← Humanitarian Frontiers
Loading summary
Transcript87 lines
- [00:03]
A
Welcome to Humanitarian Frontiers in AI, the podcast series where innovation meets impact. In each episode, we dive deep into how artificial intelligence is reshaping the future of humanitarian work. From enhancing crisis response to making aid delivery smarter and more effective, AI is opening new doors in the way we support communities in need. In this series, hosts Chris Hoffman and Nassim Motelaby bring you thought leaders from academia and the tech industry to discuss not only the vast opportunities AI offers, but also the ethical considerations and risks we all must navigate. Join them on this journey as they explore AI's potential to transform lives and address humanity's most pressing challenges.
- [00:49]
B
All right, well, we're here. Naseem, nice to see you. It's our first episode.
- [00:54]
C
Hi, Chris. I'm so excited. We are kickstarting our amazing podcast with really good speakers today. Can't wait to introduce them.
- [01:02]
B
I know, it's going to be so exciting. I'm a little nervous. We've got 10 more episodes to get through. This is the beginning of the first one, so let's work on it. I mean, we've got some great people, as you say, and I just couldn't be more excited to have this great panel on with us.
- [01:18]
C
Yes, I'm sure that there's going to be a great discussion with really good questions coming out of this specific episode. We're going to talk about how do we strategically think about AI? What are the problem areas, but how can we be realistic and actually make progress? So let's see how it goes.
- [01:35]
B
I can't wait to see how it goes. Look, we've got Nick Thompson from the Atlantic, we've got Hova Gatimezian from unhcr, and we've got Michael Chave from former Microsoft and now with his own consultancy on humanitarian AI. So we've got a great team here. Let's do this.
- [01:50]
C
Let's go for it.
- [01:52]
B
Yeah. Hello, everybody and welcome. I'm Chris Hoffman. I'm your co host here of Frontiers in AI in the humanitarian sector. And I'm joined by a very distinguished panel for this first podcast of the series. We're going to do a 10 podcast series on AI in the humanitarian sector and really look at the different aspects of AI and how it can be addressed. What are the things we need to be thinking about as a sector and to really help those influencers at the highest level start to understand more clearly what are the potential impacts of AI from a positive standpoint and being able to help people at scale, but also what are the things we need to be aware of? What are the things we need to consider and what are those baby steps we need to take to get to scale? And so I'm joined by Nasim Motalebi here as my co host. And Nassim, I'm going to hand it over to you and first question's on you. First question of the podcast. We've got a long way to go over these next 10 sessions, so over to you. Start us off.
- [02:47]
C
Thanks, Chris. I'm super excited to kickstart the discussion and we have a very diverse panel and I think we can bring different perspectives into this conversation. But I wanted to actually start with Nick and I have a question for you to get us thinking and perhaps transition into a heated discussion around AI. But my question for you is when and where was this turning point for AI? I think you have a lot of good thoughts and opinions, I should say, around AI. So I wanted to actually ask you, when do you think this discussion around started? Especially that for a long time we have longed for artificial general intelligence AI robots that can be human companions or mimic human behavior. We've anthropomorphized AI for such a long time across diverse tasks, and we've changed into this area where we can see a shift that we want AI to just do automation now or do what we don't want to do, or make our lives easier in some way by doing things that we can't do, like insane computations. So what is this turning point? Or do you see it as a turning point? What do you think fascinates people around AI these days generally? And is it human likeness anymore? Or is it that it provides a powerful engine for us to do impossible computations? And do you think we still strive for that robot or human computer interaction like we used to in the 60s or in the 70s? So, yeah, over to you, Nick.
- [04:23]
D
Okay, that's a big, great question. So the fundamental ideas date back to the 1940s and the 1950s. The fundamental work in building the architecture dates back to the 1980s. But it was really in 2015, 2016, where the work started to become impressive, maybe a little before then, when suddenly AI could solve really complicated problems in cognition better than humans could, when it could recognize images as well as humans are better than humans. And then, of course, it was in December of 2022 when OpenAI released ChatGPT. And suddenly it was apparent that there had been about 10 years of progress in one year, and everybody knew that moment was coming. But it happened much more quickly than expected. And since then, we've seen steady Progress in generative AI over the last two years, it may be leveling off, which would probably be a good thing, even though I'm an optimist, because you don't want the speed of change to come faster than society can handle it. So where we are right now is we have these incredible machines, these large language models that can, for better or worse, mimic human intelligence in many domains, exceed human intelligence in certain domains, and that are accelerating rapidly. And the challenge for society is to get as many of the benefits, the ability to learn more quickly, to solve comprehensive cognition problems, to organize data in massively more efficient ways, to get that without losing many of the safeguards and protection that society has, many of the things that hold us together as humans, individuals, and many of the protections of civil society that will all be threatened when suddenly machines that can mimic and act just like humans, or when you have computers that can talk just like humans and systems that can make voices that sound just like humans. So we're entering a wonderful, beautiful, exciting, also scary era.
- [06:35]
B
Totally. I mean, Michael, when you were working at Microsoft and now what you're doing, there's always been this question, and Hovig, probably from your side too, that beneficiary facing versus internal tools. Right. And there's this wide acceptance that we've been using AI in the sector. Right. Already in a lot of the tools that we use, whether it is through Microsoft or Google, that we're interacting with it, whether we know it or not. And we're okay with that sort of in the sector. But when we start talking about people using it, and Nick was just alluding to the fact that how are we doing it and how are people interacting with it? But there's a lot of reticence that I'm hearing in the sector. They're like, I'm fine with it, doing my analytics, but don't let it talk to somebody in South Sudan. And so what are you guys feeling about that? Is that a. Is the pace that Nick is talking about? It's still happening, but the humanitarian sector is like, whoa, whoa, whoa. Breaks, breaks, breaks, breaks. Or are they ready to also dive in?
- [07:32]
E
Yeah, good question. And there are many dimensions to that. I think you're right that there is healthy caution across the humanitarian sector around the use of AI, which is really good to see. There are some organizations, individuals that are sort of jumping in and sort of really trying it out, and that is also good. But of course, you really need to understand sort of how AI works, how it makes decisions, how it makes mistakes. So that you can proactively catch some of the errors that you will definitely see with AI. It's imperfect, it will always be imperfect. So the healthy skepticism I think I have seen primarily or sort of characteristically from the humanitarian sector is good. But it's also, I think it's important that the sector at large finds a way to lean in on the potential of what AI can provide. You're absolutely right that it can provide a lot of back office staff efficiency in terms of just improving or automating partially processes across the teams, but it can also enable completely new scenarios that are essentially impossible without AI. So it's finding the right balance between making sure you lean in on this opportunity, on this moment, on what AI is good at, but you also understand what it is you're interacting with and you make the point of the internal looking. This is what we use as, say, productivity tools within the team, within the organization. There are obvious benefits there and the ramifications, the potential negative consequences of applying an AI powered solution into an ecosystem, including how it impacts the beneficiaries, the populations you serve as a humanitarian entity there you really need to think more deeply about what can go wrong. How can you really benefit from AI, but also what happens when the output, the generated output of AI is incorrect?
- [09:33]
F
Yeah, for sure.
- [09:34]
B
Hovik, what's going on at hcr? What are they thinking? I talked to Samira in the New York office the other day and she's like, well, we've got a policy that tells us what we can and can't do, and we've got a training that tells us what we should be doing. But that's kind of where we are. Right. Are you seeing from a strategic point of view that we're moving beyond that, or is it really just garden fencing happening right now?
- [09:56]
F
I mean, it's a very interesting space that I think we in the humanitarian sector have been struggling in the analog world with our do no harm principles. So were we done with us applying our do no harm principles in the smartest way possible? Sometimes yes, sometimes no. And so now transitioning into the do no harm principles in the digital space at the same time looking at the opportunities that any change and transformation brings into helping us do better? Right. And I think we are like we were in the analog world in this dual space where there's opportunities to do good and we have to be careful not to do harm. There was this transition phase, pre AI, I would say, where we started using digital tools to better work with communities. And have we made mistakes in that space, yes. And now we are at a faster speed, I would say in that space of basically both opportunities and also kind of challenges in respecting our dual ARM principles and translating those into digital space. So I think we are there now. Are things moving? Yes. On the self serving parts, there's a lot happening in this space, but I don't think we need to talk about that. I think you rightfully pointed out let's talk about the people we serve and the communities we work with. And there I think I would say we have to be clever about how we approach communities. I think part of the risks I'm seeing is oh, cost saving measures. So let's cut corners, let's cut costs. We don't need that many staff anymore. We can just have the AI kind of replace staff members. And I mean you start hearing these kind of thoughts on efficiencies. Whereas I think maybe the premise is what is the problem we're trying to solve and be a bit more tech agnostic, if that's even possible in this world now. But just to say we have challenges, can we use technologies for good to solve those problems? Challenges and where do refugees and displaced populations fall in this dynamic? And I think how involved are the people we serve in this process? I think that is a key factor in kind of sense checking what we're doing. I think there have been mistakes and we continue making mistakes on testing technologies on people as opposed to finding a challenge with people and use those technologies with them, not on people. And I think it's important distinction to make especially when we're talking about the use of AI and technologies to improve humanitarian work and refugee in the space facing solutions. How much is the community involved? How much are they aware? How much are they part of the design of this? And I just want to give one example. So for us to be also positive about this. You know, Chris, we have a refugee led innovation fund that funds refugee led organizations in designing their own solutions to the problems they identify. We actually have one of the projects in France led by a refugee led organization that is using AI to improve access of refugees to the banking sector in France. So that's one example of refugees themselves leading the process and us acting as support. If we have more of that in the sector, you then calibrate that do no harm principle. Right. And you bring the real clients into the picture. I think the problem we have in our sector is that our main clients are not the ones who are procuring the services, they're the ones who are receiving the services. Through another party, which is us, who's procuring the services on their behalf. So how do we break that and say, well, no, even if they're not paying for it, they are the clients, so they're part of the design of the solution? It's not a theoretical exercise. I think there are means and ways of doing that. I spoke too much, but, yeah, things are happening. I think it's in the right direction. With the caveat that we have to be very careful on how much we're including communities.
- [14:14]
E
Just to follow up to this, I think the do no harm principle to humanitarian actions is so topical, so critical to this discussion around AI. When you're working with modern AI, when you're working with technology that's disruptive at the societal level, you really have to think about how you use it responsibly and ethically and that it is used in an informed manner. And so a key aspect I want to focus on is the cost of error. So, as I mentioned, AI makes mistakes. It will always make mistakes. Even as it gets better, we should anticipate and expect it to continue to make mistakes. But the cost of error means that you need to take into consideration what happens when it goes wrong. What are the downstream consequences of an incorrect AI recommendation or AI output? So, for example, if you have AI behind that, what shows, let's say, in your Netflix queue, what it's expecting or recommending you see next, or in your Amazon shopping experience, what it's recommending that you would buy? Errors, there are very low cost, little consequence. But if you use AI, for example, to help identifying areas where landmine operators should go out for land clearance, an AI recommended error there could be catastrophic.
- [15:35]
F
Right?
- [15:36]
E
So you need to make sure that you understand what's the cost of an error. And then when it has significant negative consequences potential, then make sure you have a human expert in the loop so that the final decision is a human makes that decision, that the AI generated output is just a recommendation, but it's actually a decision that's made by human.
- [15:57]
D
Michael, this is a super interesting example because don't you also have to weigh what are the benefits? So take the landmine and let's say that a human looking at a field, taking all the data is 90% accurate. And an AI looking at the field, again, it's a hypothetical, is 99% accurate. And so the AI is still not perfect, but it is better than the human. Would you be willing to deploy the AI in that situation?
- [16:22]
E
I would say it's more of a Question of at which point do you maintain a human expert in the loop? If it's very low consequence, like recommendations for what to watch next, you don't really need a human expert there. But if there is significant negative consequence of acting on an AI recommendation without human interaction, then that's where you need to keep humans in Loomis. So yes, you can have the false positive, false negatives, you can optimize for that, given the data, the quality of the model you have. But understanding when we talk about do no harm, what is the possible harm that can come of. I don't want to say blindly necessarily, but acting on an AI recommendation without using human expertise.
- [17:02]
F
This is super exciting, Chris. So we might just hijack the whole show and nothing. As I'm listening to both of you speak, I think maybe also two examples, maybe to talk about what you should do, what you shouldn't do, how much the human is involved in this. You know, in, in the Middle east we have quite advanced vulnerability assessment frameworks where we actually assess the vulnerability of the people we serve. And this is done together with, it's a huge human effort of getting to know the families, meeting them face to face, talking to them, build trust so that you understand and after a few times you meet with them, then you really truly understand the family, the composition, the people. And they're just not numbers, they're humans. So do we want to let go all of those staff who are interacting with the people and building the trust and have an algorithm just decide based on sample data? I would say probably not. But can we use advanced AI to better analyze the data collected? Absolutely. We already are using chatbots to interact with refugees who call, who have basic questions. So maybe the basic questions that have no sensitivity can be answered. But the moment there is an advanced question and there's a human interacting with the person, so you refer, so you to develop knowledge, enough knowledge to decide where do you stop using it, where do you start when you stop using it. We're also looking at, for example, we do refugee status determination that has a life changing consequence on asylum seekers, whether they are recognized as a refugee or not. Some governments are using already advanced AI tools sometimes to implement conservative immigration policies. So if the computer makes a decision, no one asks any questions and that has a life changing experience on the applicant. Maybe we use AI to better do research on the country of origin information, but it's a human that is interacting with the human to understand their story. So that that decision, ultimately the human takes also responsibility of this decision. The Point I'm trying to make is how advanced are we in determining where do we start and where do we end in the use of AI? The more we are actually becoming experts in that, I think the better we will be able to make those decisions. When you use and when we use, where we stop the use.
- [19:37]
C
I just love this conversation and it's just ringing a lot of bells. One thing is that it points to the conversation around how the humanitarian space for a long time has been a testing lab for a lot of technologists and technology corporations. And I think we speak about the harms of AI in this context. One thing that I think we don't speak enough, especially when we're using AI to solve beneficiary problems or affected populations issues, so on and so forth, is that we don't actually know the value of AI in this context. We haven't even done needs assessments, and we don't even know what value can AI bring to the affected populations. And I think it goes hand in hand with what Hovik is saying, is that we need to design with the people and understand what they would want. And I want to tie this back into why the adoption of AI in the humanitarian space right now is flourishing. And I think it's because we know what we want as organizations, oftentimes, which is efficiency, which is targeting. And a lot of the AI solutions that we deploy, I think it's because we want better systems for ourselves when it comes to aid delivery, versus better lives and livelihoods for those who we're working with. So I think that juxtaposition for me is always interesting. And I wonder if this ties into this whole idea going back to the speed of AI development or technology development generally. There are two schools of thought, right? Some people think that we should stay back and watch technology unfold, and then we, when it goes to the test of time and proves itself valuable, then we deploy it, and then that's, I think, when we get into the testing lab scenario. And I'm curious to know what we think in this room. But the other school of thought is we should be pioneers and decide where technology and how technology should be developed for the humanitarian sector. So I just wonder, what do you think generally around this? How can we stop playing catch up and if we should even do that, or should we say, okay, let's assess the harms of AI, the value of AI and then move more strategically around this.
- [21:42]
D
I'll just make a quick, strong view that I think that the humanitarian sector should find the most beneficial use cases that it can think of, should push very hard, should be as innovative as possible, and should try to drive AI that way. And I think that there are lots of places, whether it's language translation, whether it's figuring out how to give information to refugees about how to get the right paperwork, whether it's about upskilling nurses or family members who have to take care of people in crisis situations, where the ability of AI to massively increase the capability of the people who use it can be extremely helpful. And that my fear would be that the humanitarian world, because it's so intrinsically cautious, because it's so intrinsically skeptical of big technology, ends up leaving these powerful tools in the hands of everybody else, including the entities that are causing the humanitarian problems. And so I'm definitely on the side of use it carefully, use it smartly, but use it aggressively.
- [22:54]
B
Nick, just to follow up a little bit, because we're talking more about the humanitarian side of things when we think about crisis, disasters, et cetera. But then we take the other side of this, which is human rights, which is information as aid. Other sides of things, where we're trying to make sure that people have a voice that they're able to express themselves, they're able to get the right information. Given a lot of the crises that are happening just right now, there's a lot of disinformation that's out there. And so being that you're in that space of journalism and reporting and information, what are things that we need to be thinking about as we create these tools? Because you used an example of helping refugees. And I've built actually these examples in real life before on where people can get legal information if they're in Lebanon and want to find out what's the best way to register their child. And building a bot around that so they can understand that and know. Know what rules there are. But there can be misinformation out there as well around these types of things. And so what are you seeing in your sector, even within the journalism side of things? And how is this panning out for you guys?
- [23:57]
D
Yeah, I mean, it's very interesting, right, because we think of AI as all powerful, but actually it's not. It's a word completion engine trained on. We think of generative AI, mostly these chatbots, which is what we're talking about. There are lots of other AI use cases, but let's focus on these instance that what would build that information bot for a refugee or someone in Lebanon. What you have to do is you have to know what These things are good at what these things are bad at and what they're limited at. And they absolutely hallucinate and they sometimes give you false information. And if they have been trained on false data, then they will lead to all kinds of problems. And so the trick would be having people from the humanitarian world who are well versed and can, for example, make sure that there is a. Again, let's go to a bot to help refugees in Lebanon, that it is actually trained and that it accesses the proper information sources, that it is set up in such a way that there's a second bot that checks for common hallucinations or errors, that someone really trained in the system, sets it at its core, and that it has some kind of self corrective mechanism. That's a fairly complicated answer. But the point is we need to have people who love AI, appreciate AI and understand AI building these tools. And then it can be immensely powerful because then suddenly you can have thousands of people getting vastly more accurate information than they would during a refugee crisis.
- [25:27]
B
Absolutely. Hovi.
- [25:29]
F
I'm thinking of other examples of how you poorly design and roll out the solution. So if we park AI for just a second. I remember when I was managing Zaatari camp in Jordan, we had a company who heard about the refugee Syrian refugee crisis and they brought machines to 3D print prosthetics and they wanted to set up shop in the camp and print prosthetics. And having done zero homework on the simple fact that prosthetics is not about just the equipment, but rather it's a whole health protocol. And you have to work with the line ministry, you have to work with the population, you have to have a holistic approach to it. And we said, sorry, but like first you go talk to the Ministry of Health, you know. So I think it's also about having the patience and the time and the resources to actually run sound, well designed projects where you define a challenge and go through the process of identifying the best solution. And Nick, to your point, because we have these funds where we're receiving all these ideas, ultimately we select only 2% of the ideas that come to us. And many of them are AI based. We spend between the 3,000 ideas we get once a year in April, and the 2% we select, we spend eight months studying these until they reach a final phase where they're selected. So we only select 2% and then they are given the resources and the time average 18 months, but it spans around three years to be able to actually go through a process of designing and Implementing a project. So I think also in our sector, a lot of times there is this lack of patience that oh, in 24 hours we're going to just take a technology and roll it out. And I think that also is something that we need to also infuse the idea that you want to innovate, you want to do something new. You have to define your challenge, go through a process, dedicate time and energy to make the right decisions. And whether it's AI or not becomes almost secondary. Because if it's an AI project, you would need the expert skills within the project. If you don't have it, you hire it. And that's a bit the methodology we're using. You know, focus, create a space, create a sandbox, put the resources in place, take the time to make it work properly. That I think is a way to also mitigate these risks by bringing the experts. A lot of times we don't have the expertise. So knowing that you don't know and bringing in experts also helps kind of not make amateur mistakes.
- [28:09]
E
Yeah, it's clearly not just an all or nothing for the sector. I think it's important to start small. There's a lot of low hanging fruit, a lot of AI capabilities that can help many different parts of what you do as part of your mission, as part of your operation. And so building some familiarity with the AI capabilities, building some trust in the technology, setting up something initially within a smaller team within the organization, and then getting feedback, iterate on that process before you sort of fully expand and even consider something larger scale. So I really think getting comfortable with the technology, understanding its limitations, and make sure that you are able to design something that's both of value, something that's useful and usable, and co designing to Nassim's point, make sure that you bring in the actual context of either the people who will be using it, or the people who will be impacted by the output from AI.
- [29:09]
B
So the one point, Mike, you brought up cost of error. And there's also another cost. And Nassim and I have talked about this before and I would love to hear from everybody what you've heard or what your thoughts are. But it's really around this idea that we've talked about creating efficiencies, we've talked about the human to human contact. But there are also these other intrinsic parts of humanitarian work that we really sometimes don't dive into, which is if we do create enormous efficiencies, what happens to the labor of refugees that are meaningfully employed in a refugee Camp because they're there for a long period of time, you know, eight years, 10 years, 15 years. Now those jobs that they would traditionally do face to face are kind of going away. Some not. All right, but there's a reduction, let's say so to say there's potential for those things. So there are these other knock on effects. It's not just the cost of error, it's the cost of efficiency. And what are your guys thoughts on that? Because I have my own personal view, but I would love to hear what you guys think.
- [30:06]
D
I think it's still quite undetermined. It depends on every situation. But in many cases becoming more efficient actually leads to increased demand. Because if somebody can do a task more quickly, well then you give them more tasks. Sometimes if there's a cap on the number of tasks that need to be completed, then it leads to fewer jobs. But the example I give here at the Atlantic, where I work in media is people talk about translation. Once AI can translate well, then you won't need to employ any translators. Well, the truth is that translation is so expensive that up until last year we did no translation. But now AI has made it so much quicker that we actually pay translators to look in the final stages after AI has done the first bit. It's AI creating additional bits of work. And you can imagine in the example of the refugee camp, you can imagine situations where AI can do the paperwork so much more quickly that you need fewer people to do the paperwork and so then they don't have jobs, which is, which is quite bad. Or you can imagine situations where AI for example, helps the translation and language and helps communication so much that it actually creates more work. So I would just challenge the assumption that an increase in efficiency necessarily leads to a decrease in human work. Because I think that's still a really interesting unanswered question and will be very context and situation dependent.
- [31:35]
F
Actually I agree with Nicholas. You know, before we opened, let's say third party messaging apps to have a dialogue with refugees and distressed populations online, refugees had to line up in front of our office in the gap with long waiting hours and an X number of people can get in every day. And so people wait and then they go back and come back the next day, etc. Fast forward. We launched a line that was the first corporate test of use of third party messaging app in Ecuador. We went from obviously zero users to 20,000 users in two months. Now some parts of the house are saying, well, you know, if we have chatbots, we actually would need less people Actually, quite the contrary, because the amount of people interacting us online is so much higher now. So we actually have more work to.
- [32:34]
B
Do, a lot more work.
- [32:35]
F
Tell senior managers that this is not a cost saving measure. We now have more work because more people are talking to us and more cases are coming forward to us than they would before, especially in urban settings where people are scattered. So it's very difficult for them to also travel distances to come to our office. Now. They can contact us. Well, when someone contacts you, then you have to serve them. If they don't reach out to you, then it's one less client. So I think I agree that the verdict is not out yet, but I think the second thought I have to your point, Chris, if communities are involved in the design of the solutions, then you have a different game. I think the challenge and the risk is if we adopt technologies from the intention or the end goal being let's cut costs. If our end goal is to cut costs, then the outcome obviously will be very different than if our end game is to actually improve services and better serve the people we are paid to serve. If our end goal is to better serve, then you design it with them. I think the outcome would be vastly different than, oh, we need to cut a bunch of staff, so let's just roll out the software. So what's our intention? If intention is cost saving, then probably the answer is, yeah, we're going to cut corners and lay off a few people. If the intention is let's find solutions, we actually risk having more work to do, not less.
- [34:08]
E
Yeah, for sure.
- [34:09]
F
Over to you, Mike.
- [34:10]
E
I agree. AI for a while now, in particular the past two years since generative AI went mainstream, AI is definitely having a very direct impact on the job market at large. Some jobs will disappear, others will see the light of day. But it might be a case of that some tasks will be automated, many jobs will change. So you need to have that sort of continuous upskilling and learning new capabilities with the tools you have access to. I don't think it's necessarily. It's not given that it's a net loss in terms of number of jobs, but I do think many, many more jobs will be impacted in one way or another. It's in terms of new things you need to learn. But I think for the humanitarian sector specifically, the sector's being asked to do more with less. Last year, 2023 was the year with the biggest gap between funds needed for the humanitarian challenges needing to be addressed and funds actually provided. And that gap is expected to Increase. So given that AI obviously offers at least an opportunity as part of a solution to address the need to having to do more with less. But of course, there are smart ways of doing it, and there are sort of somewhat more irresponsible ways of just diving in without thinking too much about what the consequences may be. But I agree it's not a given that it's a reduction in workforce needed.
- [35:34]
B
Yeah.
- [35:35]
F
Just to point out the Nicholas point on languages, we used to have our fund. The four funds we had, we used to have just one fund, and it was only in English a few years ago with high costs on translation. Now, the funds that we use to fund UNHCR teams are in four languages. English, French, Spanish and Arabic. And the Refugee Innovation Fund is over 20 languages. And AI is helping us to actually communicate with a much broader audience through fast translation, which means we used to receive on average 100 applications a year. Now we receive 3700 applications a year. And the vast majority are from refugee organizations, and the absolute majority are not in English. I would argue that people would not have applied if we only had it in English. Now, I happen to lead my first application in Armenian, so when we advertised our fund, it was done in Armenia, so people applied in Armenian. But then there's a human eventually who needs to read the application. And so now we have way more work to do because we have way too many applications. Not way too many. We have many applications because we were able to use AI and open up and localize more. So to, again, to the point that actually we have more work at our hand because of our use of technologies for good.
- [37:01]
E
Adding to that, I think the localized aspect, the local relevance to the discussion here is really critical as we receive a lot of excitement around modern foundation models. Generative AI over the past couple of years, in particular, improved productivity. A lot of buzz around this. It's important to keep in mind that this works very well in English and a few more languages. Once you get, let's past say, 10 languages or so, the quality, the accuracy, the value of modern AI really starts to drop. And that means that the large majority of the world population do not see any benefit at all from this. So obviously that has the risk of further deepening inequities across the world. And it does really require explicit focus on countering that. And so there's a lot of work needed, there's a lot of work happening to make generative AI accessible to more users, but we're definitely not there yet.
- [37:59]
C
Thank you. We talked about two, three different costs so far. Right. One was the costs and errors of AI just deploying it. One was the costs of job displacement. And I wanted to also think about this other cost, which is quite financial and resource intensive. Technical resources around evaluation and the deployment of AI in the context of humanitarian work and also building the systems that Hovigen has been mentioning, the capacities we require to actually deploy AI. So in that regard, I have three, I guess, final questions for each of our guests. For Nick, I'm thinking, what do you think of the costs regarding evaluation of AI? Do you think it's going to increase, decrease, or who is responsible for this? Right. We now are presented by corporate companies these AI tools, but when it comes to deployment, now we have to spend these resources on actually the adoption. Right. So that's the first question for Nick, for Hovik, I'm thinking, do you think in the humanitarian space, do we have the capacity to do this? We talked about AI experimentation and I think that's a very valuable skill to have. But I think we don't have the capacity to do it oftentimes. Right. We want something that is already shown proof and value and we can deploy it, especially with humanitarian work. Right. In response and preparedness. So how much resources do we have to spend on experimentation and evaluation? And for Michael, I was wondering, what are your thoughts on private sector partnerships with humanitarian organizations? How can we create a space where we can actually collaborate more effectively? We have these very bureaucratic systems in the UN partnerships, vendor relationships, generally. Where do you think we can build the capacities where we can benefit from private sector and private sector can benefit from us? I think sometimes we forget that second component. So maybe we can go back to Nick and over to you.
- [40:07]
D
These are all great questions. AI is very expensive. It's very expensive to build the models. It is expensive to use the models. There are external costs as well. The environmental costs of AI are massive which can also contribute to humanitarian crises. What I am hopeful for is that the open source and free and on device large language models, or sometimes they're called small language models, will become essentially as efficient. There will be a convergence. There are a number of free to use open source models built by Hugging Face, built by Llama. And I would imagine that the future for humanitarian organizations that don't have the same budgets will be to these models that are right now not quite as good as the ones you can get accessing the cloud, getting to Anthropic or getting to GPT4, but will be very close. Some of them will also be locally tuned and built and use local inference in the communities they are. So I think the growth of open source models is really good for the humanitarian world and I would expect that's where a lot of the use will come and ideally the cost will be far lower than we think.
- [41:22]
F
Right now maybe it's my Internet, I'm gonna cheat and I'm gonna try to answer also the third question because it's linked, do we have the. I think ultimately first is the financial resources to do this to the next point, it's expensive. What we're trying to do is to actually get those financial resources place through the owners to be able to invest. So I think that's one specific element, the funding. We need the funding to be able to do this. The second thing, and I want to give an example, we're partnering up with arm, which is a chip design company. And not only do they fund us, but they also give us time. They're experts. But towards the challenges that we have identified. As opposed to a lot of companies who declare willingness to support only to find ourselves, that they see us as another client and they want to sell us their products. I think there is a space for us to partner up with private sector companies who are willing to work with us on the challenges that we have identified and work together solution. And I think pro bono expertise that is not just selling us their products could help us a lot. And I think I alluded to this before, being humble enough to say we don't know and to seek those who can help us find answers. I think it's key. So what we do a lot of times actually is we to our innovation projects, we're actually hiring experts and paying for their time to be able to do the right thing as opposed to pretending that we know. So I think that our ability to be efficient in deploying solutions also depends on our humility to admit that we don't know and reaching out and partnering up, ideally pro bono, but a lot of times we'd have to pay for it. The expertise, also the hardware and the software that comes with it to be able to make a dent in the system.
- [43:16]
C
Good points. And Michael, over to you. I'm sure you have some thoughts hearing both.
- [43:23]
E
Yeah, great discussion. And so definitely from my work from within Microsoft philanthropies, every single interaction I've had with humanitarian organizations, I will always learn something. Not just about their context, their mission, what they're working on, but also about the technology that we create. Learning about its limitations, coming across unexpected use cases or ways of thinking about the technology that we can then bring back to the drawing board, bring back to the product teams to help them build more robust capabilities. So in terms of partnership with the private sector and humanitarian organizations, there's a lot to benefit from on both sides. When you lean in on the respective areas of expertise, I think there's a lot of value in. We talk about cost as well, the intuitive interface of what we've seen with generative AI. I'm a linguist at heart. My entry point into AI was through linguistics. And I really enjoy the fact that now, today, language as an interface is sort of the way a lot of interaction, the use of technology happens. And I think that means that capacity building is absolutely required, but it also means that the ability to use very advanced AI, the barrier to be able to do that, has significantly lowered. And that means that we get a significantly broader range of users or broader demographic that is exposed to AI, that uses AI both in their personal lives and professional lives. And that means that we have more people participating in the discussion at a societal level around what we want from AI.
- [45:03]
B
So we're going to start to wrap it up a little bit here, and I wanted to bring in kind of two small aspects. The first one is around strategy, right? We wanted to talk about what strategy means and what strategy looks like. And what I'm taking from the conversation were kind of three big pieces. First is the risks. Organizations, as they're developing their strategy, they need to understand the risks. Now, I know that that sounds a bit trite maybe, but. But I mean, the reality is, is that I don't know if anybody has any clue what the risks are, right? And there are just so many, and it's so outside of their box that this is really going to be a big journey on that risk. Because the cost of error, all these different costs that we talked about, we haven't even gotten into data privacy, data protection, we haven't gotten into cybersecurity, all those different pieces of this puzzle. The second one is around design principles. So just pulling from what everybody's saying is they've got to kind of come up with a way of creating this opportunity for the design principles to be fit for purpose, right. And understand that there might be different tiers of the design principles on engagement. You know, there's the long tail, as you said, Hovig, three years, but there could be the short tail, which is, hey, this tool already is there. We're just going to create a number of different agents, parent child agents that are able to Go through a process and it's quite easily defined and quite easily used. The last one is not tech for tech's sake.
- [46:21]
D
Right.
- [46:21]
B
We hear this all the time in the sector and it's when even we are creating strategies, when the powers that be start to create these strategies, are they being pragmatic about their use?
- [46:32]
F
Right.
- [46:32]
B
Not just designing the strategy. Because now we must all use AI, which tends to direct things on the technology side, but actually, how can it be applied? So I just think it's been a really impactful and great conversation for all of you to spend this time. And I wanted to go and Nassim, I'm including you in this one. I had four quick questions and it's just a one sentence answer for everybody and each question is different, so they're super easy. But Nassim, for you, I'm going to start with you. What is your one wish for AI in the humanitarian sector?
- [47:06]
C
I wish for AI to actually equalize the knowledge space. That's what I want.
- [47:10]
B
Democratization. I love it. Nick, what's one quick win in the humanitarian space that we can. Given all the the knowledge that you glean from all the different people you talk to and all the different work that you're doing, what's one quick win for us in the sector?
- [47:24]
D
Translation between two people who speak rare languages. Once you can get AI to translate not from a rare language into English, but from a rare language into another rare language, that will be beautiful.
- [47:39]
E
I love it.
- [47:40]
B
All right, Hovig, question for you. What's one frustration in the sector around AI and the use of AI?
- [47:49]
F
My frustration with our sector is the fear of missing out, driving decision making, not having the right reasons why we use technologies, but just the resemblance of being cutting edge as opposed to saying, we have major problems to solve. Let's focus on them.
- [48:08]
E
Yeah, I love it.
- [48:10]
B
And finally, Michael, for you, who is one person we should talk to next.
- [48:14]
E
You should talk to people within your organization about how you think about AI and get started. I think it really is important to get started, to explore the opportunities, but to make sure that you are aware what the technology is, not what it does, but take the first step. Yeah, I love it.
- [48:32]
B
Well, this is going to be part of a two part series, to be fair, because it was tough to bring in five people onto one call and actually get them all at the same time in the 16 different times that we all operate on. So we're going to have a call with Yuri and Lar and Ricardo Rencon from EFAD coming up in the coming weeks as well and discuss with them. And I'm really excited. There's so much more to talk about here. There's so many questions, more questions than I had. Nassim. Naseem did so well at putting together all these amazing questions, and I think we got to like, three of them.
- [49:03]
D
Naseem, no, your questions are fantastic. And your summary was great.
- [49:08]
E
Absolutely.
- [49:09]
C
It was such a pleasure talking to you.
- [49:11]
B
Yeah. Well, guys, listen, like I said, I can't thank you all enough. I am a bit in awe of everything that all of you have done. Hovig alluded to it. Hovik and I met in the desert with about 37,000 refugees coming over the Libyan border right before Gaddafi was killed. And we were like, okay, dude, what do we do next?
- [49:29]
E
That's where our relationship started.
- [49:32]
B
And Nassim has been so kind to. To join on this journey. She was like, who is this guy reaching out to me on LinkedIn that wants to do a podcast on AI in the humanitarian sector? So super thankful that you're hosting this with me. And part of this. And Michael, I don't know if you remember, we were on a panel.
- [49:47]
E
I do.
- [49:48]
B
A couple months ago. Yeah. At hnpw, which was great. And I really liked a lot of the things that you're saying. And I really want to hope to bring you on the ethics call because I know that you wrote a lot of the ideas around ethics and AI and the humanitarian sector. So you might be a second, a double up for us, maybe when we get into that.
- [50:05]
E
Hold me in whenever. Absolutely.
- [50:07]
B
Yep. And Nick, my own personal thing, you inspire me with everything that you do, especially your running and your writing. So it's been a great pleasure to have you join with us.
- [50:17]
D
It's so much fun. I love talking with all of you. It's just great.
- [50:20]
F
That's awesome.
- [50:20]
D
I learned a lot from this conversation. Thank you.
- [50:22]
B
Well, thanks, everybody and big hugs. I'm going to stay on here with Nasim. I've got to do some of the other reading on thanking everybody and everything so that can go down the recording. But I'll leave of you guys to your day. We. We did an hour and I really appreciate all of your time, so thank you all so much.
- [50:37]
F
Thank you.
- [50:37]
E
Thanks, everyone. Have a good rest of your day.
- [50:39]
C
Have a good day.
- [50:41]
B
We did it, Naseem. We did it. We did our first episode. Congratulations.
- [50:46]
C
Congrats to you as well. What a great conversation that was. I personally learned a lot and I can resonate with all of the perspectives that were shared Yeah, I think it.
- [50:57]
B
Was a really pragmatic group that was able to look at both sides of the story. Nobody was evangelizing either way. Right. They were really trying to look at both sides because it's so complex. The humanitarian sector is complex. The situations are complex. The people that we're engaging with come from varied cultures and varied languages, varied backgrounds. And so it's important to not be on one side or the other, but to kind of be able to understand both perspectives. So I thought it was really good.
- [51:23]
E
Good.
- [51:24]
B
Yeah.
- [51:24]
C
I think. I love the fact that what Hovik mentioned around the FOMO in the humanitarian space around AI, we either try to just go for it super quick and just do it as fast as we can, catch up with the trends, but at the same time, we have people who have completely been in an idle state because they're trying to see how can they control the risks of AI. And I think having that middle ground is really critical in this space. I'm not being so afraid of it because we also talked about how resource intensive it is to even start doing AI. So perhaps we should just start thinking about it and strategize our way forward.
- [52:03]
E
Absolutely.
- [52:04]
B
And that's why starting off, I mean, I think both you and I agree that starting off with the strategic discussion was going to be really important around what kind of strategies organizations need to start to look at adopting or writing to get there. But I'm done. I think I need a glass of water. Let myself calm down after that great conversation. I can't wait to hear it back and. And I can't wait to do the next episode with you. Coming up in just another few weeks here, we'll be doing our second episode and can't wait to see you there.
- [52:30]
C
Yes, absolutely. Talk to you soon.
- [52:32]
B
See ya.
- [52:34]
A
Thank you for joining us on humanitarian frontiers in AI. We hope today's conversation gave you new insights into how AI is transforming humanitarian efforts and the steps we need to take to ensure it's done ethically and effectively. If you enjoyed this episode, be sure to subscribe and stay tuned for more discussions with leaders and innovators at the intersection of technology and humanitarian work. Together, we're exploring how AI can bring real change to communities in need. Keep pushing the frontiers of possibility.