 --> 00:28:48.839
not disadvantaged as the next wave of AI comes through
500
00:28:49.400 --> 00:28:52.440
and changes the jobs that we have in the near future.
501
00:28:52.720 --> 00:28:55.079
So I think that's one thing is make sure we
502
00:28:55.279 --> 00:28:58.559
help to pay and build the structures for women to learn.
503
00:28:59.519 --> 00:29:03.200
And then I think just always remembering to be critical
504
00:29:03.240 --> 00:29:05.559
about the information that you engage with and try to
505
00:29:05.599 --> 00:29:10.839
think outside of yourself and your own positionality about the
506
00:29:10.880 --> 00:29:15.400
information that you're being fed and how it might reflect
507
00:29:15.400 --> 00:29:18.279
biases that are embedded in the system. I think asking
508
00:29:18.319 --> 00:29:20.720
yourself a couple of questions about what it is you're
509
00:29:20.720 --> 00:29:24.839
reading and how it might have come to be is
510
00:29:24.880 --> 00:29:27.599
really helpful in addressing some of those critical bias.
511
00:29:28.160 --> 00:29:31.759
Jemma, if you could change one thing about the world
512
00:29:31.799 --> 00:29:34.880
of AI right now, what would it be to make
513
00:29:35.000 --> 00:29:38.559
it easier for women to be engaged or for other
514
00:29:39.160 --> 00:29:41.960
disadvantage groups to engage with AI? What do you think
515
00:29:41.960 --> 00:29:44.200
the most important change that needs to happen is?
516
00:29:45.839 --> 00:29:49.359
Yeah, I think it comes back to paying women to learn.
517
00:29:50.240 --> 00:29:53.400
We know that women's confidence levels with AI are much
518
00:29:53.440 --> 00:29:58.519
lower than men's, and women mistrust or distrust AI systems
519
00:29:58.960 --> 00:30:03.599
more than men. Do something rightly so and yes, rightly so,
520
00:30:04.319 --> 00:30:08.160
But to make an intervention in that cycle is really important.
521
00:30:08.200 --> 00:30:11.119
Training women so that they can help to build those
522
00:30:11.160 --> 00:30:14.200
systems and make sure that they're not causing harm to
523
00:30:14.279 --> 00:30:17.920
other women in the future would make a massive difference. Yeah.
524
00:30:17.920 --> 00:30:21.359
And I think it's also inspiring women to recognize that
525
00:30:21.400 --> 00:30:24.039
this is actually an opportunity to smash the patriarchy if
526
00:30:24.039 --> 00:30:24.759
we get this right.
527
00:30:24.880 --> 00:30:27.920
I really believe that, right, so.
528
00:30:28.200 --> 00:30:32.119
Rishena, what about for you one practical change I think.
529
00:30:31.920 --> 00:30:36.720
We need strict laws and regulations and enforcement of those
530
00:30:36.799 --> 00:30:40.880
laws and regulations. You know, our Prime Minister talked about
531
00:30:41.200 --> 00:30:44.799
the Office of AI and the announcement around the National
532
00:30:44.839 --> 00:30:48.960
Standards for AI and made it very clear that if
533
00:30:49.000 --> 00:30:52.799
you've got you know, infringement of the copyright of Australian artists,
534
00:30:52.799 --> 00:30:55.359
that it is no less than theft. And what I
535
00:30:55.400 --> 00:30:58.480
feel is that we need the same level of clarity
536
00:30:58.799 --> 00:31:04.240
and firmness guarding harm that is being perpetuated against women
537
00:31:04.519 --> 00:31:07.759
by AI. So this is non negotiable. We have to
538
00:31:07.799 --> 00:31:10.599
have strict laws. Now again, we have the stats I
539
00:31:10.599 --> 00:31:16.680
mean the International Group of Scientists on AI recently the
540
00:31:16.759 --> 00:31:21.160
release of support which shows ninety nine percent of deep
541
00:31:21.240 --> 00:31:25.960
fakes are actually targeted against women. We cannot have this.
542
00:31:26.720 --> 00:31:29.400
We cannot have this, and this is a non negotiable.
543
00:31:29.519 --> 00:31:31.640
So this is where the laws need to come in
544
00:31:32.119 --> 00:31:35.519
to actually penalize people who are doing this and who
545
00:31:35.559 --> 00:31:38.839
are using this. So I would strongly try and put
546
00:31:38.880 --> 00:31:42.240
this across, which is in our national standards which are
547
00:31:42.279 --> 00:31:45.640
supposed to be coming up, that we include this idea
548
00:31:46.279 --> 00:31:50.160
that women are being left behind and actively look for
549
00:31:50.279 --> 00:31:53.880
legislation and laws that protect women from the harms caused
550
00:31:53.880 --> 00:31:57.400
by AI. That's extremely yeah, serious and important.
551
00:31:58.200 --> 00:32:00.920
Yeah, And I think there's so much reshina in that
552
00:32:01.119 --> 00:32:04.960
about actually enforcing the existing laws. And many of these
553
00:32:04.960 --> 00:32:08.279
things are already prohibited under Australian law, right, it's just
554
00:32:08.359 --> 00:32:11.799
that we're not enforcing them. And whether that's through courageous
555
00:32:11.799 --> 00:32:15.240
people who decide to make take test cases, whether it's
556
00:32:15.240 --> 00:32:16.279
through regulators who.
557
00:32:16.240 --> 00:32:17.039
Are enforcing it.
558
00:32:17.519 --> 00:32:20.680
But I do think with the approach that the government
559
00:32:20.720 --> 00:32:23.440
is taking with the establishment of the AI Office and
560
00:32:23.480 --> 00:32:26.799
the Prime Minister's Department, they've also made very clear that
561
00:32:26.880 --> 00:32:30.759
the existing mandates of the regulators, the existing laws apply.
562
00:32:31.440 --> 00:32:34.839
But the thing we haven't seen is increased funding for
563
00:32:34.960 --> 00:32:39.000
regulators to actually be able to apply those laws and
564
00:32:39.160 --> 00:32:43.240
enforce those laws against AI companies. And I think unless
565
00:32:43.279 --> 00:32:47.839
we get a few quick follows of enforcement action, which
566
00:32:47.880 --> 00:32:50.200
inevitably will take years to go through the courts, but
567
00:32:50.440 --> 00:32:55.559
to actually initiate those processes, this statement of existing law
568
00:32:55.599 --> 00:32:58.039
applies becomes an empty statement.
569
00:32:58.279 --> 00:32:59.839
And I say that as someone.
570
00:32:59.680 --> 00:33:02.359
Who will on the record when the government first made
571
00:33:02.400 --> 00:33:04.799
that statement a number of years ago saying this is
572
00:33:04.839 --> 00:33:07.000
the most powerful thing a government can do say the
573
00:33:07.079 --> 00:33:09.799
existing laws apply because you're not having to make new
574
00:33:09.880 --> 00:33:13.640
laws for artificial intelligence, but we have to enforce them,
575
00:33:13.720 --> 00:33:14.680
and to me that.
576
00:33:14.799 --> 00:33:16.200
Is absolutely crucial.
577
00:33:21.359 --> 00:33:26.200
We had this panel at the AI Safety Forum. It's
578
00:33:26.240 --> 00:33:28.440
been a week or a couple of weeks. It'll be
579
00:33:28.640 --> 00:33:31.480
even more so by the time this actually drops. I
580
00:33:31.519 --> 00:33:35.079
attended the Safety Forum eighteen months ago and made up
581
00:33:35.079 --> 00:33:37.920
my mission to ensure that there was greater diversity at
582
00:33:37.920 --> 00:33:40.440
this safety forum. So we had a women in AI
583
00:33:40.519 --> 00:33:43.039
mentoring network. Both of you were mentors, So thank you
584
00:33:43.079 --> 00:33:46.440
again for being mentors in that network. What are your
585
00:33:46.559 --> 00:33:50.119
observation first about the people in the room and then
586
00:33:50.279 --> 00:33:53.519
also your key takeout from that AI Safety Forum, which
587
00:33:53.559 --> 00:33:56.640
really was a key convening in terms of the researchers
588
00:33:56.680 --> 00:34:00.160
and experts who really are working on these issues in Australia.
589
00:34:00.480 --> 00:34:03.039
So I think it was such a great idea to
590
00:34:03.119 --> 00:34:06.960
have that mentor women mentoring program because and to have
591
00:34:07.039 --> 00:34:09.159
it before the forum, right, So we met in the
592
00:34:09.199 --> 00:34:12.639
evening and we had a lovely time getting to know
593
00:34:12.719 --> 00:34:15.280
each other, some of the women in that room, but
594
00:34:15.400 --> 00:34:18.760
also the mental MENTI relationship meant that some of the
595
00:34:18.800 --> 00:34:21.679
more early career people could know that they have at
596
00:34:21.760 --> 00:34:25.480
least one other person, you know, backing them. So when
597
00:34:25.519 --> 00:34:28.760
we actually ended up in the AI Safety Forum the
598
00:34:28.840 --> 00:34:32.760
next day, it was way better experience than it would
599
00:34:32.800 --> 00:34:36.239
have been, I think, because I knew these spaces right,
600
00:34:36.360 --> 00:34:39.320
and I was like, oh, hey, hello, and then you
601
00:34:39.400 --> 00:34:41.880
have the sense of you know, you start to yeah, sure,
602
00:34:41.920 --> 00:34:44.159
it's still about say twenty thirty percent. I don't have
603
00:34:44.159 --> 00:34:47.079
the numbers on how many women ended up attending.
604
00:34:47.360 --> 00:34:48.119
I'll have to ask them.
605
00:34:48.119 --> 00:34:52.719
Actually, yeah. And it's always funny because when you sometimes
606
00:34:52.800 --> 00:34:55.400
when you asked this and this happens in our classrooms,
607
00:34:55.920 --> 00:35:00.679
my male educator counterpart would say, roshina and as nice
608
00:35:00.679 --> 00:35:04.559
as so many women in the class, And I was like, wait,
609
00:35:04.599 --> 00:35:06.960
I was about to exactly say the opposite, where are
610
00:35:07.000 --> 00:35:11.360
the women in the class? So it's it depends on
611
00:35:11.440 --> 00:35:13.719
you know what you're looking out for. But at the
612
00:35:13.840 --> 00:35:16.719
II Forum, I felt, and I know from the fact
613
00:35:16.760 --> 00:35:19.679
that the year before was far less women, that we're
614
00:35:19.679 --> 00:35:23.880
on the right trajectory to actually improve that diversity. What
615
00:35:24.000 --> 00:35:26.960
I also liked in terms of diversity was people coming
616
00:35:27.079 --> 00:35:31.280
from a number of different disciplines and the number of
617
00:35:31.280 --> 00:35:35.400
different stakeholder groups were represented. So you had industry, you
618
00:35:35.480 --> 00:35:41.960
had academia, research, governance, government policy, people who were just
619
00:35:42.000 --> 00:35:44.920
you know, societal kind of not for profits.
620
00:35:44.519 --> 00:35:44.920
And so on.
621
00:35:44.960 --> 00:35:48.119
So it was such a range, and rightly so, because
622
00:35:48.159 --> 00:35:52.159
AI is touching every aspect of our lives, and so
623
00:35:52.559 --> 00:35:55.800
to hear what the others are bringing to the table
624
00:35:55.880 --> 00:36:00.559
from their disciplines was very refreshing. And then be able
625
00:36:00.639 --> 00:36:03.039
to be in the same space at the same time,
626
00:36:03.119 --> 00:36:06.599
and we could have these conversations where we're bringing our
627
00:36:06.679 --> 00:36:10.079
own piece and we can combine these ideas, so it
628
00:36:10.159 --> 00:36:13.880
becomes the greater than you know, some of the parts situation.
629
00:36:14.400 --> 00:36:18.639
So it is a really good forum to encourage the
630
00:36:19.000 --> 00:36:23.519
interdisciplinary approach that AI actually needs in order to solve
631
00:36:23.559 --> 00:36:27.000
some of it problems. So I was super excited and
632
00:36:27.320 --> 00:36:31.280
very very grateful for the invitation and very hopeful that
633
00:36:31.639 --> 00:36:34.280
from the whole experience, looking at all, right, A, we're
634
00:36:34.280 --> 00:36:37.920
having the right conversations, We're talking about the latest reports
635
00:36:37.960 --> 00:36:40.840
and so on from the UN as well, and we
636
00:36:41.000 --> 00:36:44.119
are we have what seems like the right people. The
637
00:36:44.159 --> 00:36:46.119
only thing I would add, I think next time we
638
00:36:46.159 --> 00:36:49.440
can also include the younger generation and a bit more
639
00:36:49.480 --> 00:36:53.639
of the older generation, because yes, we need to hear
640
00:36:53.679 --> 00:36:54.599
from them as well.
641
00:36:54.480 --> 00:36:58.280
Yeah, yeah, And for anyone who's listening to Ushina describe
642
00:36:58.320 --> 00:36:59.360
that forum going.
643
00:36:59.159 --> 00:37:00.559
Gosh, I wish I had a been there.
644
00:37:00.719 --> 00:37:02.559
We will put a link in the show notes because
645
00:37:02.599 --> 00:37:06.599
a lot of the sessions were recorded and you will
646
00:37:06.639 --> 00:37:08.360
be able to access them.
647
00:37:09.000 --> 00:37:10.199
Jimma, what about for you?
648
00:37:10.280 --> 00:37:12.920
What was your observation about the people in the room
649
00:37:12.960 --> 00:37:14.079
and key takeouts?
650
00:37:14.599 --> 00:37:17.719
I think on the networking night was amazing and it
651
00:37:17.760 --> 00:37:20.599
was great to be able to spot people again. But
652
00:37:21.280 --> 00:37:23.599
the very first day when I came in, I arrived
653
00:37:23.639 --> 00:37:26.320
a little bit late so as talk was already happening,
654
00:37:26.639 --> 00:37:29.400
But the first thing I saw was a woman holding
655
00:37:29.480 --> 00:37:32.639
a baby in the Yeah, And I'm always like, if
656
00:37:32.639 --> 00:37:36.599
there's babies in the space, then it's an inclusive and
657
00:37:36.760 --> 00:37:40.760
open space because it's so easy to shut out mothers
658
00:37:40.800 --> 00:37:44.320
from spaces that if someone's made the effort to include
659
00:37:44.320 --> 00:37:47.840
a mother, then there's something good happening there. So I
660
00:37:47.880 --> 00:37:51.559
think that was really promising, and I really liked, as
661
00:37:51.760 --> 00:37:56.079
Rashina said, like that interdisciplinary approach and people from across
662
00:37:56.119 --> 00:38:00.079
the spectrum, because it felt like the conversation about the
663
00:38:00.119 --> 00:38:04.000
harms to women, the harms to other marginalized groups was
664
00:38:04.039 --> 00:38:07.159
embedded in a larger conversation about the ethics of AI
665
00:38:07.880 --> 00:38:10.599
and what we wanted to do, rather than this niche
666
00:38:10.639 --> 00:38:14.679
conversation which often happens where the women have to go
667
00:38:14.719 --> 00:38:17.519
off into a corner and talk about the women's problems.
668
00:38:17.599 --> 00:38:19.440
But it really felt like that that was part of
669
00:38:19.480 --> 00:38:22.679
the central conversation, not divided off.
670
00:38:23.440 --> 00:38:26.280
Yeah, and I was super impressed too. Our panel was
671
00:38:26.320 --> 00:38:30.599
sort of this first set of panels after the plenary discussions,
672
00:38:30.639 --> 00:38:33.559
and we were up against Helen Tona, who was the
673
00:38:33.599 --> 00:38:36.960
former board member at Open AI, and we had a
674
00:38:37.079 --> 00:38:40.480
full room full of people who came and listened to
675
00:38:40.480 --> 00:38:42.960
our panel. So I was particularly impressed by that. But
676
00:38:43.039 --> 00:38:45.880
also we shouldn't be impressed by that. That should be normal.
677
00:38:49.800 --> 00:38:52.440
We've spoken a lot about the problems and the challenges,
678
00:38:52.760 --> 00:38:55.960
but particularly in the exchange that Jema and I had
679
00:38:55.960 --> 00:38:57.679
a little bit earlier, where we're talking about we can
680
00:38:57.760 --> 00:39:00.719
actually use AI to help smash the patriot right, this
681
00:39:00.800 --> 00:39:02.679
is it doesn't have to be this way, it.
682
00:39:02.559 --> 00:39:05.760
Can be something different. Can you help us to paint.
683
00:39:05.519 --> 00:39:09.960
A picture of what that positive future could look like
684
00:39:10.239 --> 00:39:12.719
if we do, you know, roll up our sleeves and
685
00:39:12.800 --> 00:39:16.199
get involved in shaping AI to be different than it
686
00:39:16.280 --> 00:39:18.800
might currently be. Jemal, why don't I put you on
687
00:39:18.840 --> 00:39:19.400
the spot first.
688
00:39:19.440 --> 00:39:22.679
I think the ideal is that we have a genuine
689
00:39:22.719 --> 00:39:27.199
safety bay design approach to AI, in which any AI
690
00:39:27.320 --> 00:39:32.119
system that is accessible in Australia has had the harms
691
00:39:32.199 --> 00:39:35.920
thought about and managed beforehand, so that it's not used
692
00:39:36.320 --> 00:39:40.519
to ruin people's lives or make them worse, or make
693
00:39:40.679 --> 00:39:45.320
existing harms worse, And that we have genuine accountability mechanisms,
694
00:39:45.360 --> 00:39:50.119
and that those accountability mechanisms are accessible. It's like you say, Johanna,
695
00:39:50.239 --> 00:39:53.039
that we have these laws but we can't use them,
696
00:39:53.039 --> 00:39:57.760
and I think in the ideal situation, those things can
697
00:39:57.800 --> 00:40:01.440
be utilized to hold both people that use the systems
698
00:40:02.000 --> 00:40:06.280
for harm accountable and also the technology companies that produce
699
00:40:06.360 --> 00:40:09.599
systems that are used for harms accountable as well.
700
00:40:09.920 --> 00:40:12.880
What about for you, Rishana, what does a world look
701
00:40:12.920 --> 00:40:14.000
like if we get this right?
702
00:40:14.360 --> 00:40:17.840
I can imagine a world where in the entire AI
703
00:40:17.880 --> 00:40:22.280
ecosystem there is diversity throughout right from the mangoes which
704
00:40:22.280 --> 00:40:26.440
are not which are pretty much shaping the future of
705
00:40:26.519 --> 00:40:30.199
these models, right through to the development teams and the
706
00:40:30.239 --> 00:40:33.199
design teams and the governance teams and the auditing teams
707
00:40:33.880 --> 00:40:36.239
in fact also into it. And I didn't mention this before,
708
00:40:36.280 --> 00:40:40.119
but the investments and the investors and the venture capitalists,
709
00:40:40.119 --> 00:40:42.400
and we know that. And again that is a problem
710
00:40:42.800 --> 00:40:46.320
of only about two percent of women only teams get funded,
711
00:40:46.639 --> 00:40:49.840
even though money is being poured into AI innovation, So
712
00:40:50.079 --> 00:40:52.599
we're being left out there as well. So there's an
713
00:40:52.679 --> 00:40:55.679
innovation and investment part. And then of course when it
714
00:40:55.719 --> 00:40:59.400
comes to women using AI, that they're not actually paying
715
00:40:59.440 --> 00:41:03.320
that AI competency penalty, and they're actually being seen as
716
00:41:03.400 --> 00:41:06.880
progressive in their responsible use of AI, just like the
717
00:41:06.920 --> 00:41:11.400
men are. It's throughout that AI ecosystem that I would
718
00:41:11.519 --> 00:41:14.639
like to see more diversity across all the teams and
719
00:41:14.760 --> 00:41:15.559
all the users.
720
00:41:15.760 --> 00:41:19.360
And I might add one additional vision for us, which
721
00:41:19.440 --> 00:41:22.960
is that I hope that we use AI to identify
722
00:41:23.280 --> 00:41:27.199
bias and discrimination in our society, to help us to
723
00:41:28.519 --> 00:41:32.639
uncover places where we perhaps don't see it happening as much,
724
00:41:32.880 --> 00:41:35.400
and to eliminate that. And I think that is a
725
00:41:35.440 --> 00:41:38.440
real possibility. This is not pie in the sky stuff.
726
00:41:38.519 --> 00:41:43.559
This is within our power to demand that artificial intelligence
727
00:41:43.960 --> 00:41:46.480
be designed in a way that helps us to do that.
728
00:41:47.119 --> 00:41:50.360
I've loved this conversation. Thank you so much, Jemma, Thank
729
00:41:50.360 --> 00:41:54.639
you so much. Forshina will pop some links to both
730
00:41:54.639 --> 00:41:58.599
of your really excellent work and research into the show notes.
731
00:41:58.760 --> 00:42:01.880
And thank you so much for being mentors, for being
732
00:42:01.920 --> 00:42:05.039
part of the AI Safety Forum, and also for giving
733
00:42:05.119 --> 00:42:07.079
your time again to have this conversation.
734
00:42:07.360 --> 00:42:09.599
Hopefully it's the first of many more that we have.
735
00:42:09.840 --> 00:42:10.559
Thank you so much.
736
00:42:10.599 --> 00:42:11.000
Thank you.
737
00:42:19.119 --> 00:42:21.920
Well, that's it for this episode of Tech Mirror, which
738
00:42:21.960 --> 00:42:24.800
is brought to you by the Tech Policy Design Institute.
739
00:42:25.239 --> 00:42:27.480
We're based here in Canberra on the lands of the
740
00:42:27.519 --> 00:42:31.559
number all Nambri people. If you found today's conversation useful
741
00:42:31.639 --> 00:42:34.559
or thought provoking, please do share it with a friend
742
00:42:34.639 --> 00:42:37.760
or a colleague, or leave a review and subscribe wherever
743
00:42:37.800 --> 00:42:41.400
you get your podcasts. If you're watching, please don't forget
744
00:42:41.440 --> 00:42:44.440
to like and follow. For show notes, you can visit
745
00:42:44.480 --> 00:42:48.599
tech Policy dot Au Forward Slash podcast and this podcast
746
00:42:48.679 --> 00:42:52.639
was made possible with thanks to the generous contributions from government,
747
00:42:52.800 --> 00:42:56.880
industry and philanthropy to the Tech Policy Design Fund, the
748
00:42:56.920 --> 00:43:00.360
full details of which are available on our website. The
749
00:43:00.440 --> 00:43:03.519
team at Audiocraft produced this pod on the lands of
750
00:43:03.559 --> 00:43:04.440
the Gadigul people.
751
00:43:04.480 --> 00:43:05.480
Of the Eora Nation.
752
00:43:05.880 --> 00:43:09.880
Music is Vitalia scopellis A big thank you also to
753
00:43:09.960 --> 00:43:13.199
the team at the Tech Policy Design Institute, without whom
754
00:43:13.320 --> 00:43:16.960
this pod would not be possible. Thank you for joining us,
755
00:43:17.360 --> 00:43:20.480
and as always, get in touch and get involved.