AIxMultilateralism: When AI Speaks for Nations - The Ethics of AI Use in Multilateral Deliberations, with Eleonore Fournier Tombs

WEBVTT
00:00:15.838 --> 00:00:21.538
Welcome to AI and Multilateralism, a playlist of conversations here at the Commons,
00:00:22.018 --> 00:00:27.198
our space at the UN Library and Archives Geneva for sharing knowledge on multilateralism.
00:00:27.778 --> 00:00:33.218
In this series, we're joined by experts who help us unpack the many ideas and
00:00:33.218 --> 00:00:37.538
issues at the nexus of AI and international cooperation.
00:00:38.198 --> 00:00:45.058
Today, we're asking the question, is it ethical to use AI in multilateral deliberations.
00:00:45.278 --> 00:00:51.438
We're joined by Eleanor Fournier-Toombs, Head of Anticipatory Action and Innovation
00:00:51.438 --> 00:00:57.438
at the UNU CPR, the United Nations University Centre for Policy Research.
00:00:57.838 --> 00:01:02.478
There, she leads a team providing strategic advice on AI risk,
00:01:02.838 --> 00:01:09.158
policy, and technology strategy to international agencies, regulators, and executives.
00:01:09.498 --> 00:01:14.678
As a data scientist, she's worked in the private sector and across the UN system.
00:01:14.918 --> 00:01:20.778
And we're diving into her research on the ethics of using AI in multilateral fora.
00:01:29.958 --> 00:01:33.578
Eleanor, thank you so much for your time today. Thank you so much.
00:01:33.758 --> 00:01:37.178
We're recording this during the 80th session of the General Assembly.
00:01:37.498 --> 00:01:41.838
It's a busy week there in New York and a lot of issues are being deliberated,
00:01:42.058 --> 00:01:46.538
one of course being about AI and how it can be governed collectively.
00:01:47.038 --> 00:01:50.458
Later today is going to be the launch of the Global Dialogue on AI Governance
00:01:50.458 --> 00:01:54.238
at the UN, which is one of the two major mechanisms recently established by
00:01:54.238 --> 00:01:58.058
the UN to promote international cooperation on AI governance.
00:01:58.338 --> 00:02:01.078
I'm sure you're very interested to see what happens at the launch today,
00:02:01.258 --> 00:02:04.098
but is there anything else you'd like to share about what's been happening this
00:02:04.098 --> 00:02:05.538
week in discussions on AI?
00:02:06.097 --> 00:02:11.117
Yes, of course. And it is a very exciting time for global AI governance.
00:02:11.797 --> 00:02:16.637
Really, the UN started severely leaning into global AI governance in 2023.
00:02:17.397 --> 00:02:22.997
Before that, we had UNESCO that kind of led with the recommendation on the ethics of AI.
00:02:23.197 --> 00:02:27.757
And then it became much more mainstream in terms of the secretariat's leadership
00:02:27.757 --> 00:02:31.517
and the tech envoy's leadership around the summer of 2023.
00:02:31.517 --> 00:02:36.357
And that's when they established the high-level advisory body on AI,
00:02:36.637 --> 00:02:39.577
which our team provided a lot of research support for.
00:02:39.817 --> 00:02:43.317
And the advisory body came together.
00:02:43.317 --> 00:02:48.697
There was 38 members and they came together for one year where they developed
00:02:48.697 --> 00:02:55.237
recommendations on how the UN could take a leadership role in the risks and opportunities of AI.
00:02:55.237 --> 00:02:59.817
And they came up with seven recommendations, a few of which were adopted in
00:02:59.817 --> 00:03:01.737
the Global Digital Compact last year.
00:03:01.877 --> 00:03:05.097
And now, finally, we're at the implementation of these recommendations.
00:03:05.317 --> 00:03:09.857
And those are the Global Dialogues and the International Scientific Panel on AI.
00:03:10.357 --> 00:03:14.437
So very briefly, the International Scientific Panel on AI, which has also been
00:03:14.437 --> 00:03:18.777
announced and adopted, and now we need to select the 40 members of this panel.
00:03:19.557 --> 00:03:25.777
Its aim is to provide global technical consensus on the risks and opportunities of AI.
00:03:26.417 --> 00:03:30.937
And that's actually quite important because it provides a foundation for all
00:03:30.937 --> 00:03:33.337
of the other policy work that's done by member states.
00:03:33.737 --> 00:03:38.377
We do a lot of this kind of scientific consensus, technical consensus at the UN.
00:03:38.777 --> 00:03:43.517
And although there's no kind of strict policy recommendations that come out
00:03:43.517 --> 00:03:47.577
of it, it's not a convention from part of the human rights framework, for example.
00:03:47.577 --> 00:03:52.857
It's still really important work so that we all kind of agree globally on where
00:03:52.857 --> 00:03:55.557
this technology is going and what we want to do about it.
00:03:56.097 --> 00:04:00.197
And then, as you mentioned, today they're going to launch the Global Dialogues
00:04:00.197 --> 00:04:06.737
on AI, which will be a forum for discussion and deliberation every year and on a regular basis.
00:04:07.421 --> 00:04:11.601
To talk about all the different issues around AI. So for me,
00:04:11.881 --> 00:04:15.561
this week, I've spoken a lot about AI and the environment.
00:04:15.861 --> 00:04:19.641
I think this is a kind of upcoming issue that we're starting to look at very
00:04:19.641 --> 00:04:22.101
seriously, especially with generative
00:04:22.101 --> 00:04:25.521
AI tool that consumes so much more energy than other kinds of AI.
00:04:26.081 --> 00:04:30.161
We've also talked a lot about international standards. So to me,
00:04:30.361 --> 00:04:35.521
international standards in AI are a source of a lot of hope for multilateralism
00:04:35.521 --> 00:04:37.861
because Because although we tend
00:04:37.861 --> 00:04:42.781
to see a convergence on the risks and opportunities of AI more broadly,
00:04:43.121 --> 00:04:46.141
there's not a lot of argument when we say, you know,
00:04:46.541 --> 00:04:51.641
there are potential biases in AI or there's privacy considerations in AI.
00:04:51.961 --> 00:04:57.561
And we also say when we want to use AI for the public good and to improve people's
00:04:57.561 --> 00:04:59.921
lives, we're not arguing immensely.
00:04:59.921 --> 00:05:04.081
But I do feel that there's so much fragmentation when it comes to the solutions to that.
00:05:04.081 --> 00:05:08.641
And international standards are a way forward where a lot of parties,
00:05:09.041 --> 00:05:12.121
whether it's governments, private sector, civil society.
00:05:12.821 --> 00:05:17.041
Different member states with completely different legal philosophy and kind
00:05:17.041 --> 00:05:22.001
of historical approach to technology, they're all kind of coming on board and
00:05:22.001 --> 00:05:24.201
agreeing to put out these standards.
00:05:24.201 --> 00:05:29.021
And so I find that this is a time of a lot of hope for AI governance.
00:05:29.221 --> 00:05:32.501
I think it's bringing a lot of countries together and I really hope that we
00:05:32.501 --> 00:05:37.621
can continue making big strides in international cooperation on these topics.
00:05:38.191 --> 00:05:42.331
It must be interesting for you to see how, based out of the AI advisory body.
00:05:42.911 --> 00:05:46.971
These key recommendations for a scientific panel, a global dialogue,
00:05:47.311 --> 00:05:51.271
this has been adopted by the UN and is being really launched and discussed.
00:05:51.751 --> 00:05:55.631
We're going to be talking a lot in this series about this element of the global
00:05:55.631 --> 00:05:57.951
governance of AI and what that means for multilateralism.
00:05:58.191 --> 00:06:02.311
But a lot of your work and your research is actually looking also at how AI
00:06:02.311 --> 00:06:06.251
is changing the practice of diplomacy, the practice of multilateral deliberation.
00:06:06.251 --> 00:06:11.111
And so I wanted to really touch on that today, the use of AI in multilateral deliberations.
00:06:11.611 --> 00:06:17.691
Can we firstly define deliberation? What does this mean? And why is this important to multilateralism?
00:06:18.271 --> 00:06:22.571
Great question. And I've been working on the topic of deliberation for a really long time.
00:06:22.751 --> 00:06:27.091
And what we're discussing today, the use of AI in global deliberations,
00:06:27.311 --> 00:06:32.531
is part of work that is going to be published in Ethics and International Affairs
00:06:32.531 --> 00:06:39.751
in a few months, which was developed with the help of my deliberation co-author, Michael McKenzie,
00:06:40.071 --> 00:06:42.751
who teaches at Vancouver Island University.
00:06:43.151 --> 00:06:47.271
And I'm mentioning him because we've done a lot of publications and projects
00:06:47.271 --> 00:06:48.651
together on deliberation.
00:06:49.131 --> 00:06:53.391
Deliberation is almost everything that really takes place at the UN Secretariat.
00:06:53.391 --> 00:06:59.731
It's a really core function of setting global goals and objectives and norms.
00:06:59.731 --> 00:07:05.491
And it's a process that member states go to, to exchange ideas and perspectives
00:07:05.491 --> 00:07:10.631
so that they can come together in a type of convergence or decision.
00:07:10.911 --> 00:07:14.471
And so we have deliberation in a lot of political processes.
00:07:14.551 --> 00:07:19.551
When you look at national level government, you know, parliaments,
00:07:19.911 --> 00:07:25.431
that's all deliberation. We start to see also deliberative processes in a kind
00:07:25.431 --> 00:07:27.411
of different way taking place online.
00:07:27.671 --> 00:07:32.131
So a lot of people study the way in which we participate and exchange ideas
00:07:32.131 --> 00:07:36.871
on social media and grow our political opinions.
00:07:36.871 --> 00:07:43.351
And so technology has really been closely intertwined with deliberation for
00:07:43.351 --> 00:07:48.931
about 20 years now because we've been using so many technological platforms
00:07:48.931 --> 00:07:50.751
for this shaping of ideas.
00:07:51.071 --> 00:07:58.951
And so with AI, it's really a fascinating topic because AI also generates new ideas and content.
00:07:58.951 --> 00:08:05.911
And I've been very interested in figuring out how that impacts the global deliberation,
00:08:05.911 --> 00:08:10.471
because it seemed to me that we might be losing a little bit of our autonomy
00:08:10.471 --> 00:08:15.491
and our agency by inserting AI into deliberative processes.
00:08:15.871 --> 00:08:19.251
That's kind of where this article is going and this research is going.
00:08:19.781 --> 00:08:24.921
You mentioned that technology has been intertwined and used in global deliberation
00:08:24.921 --> 00:08:27.121
for already about 20 years.
00:08:27.461 --> 00:08:32.421
What is the state of play now? How is AI being used in global deliberations?
00:08:33.061 --> 00:08:36.961
So now we have, we're about two years in for generative AI.
00:08:37.401 --> 00:08:43.281
I would say that before then, the only way that we would really inject AI in
00:08:43.281 --> 00:08:47.721
global deliberations is through the use of predictive analytics,
00:08:47.721 --> 00:08:50.801
where we could have data and bring data to the table.
00:08:50.961 --> 00:08:55.961
So it's very, very different and not really a source of immense concern.
00:08:56.101 --> 00:09:01.481
I think for generative AI, it's a little bit different because we have what
00:09:01.481 --> 00:09:05.961
I've kind of categorized as these formal uses and informal uses.
00:09:06.221 --> 00:09:12.061
And for formal uses, I mean uses of AI that are explicit, that are intentional
00:09:12.061 --> 00:09:13.861
and that are kind of bounded.
00:09:14.801 --> 00:09:19.581
So that could be, for example, the use of an AI tool as a moderator.
00:09:19.901 --> 00:09:25.901
So we've all seen in Zoom calls and in Teams calls that AI agents can attend
00:09:25.901 --> 00:09:28.441
and take a summary of the meeting.
00:09:28.441 --> 00:09:33.901
There's also been some experiments which have been very fascinating on using
00:09:33.901 --> 00:09:41.141
AI in a more separate platform from our regular tools and having the AI tool not only summarize,
00:09:41.141 --> 00:09:46.761
but point out areas of convergence and divergence so that the participants would
00:09:46.761 --> 00:09:53.141
be able to go deeper into these topics and understand where is the position of different players.
00:09:53.141 --> 00:09:58.821
We're very big on inclusion here at the UN, and it is difficult for people to
00:09:58.821 --> 00:10:02.181
come to deliberative forums in person.
00:10:02.461 --> 00:10:07.861
And it's also challenging to participate, everybody at the same time,
00:10:08.081 --> 00:10:10.921
when there's hundreds of people waiting for their turn to speak,
00:10:10.921 --> 00:10:15.301
and are people really discussing amongst themselves, or are they doing kind
00:10:15.301 --> 00:10:17.921
of what we call in deliberation, positional politics.
00:10:17.921 --> 00:10:22.881
So we see a lot of that at the UN this week is that every party is coming in
00:10:22.881 --> 00:10:26.561
with their written script and they're just speaking and then the next person
00:10:26.561 --> 00:10:28.861
speaks and they're not referring to each other's content.
00:10:29.081 --> 00:10:32.921
And that's not good deliberation. That's just kind of positional.
00:10:33.061 --> 00:10:39.681
And so formal uses of AI have a potential to support some of these and just
00:10:39.681 --> 00:10:44.141
improve the quality of deliberation and improve participation and inclusion
00:10:44.141 --> 00:10:45.661
and diversity of perspectives.
00:10:45.661 --> 00:10:48.301
So in that way, I am optimistic.
00:10:48.976 --> 00:10:54.396
Where I wanted to raise the alarm a little bit and just bring diplomats and
00:10:54.396 --> 00:11:00.096
other policymakers on the global stage into this conversation is the informal uses of AI.
00:11:00.636 --> 00:11:07.776
And where I think it's very critical to reflect is where AI tools are writing
00:11:07.776 --> 00:11:15.336
people's speeches and interpreting the speeches and the contributions of others without any filters.
00:11:15.336 --> 00:11:20.616
And we started to see this a couple of years ago and the way in which I started
00:11:20.616 --> 00:11:22.516
to observe this is in duplication.
00:11:22.916 --> 00:11:28.236
So you would see a senior policymaker who maybe had not had time to write speaking
00:11:28.236 --> 00:11:33.376
points, you would hear them speak and then you would hear the next person say the same sentences.
00:11:33.796 --> 00:11:38.836
And the reason why this happened is because, you know, AI tools,
00:11:38.996 --> 00:11:43.316
generative AI tools actually only have a very limited creativity and they can't
00:11:43.316 --> 00:11:45.636
replace what is happening inside your brain.
00:11:45.836 --> 00:11:51.156
And so it would kind of regenerate the same talking points over and over again.
00:11:51.316 --> 00:11:52.956
And so people were repeating themselves.
00:11:53.176 --> 00:11:56.996
That's the only way that you would really know that this is happening.
00:11:57.276 --> 00:12:03.276
So for me, it actually brought up a whole bunch of different issues that are
00:12:03.276 --> 00:12:07.896
a little bit beyond what we speak about with regular AI risk conversation.
00:12:08.216 --> 00:12:11.316
So with regular AI risk, we're talking about things like bias,
00:12:11.516 --> 00:12:13.536
exclusion, we're talking about hallucinations.
00:12:13.776 --> 00:12:18.456
And I think these are all continue to be challenges that we face in AI,
00:12:18.736 --> 00:12:23.636
but there's something around AI in global deliberations that's very unique.
00:12:24.116 --> 00:12:29.216
You mentioned that a lot of discussion today is on the technical risks of AI,
00:12:29.616 --> 00:12:35.436
like bias, exclusion, a lack of languages, hallucinations, environmental issues.
00:12:35.916 --> 00:12:40.216
But you also, in your research, look much deeper at the ethical issues.
00:12:40.436 --> 00:12:45.756
And some that you refer to in your work are ideas like technological determinism,
00:12:46.056 --> 00:12:50.556
sovereignty and self-determination, the actual legitimacy of global deliberations
00:12:50.556 --> 00:12:55.416
themselves, if we continue down this way of mostly informal use of AI and AI agents.
00:12:55.896 --> 00:13:00.556
Could we explore that a little bit further? What do we mean by technological determinism?
00:13:01.183 --> 00:13:06.123
In this work, we've identified three different areas of concern.
00:13:06.283 --> 00:13:08.443
And as you mentioned, there's determinism.
00:13:08.623 --> 00:13:12.963
We also have this idea of state sovereignty and self-determination,
00:13:12.963 --> 00:13:17.123
which I think is a little bit compromised, with AI and authenticity.
00:13:17.583 --> 00:13:22.863
So for technological determinism, that's the idea that all events,
00:13:23.063 --> 00:13:27.803
including human actions and thoughts, are caused by forces external to free will.
00:13:28.003 --> 00:13:33.743
So it's a kind of a philosophical idea, and particularly are caused by technology that we're using.
00:13:33.943 --> 00:13:38.943
And so the idea is that the medium itself and a technological tool that is being
00:13:38.943 --> 00:13:42.723
used is shaping people's ideas and people's free will.
00:13:43.143 --> 00:13:45.203
So to me, this is extremely relevant
00:13:45.203 --> 00:13:51.623
today, because we're having our ideas shaped actively by AI tools.
00:13:51.823 --> 00:13:57.623
And my concern is that because the UN is focused on developing global norms
00:13:57.623 --> 00:14:02.843
and objectives and decisions that are going to shape the future of humanity and of the earth,
00:14:03.043 --> 00:14:09.643
to what extent are these ideas shaped by the technology that we're using itself?
00:14:09.843 --> 00:14:14.383
And I think it's really important to take a step back and say every single action
00:14:14.383 --> 00:14:17.263
where we just put in a prompt and say,
00:14:17.583 --> 00:14:22.403
write my whole speaking points and don't check and don't apply a critical perspective,
00:14:22.403 --> 00:14:27.663
you're not the one who is contributing those ideas into the deliberation.
00:14:27.823 --> 00:14:34.363
An AI tool is contributing those ideas based on training data and random things
00:14:34.363 --> 00:14:38.903
on the internet and whatever guardrails are in place by the generative AI tool.
00:14:39.143 --> 00:14:43.783
And in a deliberation, the convergence and the consensus is based on all of
00:14:43.783 --> 00:14:47.623
the individual inputs and then the way in which they kind of mesh together and
00:14:47.623 --> 00:14:50.363
that people discuss and come up with new ideas.
00:14:50.643 --> 00:14:54.683
And so if each of the individual inputs start becoming compromised in this way,
00:14:54.863 --> 00:14:58.823
I think it's really a source of concern and something that we should take a
00:14:58.823 --> 00:15:02.343
step back and really think about the impact that that might have.
00:15:02.966 --> 00:15:08.066
So if we're all using AI in writing speeches, in how we deliberate,
00:15:08.386 --> 00:15:14.486
in multilateral settings, the technology itself is thinking for us. So that's my concern.
00:15:14.846 --> 00:15:18.286
We certainly do not have to use AI that way.
00:15:19.546 --> 00:15:24.266
Generative AI tools have a lot of power and can do a lot of different things.
00:15:24.406 --> 00:15:28.706
And so I'm a data scientist. I'm not against technology in any way.
00:15:28.706 --> 00:15:34.106
But I don't think that we should allow generative AI tools to do our thinking
00:15:34.106 --> 00:15:39.566
for us, especially when we're active participants in global deliberations.
00:15:39.906 --> 00:15:43.786
So a generative AI tool can help in a lot of different ways.
00:15:43.786 --> 00:15:47.626
It can help with a little bit of research. It can help to structure ideas.
00:15:47.906 --> 00:15:53.226
It can help to summarize what you wanted to say, maybe to some speaking points.
00:15:53.226 --> 00:15:58.286
But every time that people are entering a deliberation unprepared and have not
00:15:58.286 --> 00:16:02.606
really thought what they wanted to say, and they're just putting a prompt or
00:16:02.606 --> 00:16:06.566
putting a speech by another member state into AI and getting a response,
00:16:06.786 --> 00:16:10.166
that's where you're losing your own participation.
00:16:10.166 --> 00:16:17.046
And I think that then goes to another level because the UN is also a forum for
00:16:17.046 --> 00:16:18.466
deliberation for member states.
00:16:18.726 --> 00:16:25.706
And so each state, each country that is part of the UN is responsible for their
00:16:25.706 --> 00:16:30.106
own country and for the decisions that they're going to be participating in.
00:16:30.106 --> 00:16:35.486
So the UN is not a global government that has authority over its member states.
00:16:35.706 --> 00:16:40.726
The member states ultimately have their own freedom and self-determination.
00:16:40.926 --> 00:16:45.046
And so they willingly participate in negotiations. They willingly participate
00:16:45.046 --> 00:16:51.726
in conventions and agreements and in shaping ideas at the UN.
00:16:51.726 --> 00:16:58.826
And so if the positions of member states are drafted by AI agents rather than
00:16:58.826 --> 00:17:03.066
by a diplomatic representative of the citizens of that state,
00:17:03.266 --> 00:17:06.546
the states themselves are not being authentically represented.
00:17:07.346 --> 00:17:12.746
And again, what we would see if we took this exercise to the fullest,
00:17:12.986 --> 00:17:18.066
we would see that each state is actually representing the same kind of ideas
00:17:18.066 --> 00:17:20.666
mashed up from the internet.
00:17:20.746 --> 00:17:27.546
And we can also add that there is certainly only a few states that provide the
00:17:27.546 --> 00:17:32.626
generative AI tools and only a few states that are overrepresented in the training data.
00:17:33.232 --> 00:17:38.072
So that would mean that you would have an over-representation of the ideas coming
00:17:38.072 --> 00:17:43.832
from these states, these kind of AI powers, and an under-representation of ideas
00:17:43.832 --> 00:17:47.152
coming authentically from the other member states.
00:17:47.152 --> 00:17:51.292
And so you would have a very, very strange situation from my perspective,
00:17:51.652 --> 00:17:59.152
where the real needs and perspectives of the citizens and residents of UN member
00:17:59.152 --> 00:18:05.172
states would no longer be very clearly articulated and represented on a global stage.
00:18:05.172 --> 00:18:11.072
So on that note of self-determination and looking at how global deliberations
00:18:11.072 --> 00:18:16.392
are a space for all countries, all member states to bring forth their ideas,
00:18:16.672 --> 00:18:22.672
their values, their priorities for global development, what would you say is
00:18:22.672 --> 00:18:27.552
the biggest risk in terms of self-determination if we continue to use AI in these ways?
00:18:27.552 --> 00:18:29.772
I think there's two things.
00:18:30.012 --> 00:18:34.012
One of them is that we're going to run out of good ideas.
00:18:34.312 --> 00:18:38.452
I think the only way that we can have really good and long-lasting solutions
00:18:38.452 --> 00:18:43.732
at the UN is if we have real diversity of perspective and creativity.
00:18:44.352 --> 00:18:49.212
And I just don't see that with a generative AI tool at scale.
00:18:49.352 --> 00:18:50.852
I think we have to represent ourselves.
00:18:51.512 --> 00:18:56.672
And the other thing is loss of legitimacy. So if we see that member states are
00:18:56.672 --> 00:19:02.552
not really representing their citizens, the citizens are going to lose faith in the organization.
00:19:02.932 --> 00:19:07.092
And that's something we sometimes forget at the UN is that this is member state
00:19:07.092 --> 00:19:08.872
contributions that's running the UN.
00:19:09.012 --> 00:19:11.952
And that comes from income tax and from other kinds of taxes.
00:19:12.392 --> 00:19:16.532
And so if people feel that they're no longer being authentically represented,
00:19:16.532 --> 00:19:18.212
they're going to lose faith in
00:19:18.212 --> 00:19:21.892
the UN system. And so we're threatening the multilateral system at large.
00:19:22.573 --> 00:19:27.193
You mentioned authenticity. Can we explore a bit more what this means in terms of our agency?
00:19:27.373 --> 00:19:32.313
Because you also look at something called epistemic agency and genuine agency.
00:19:32.833 --> 00:19:35.693
Can you break that down for us in terms of what it means for multilateralism?
00:19:36.493 --> 00:19:41.653
Yes. So authenticity is, again, one of these kind of political theory terms,
00:19:41.853 --> 00:19:46.653
which means the degree to which a political representative is considered to
00:19:46.653 --> 00:19:50.973
be genuine and true and aligned with what they say that they're going to be done.
00:19:50.973 --> 00:19:56.073
So it's very important when in democracy, we elect political representatives
00:19:56.073 --> 00:20:02.013
that they're believable and we can consider them authentic to the ideas that they're representing.
00:20:02.413 --> 00:20:07.573
So either they're not necessarily part of a community, but they understand the
00:20:07.573 --> 00:20:11.113
community and are going to really represent their perspectives.
00:20:11.133 --> 00:20:15.813
Or often we like electing people that are a member of a certain community that
00:20:15.813 --> 00:20:20.053
we feel they've experienced, they have lived experience. And so that's where
00:20:20.053 --> 00:20:22.513
authenticity is really critical.
00:20:22.773 --> 00:20:27.173
And this idea of agency is very connected to this.
00:20:27.353 --> 00:20:33.073
Agency is the idea that we have an impact on the political decisions that are
00:20:33.073 --> 00:20:35.213
being made. We can choose. We have choice.
00:20:35.473 --> 00:20:42.053
And if political decisions are made by non-human forces, we lose agency in terms
00:20:42.053 --> 00:20:45.433
of where we're going to go as a world.
00:20:46.078 --> 00:20:51.418
And I think that's really important because we shouldn't feel that we're kind
00:20:51.418 --> 00:20:55.238
of passively watching the global and geopolitical events.
00:20:55.398 --> 00:20:59.298
And I think a lot of the times these days, there's just been a lot of change
00:20:59.298 --> 00:21:04.518
and a lot of movement and a lot of concerning events on the geopolitical stage.
00:21:04.798 --> 00:21:09.358
And many people have felt a loss of agency, both from a political perspective
00:21:09.358 --> 00:21:11.418
and from a technological perspective.
00:21:12.078 --> 00:21:18.358
So, so many conversations that I've had have been around this idea of AI is
00:21:18.358 --> 00:21:20.298
coming and there's nothing I can do about it.
00:21:20.398 --> 00:21:25.618
I'm kind of losing my sense of being able to shape the life that I have.
00:21:25.798 --> 00:21:31.658
And so it goes down not only to at the global level, but also at the very individual
00:21:31.658 --> 00:21:37.458
level where people are saying, I don't know the kind of world that my children are going to grow up in.
00:21:37.458 --> 00:21:41.678
And I don't know if I can really guarantee that I'm going to be able to pay
00:21:41.678 --> 00:21:47.818
for my rent and have meaningful work because I feel that I'm losing my sense
00:21:47.818 --> 00:21:50.098
of self and agency to these AI tools.
00:21:50.398 --> 00:21:54.458
And so this issue that we're seeing at the multilateral level,
00:21:54.618 --> 00:21:57.478
it really drills down also to the individual level.
00:21:57.618 --> 00:22:02.278
And that's why we need to understand our approach to these tools and really
00:22:02.278 --> 00:22:04.918
change our mindset when it comes to using them.
00:22:04.918 --> 00:22:10.278
When we look at how we approach it, because AI is part of our world everywhere,
00:22:10.558 --> 00:22:14.158
and you mentioned also when it's formally embedded, the real benefits it can
00:22:14.158 --> 00:22:15.738
bring to multilateral deliberations.
00:22:16.358 --> 00:22:20.098
Thinking of member states that maybe don't have the same resources as bigger
00:22:20.098 --> 00:22:24.658
ones, what are the risks that we don't use AI and how can it be used in a way
00:22:24.658 --> 00:22:26.998
that's of benefit to all in multilateral fora?
00:22:27.678 --> 00:22:32.118
So that's a great question. And actually, it's very unique to the structure of the UN.
00:22:32.418 --> 00:22:35.898
And it's something that we do bring up also, because it needs to be addressed,
00:22:36.078 --> 00:22:38.458
is the question of resources.
00:22:39.038 --> 00:22:46.838
So one of the reasons that member states might adopt generative AI is because of productivity gains.
00:22:46.838 --> 00:22:51.758
Many member states that are represented in the UN, so in New York and in Geneva,
00:22:52.098 --> 00:22:54.598
have very, very small permanent missions.
00:22:54.978 --> 00:22:59.338
There might only be one person or two people that are really following a number
00:22:59.338 --> 00:23:02.098
of different issues, and they're very, very, very strained.
00:23:02.758 --> 00:23:08.138
So what they will argue is that if they're able to use generative AI to accelerate
00:23:08.138 --> 00:23:14.078
their work, then they will be more often represented in these different deliberative
00:23:14.078 --> 00:23:16.138
forums because otherwise they can't even attend.
00:23:16.538 --> 00:23:22.958
This is an issue because the G7 countries, for example, have really big permanent missions.
00:23:23.238 --> 00:23:27.338
So this is a little bit of a trap where the under-resourced countries are going
00:23:27.338 --> 00:23:30.698
to say, well, this is increasing our capability and our capacity.
00:23:31.398 --> 00:23:34.778
However, we need to keep in mind, again, how it's being used.
00:23:35.018 --> 00:23:40.918
If we have an AI agent that is now writing the perspective of a small member
00:23:40.918 --> 00:23:45.038
state, the member state may believe that now they're being authentically represented
00:23:45.038 --> 00:23:49.178
to a greater extent and there's more inclusion on the global stage.
00:23:49.298 --> 00:23:55.238
But that is not necessarily the case because whose perspective is that agent really representing?
00:23:55.238 --> 00:23:59.478
That needs to be acknowledged, first of all, because it will be a huge incentive
00:23:59.478 --> 00:24:02.958
for adoption, particularly the informal uses.
00:24:02.958 --> 00:24:07.518
So once we acknowledge this idea and acknowledge that if it's not used properly,
00:24:07.518 --> 00:24:12.758
it could actually have the inverse effect of what smaller member states are hoping to achieve,
00:24:12.978 --> 00:24:18.438
then we can start thinking about solutions and hopefully being able to still
00:24:18.438 --> 00:24:23.918
increase productivity and use a very powerful tool, but without really undermining
00:24:23.918 --> 00:24:27.858
state representation and sovereignty and agency.
00:24:28.542 --> 00:24:33.782
So let's talk about solutions then. How can we ensure that AI can be leveraged
00:24:33.782 --> 00:24:38.042
ethically, fairly, sustainably in global deliberations?
00:24:38.202 --> 00:24:40.002
What do we need to do to make that happen?
00:24:40.422 --> 00:24:47.382
I think that addressing the specific use of generative AI tools right now is really important.
00:24:47.862 --> 00:24:51.142
There's a couple of broader issues that really need to be addressed.
00:24:51.142 --> 00:24:57.222
For example, the whole question of privacy and the type of data that the AI
00:24:57.222 --> 00:24:58.682
tools are being built upon.
00:24:59.262 --> 00:25:05.962
So increasingly, we're finding solutions where AI tools are able to be enclosed,
00:25:05.962 --> 00:25:11.062
where what happens in the relationship or the kind of interaction with the AI
00:25:11.062 --> 00:25:14.982
tool is not going back to the kind of main training model.
00:25:14.982 --> 00:25:17.602
And that's really important in terms of confidentiality.
00:25:17.902 --> 00:25:22.462
That's something that needs to be brought up as well. And that's why uses of
00:25:22.462 --> 00:25:25.862
AI need to be disclosed so that you can make sure that they're secure.
00:25:26.482 --> 00:25:31.602
There's a possibility increasingly of having AI tools that are trained on your
00:25:31.602 --> 00:25:36.062
country's own documentation and policy frameworks.
00:25:36.342 --> 00:25:39.382
That's extremely important. You can only do that if you're intentional.
00:25:39.862 --> 00:25:44.582
And so what we propose in our research is this idea of ethics by design.
00:25:44.982 --> 00:25:49.502
And to really thinking through, how are we going to avoid these traps?
00:25:49.822 --> 00:25:55.142
And for me, it's a question of intentionality. So this idea of designing the
00:25:55.142 --> 00:25:59.762
tool that is going to be used by diplomats and policymakers in such a way that
00:25:59.762 --> 00:26:05.762
it's very uniquely connected to the data and information and policy positions
00:26:05.762 --> 00:26:08.322
that you want to represent.
00:26:08.642 --> 00:26:12.982
So that's more of a technical perspective, but it can only be done if people disclose.
00:26:13.716 --> 00:26:18.236
So the question of disclosure, I think, is also important. We're in a transition
00:26:18.236 --> 00:26:23.276
period in our use of AI, where people are a little bit worried about being called out.
00:26:23.616 --> 00:26:28.336
We have a very kind of mixed approach to AI, because on the one hand,
00:26:28.476 --> 00:26:33.696
we're saying, let's have wide-reaching AI adoption, and let's have inclusion,
00:26:34.156 --> 00:26:36.556
and bridge the digital divide through AI.
00:26:36.776 --> 00:26:41.776
And on the other hand, when people are using AI in their day-to-day work, they can be shamed.
00:26:41.776 --> 00:26:47.236
So the idea of disclosure is something that I saw at the Brookings Institution.
00:26:47.236 --> 00:26:53.856
They published a really interesting disclosure policy where there's open dialogue
00:26:53.856 --> 00:26:57.536
inside a policy team and a research team on use of AI.
00:26:58.076 --> 00:27:03.416
And researchers explain and talk about where they're actually using AI so that
00:27:03.416 --> 00:27:08.436
it can be brought to light. So we don't necessarily need to disclose our use
00:27:08.436 --> 00:27:10.216
of AI in a deliberative process.
00:27:10.616 --> 00:27:13.336
You don't need to stand in front of the general assembly and say,
00:27:13.516 --> 00:27:15.496
I use AI to write my speech.
00:27:15.496 --> 00:27:21.196
But certainly thinking through how am I going to talk about this with supervisor
00:27:21.196 --> 00:27:25.816
peers and make sure that it's being used responsibly is super important.
00:27:25.816 --> 00:27:32.016
And it'll help take away the shame, which also helps use it from a very critical
00:27:32.016 --> 00:27:36.856
perspective and not kind of like as a blanket thinking for me tool.
00:27:37.036 --> 00:27:38.276
So that's really important.
00:27:38.841 --> 00:27:44.041
And I think what I really want to do with this is also start having dialogue
00:27:44.041 --> 00:27:46.081
around AI tools and deliberations.
00:27:46.361 --> 00:27:50.541
I think we need to talk more about this and be more open and honest.
00:27:50.741 --> 00:27:55.061
I do feel that this idea, for example, of AI tools now being kind of invited
00:27:55.061 --> 00:27:59.561
to a lot of dialogues at the UN, especially the online ones,
00:27:59.701 --> 00:28:03.661
some of which are formal dialogues, member state mandated formal discussions,
00:28:03.701 --> 00:28:05.021
and you have an AI tool there.
00:28:05.241 --> 00:28:10.121
People have not been choosing this. There hasn't been any conversation about
00:28:10.121 --> 00:28:11.861
what is this AI tool doing?
00:28:12.261 --> 00:28:16.701
Where is this information and data doing? Has there been any consideration of
00:28:16.701 --> 00:28:19.021
the data leakages that are happening?
00:28:19.501 --> 00:28:23.821
So that's where I think we need to have a lot more conversation and be very,
00:28:23.941 --> 00:28:28.521
very intentional and specific in the way in which we're going to use AI.
00:28:29.141 --> 00:28:32.781
I'm sure there's a lot that's going to happen in this space as you work on this.
00:28:33.241 --> 00:28:36.001
Thank you so much for sharing all of this knowledge with us today.
00:28:36.121 --> 00:28:39.101
We wish you the best as you continue this research and look at the ethics of
00:28:39.101 --> 00:28:40.861
AI use in deliberations.
00:28:41.221 --> 00:28:46.481
Is there an open source resource you'd like to share with our listeners if they'd like to learn a bit more?
00:28:47.001 --> 00:28:53.061
Yes, I'm happy to share the blog post on Carnegie Council for Ethics in International
00:28:53.061 --> 00:28:56.981
Affairs where some of these ideas began to be developed.
00:28:57.501 --> 00:29:02.621
And our paper will be published shortly. And so I'll make sure to share it so
00:29:02.621 --> 00:29:04.961
that you can add it to the podcast page.
00:29:05.221 --> 00:29:10.141
I think there'll be many more of these publications coming up soon. Really interesting.
00:29:10.681 --> 00:29:13.801
Eleanor Fornia-Toumbs, thank you so much for joining us. See you soon.
00:29:13.861 --> 00:29:16.281
Take care. Thank you so much. It was great to be here.