Artificial intelligence and anticipatory action: a conversation – part 2
In a recent workshop organized by the Risk Knowledge Action Network, the ELLIS Unit Jena, the German Red Cross and the Malagasy Red Cross Society, members of the anticipatory action community met with artificial intelligence (AI) and climate researchers to discuss the use of AI in climate risk mitigation and anticipatory action. Two of the workshop’s facilitators, Karen Dall and Vitus Benson, continued the conversation afterwards. This is the second part of their discussion, which explores some of opportunities for the anticipatory action sector to benefit from AI, as well as the potential risks. You can read the first part of their conversation here.
"Using satellite imagery as an input, AI can support more efficient mapping to identify high-risk areas – but we can go an even more granular level."
Vitus: We’ve talked about the current state of AI and some of its uses in forecasting. Which examples did you find particularly interesting during the workshop?
Karen: The big question of how AI can help us in risk mapping, as well as risk and impact analysis. What we need in humanitarian action, and anticipatory action in particular, is an understanding of where vulnerable people live. Why are they particularly at risk? Which communities and households are most at risk? For this, we need to understand a variety of risk factors, including people’s coping capacity, which is also dependent on the hazard.
For example, for droughts you might need to know where the nearest reservoirs or water sources are, or which areas people use as agricultural lands. For cyclones and floods, you might need to know which houses can withstand a storm and which are on higher ground, as well as where the closest evacuation sites are.
Using satellite imagery as an input, AI can support more efficient mapping to identify high-risk areas, but we can go an even more granular level. One example we saw in the workshop went to the household level, with Microsoft using AI to detect the roof types of settlements. This can help in identifying those that are most vulnerable to flooding and cyclones.
Of course, you need ground truthing for this; AI cannot do the job alone, it needs people on the ground that verify and help train the model. I think this is one of the challenges when we talk about bringing AI together with humanitarian action.
AI for humanitarian work: a risk or an opportunity?
Karen: Where do you see the biggest challenges for AI to be trained, and then used, for anticipatory action? How do we get from the theory of AI to making it actionable for humanitarians?
Vitus: One thing I learned during the workshop is that there is still a long way to go – and we really need to work together. One of the big issues seems to be a large gap in language between those in the humanitarian and AI spaces, and in understanding what the final output of an AI system would look like and how to evaluate it. We machine learners really like quantitative metrics, for example looking at many, many examples, or computing an average performance score. But I have the feeling that in the humanitarian context, to exaggerate a little, what really matters is that you have this case study, or this example, where it really works and where you try to figure everything out, including any issues your system might have. I guess this can only work through collaboration.
What do you think? Do you see AI as a risk or an opportunity for your work?
Karen: It’s a bit of both. It has lots of potential. There are a lot of information gaps in humanitarian action – we work in many areas that are highly data-scarce – and if AI can help fill some of those gaps, that would be amazing.
On the other hand, this data scarcity is a challenge in itself. As you said, AI needs a lot of data to be trained. So, which data do we have now, and how do we get the kind of data that we need to train the models? Maybe we need to use existing data better. For example, there are a lot of datasets that are not digitized or openly available, but just data sitting somewhere, in some office, and we do not have access to it at the moment. Can AI help to digitize those datasets and make them usable for humanitarian practitioners?
This leads me to another point: the importance of involving local actors, meaning the people who are implementing the actions. These days, everybody talks about AI, everybody wants to use it in the humanitarian sector, but it is local organizations, like the National Red Cross or Red Crescent Societies, and especially their volunteers and staff on the ground, that are going to respond to a disaster. If AI is to support humanitarian responses, it needs to work for the people who are responding, and for those who are benefitting from the response.
Another issue is accountability and transparency. For example, if you use AI to identify target communities for anticipatory action, it needs to be explained why the AI model selected one community over another. A decision based on AI needs to be justified, or at least be sufficiently transparent to be understood by different actors. Explaining the ‘black box’ – how AI creates content or comes to its conclusions – is one of the big tasks.
It reminds me of how anticipatory action started around 10 years ago. Initially, the reactions were: how can we base humanitarian action, and more importantly financing, on a forecast? It took a few years to bring forecasts into humanitarian action on this very practical, operational level. And only now, or in recent years, do donors and the different stakeholders globally trust the system. This task – of building trust in AI-generated outputs – is something that needs to be done jointly by scientists and humanitarians.
"I hope that AI can revolutionize the field in a similar way to the internet revolution: to democratize access to information and particularly to early warning information."
Forecasting a future for AI and anticipatory action
Vitus: What is your dream for the next 15 to 20 years? What will AI being doing in anticipatory action?
Karen: I hope it will fill some of the data gaps we face today, which make our work very challenging. I also hope that AI is used, designed and adapted in a way that it is not just sitting with us in the Global North, but being shaped globally and in a participatory manner. Questions of accessibility and literacy remain, which I would like to see addressed.
In 20 years, I hope all the questions around data protection in AI, and also around its energy consumption, are resolved. We will certainly have much better forecasts, and with longer lead times, thanks to the ideas that you have talked about. These ‘new’ forecasts will enable us to do anticipatory action even earlier and in a more targeted way; improved risk information will also help us plan anticipatory actions better.
However, it is not just about improving forecasts: we also need to bring early warnings, and the actions triggered by them, to the most remote communities in good time – the so-called ’last mile’. AI can help with this, but it cannot replace the efforts of local stakeholders. So, let’s harvest the potential of AI without exaggerating it. Is that a reasonable ambition?
Vitus: I very much agree with everything you said. I hope that AI can revolutionize the field in a similar way to the internet revolution: to democratize access to information and particularly to early warning information. I can envision having AI as another ‘participant’ in your crisis meetings in the situation room. This means that humanitarians on the ground, wherever they are, can get an alert for an approaching hazard from an AI system and then interact with it, sharing their own opinions, experience, local contexts and indigenous knowledge. Ultimately, this will lead to better, more informed decisions – decisions that enable us to save more lives and greatly dampen the impact that climate risks have on livelihoods around the globe.
But, let’s face it, we are still far away from that scenario! I hope we can continue this discussion to find a way to get a bit closer.