AI for Clean Air - Scaling Smarter Solutions

[Genevieve Connors]
Hello everyone, and welcome to this World Bank Group Live event. My name is Genevieve Connors, and I'm the Manager for Environment for Policy & Regulations here in Washington, D.C., in the World Bank Group’s Planet Vice Presidency. And I'm really glad you've all been able to join us today for this conversation. Let's talk about air pollution. Air pollution is still the world's single largest environmental health risk. Around 5.7 million people die from it every single year. It costs economies close to 5% of global GDP. And in South Asia alone, nearly 1 billion people are exposed to air that exceeds safe limits every single day.

But you know this. These are not new numbers. The frustrating truth is that despite decades of effort, progress has been far too slow and far too uneven. The communities bearing the heaviest burden are often the ones with the least capacity to respond. While the challenge is not new, one lesson from decades of World Bank Group engagement is that lasting progress depends on strong evidence, local leadership, and the ability to translate data into action. Data into action. This is where the landscape is changing rapidly. Artificial intelligence is opening new possibilities for air quality management, from so-called hyperlocal pollution mapping to real-time forecasting for source identification and decision support tools.

We're seeing tools, new tools, that can map pollution at the neighborhood level, that can forecast air quality in near real-time and help governments and people prioritize where to act first. This is all powered by machine learning, satellite imagery, and low-cost sensor networks. And this is not just an environmental story. Beyond cleaner air, AI-powered solutions are creating economic opportunities. New businesses, technologies, and services are generating jobs across the public and private sectors while creating demand for skills that can help equip the next generation of workers for a rapidly evolving economy.

For young people across the region of South Asia in particular, managing air pollution represents a genuine economic opportunity, not just to tackle working on the problem, but to build careers and businesses around solving it. This is one of the first events the World Bank Group is organizing specifically on the intersection of AI and air quality, and I have to say, we're genuinely excited about it. The conversations we've had putting this event together have reinforced how much is happening in this space right now and how much potential there is to move faster and smarter if and when we connect with the right people. And today is a step in that direction.

So, let's kick things off with two lightning talks. These will be short, focused, packed with substance. These will be from practitioners who are working at the cutting edge of AI-driven solutions for air quality, and I want to give them the floor to show us what is actually possible. Our first speaker is Professor Mike Bergin, Sternberg Family Professor, of the Department of Civil & Environmental Engineering at Duke University. Dr. Bergin's research focuses on the influence of air pollution on human health, climate, and renewable energy production, with a particular focus on fine particulate matter. Mike, the floor is yours.

[Mike Bergin]
I'm Mike Bergin. I spent the last 20 years of my career doing international projects, trying to improve air quality throughout the world. And I'm here today to talk about AI. One thing I'd like to say about AI is AI holds great promise in improving environmental and human health. And this is particularly true for air pollution. Air pollution, in this case, fine particulate matter, small particles that float in the atmosphere that get in our lungs, can kill millions of people a year. Recently, there's been an explosion of low-cost sensors to measure air quality. Again, the importance of measuring air quality is to determine the sources and levels that affect human health. The sensors are available and they cost hundreds of dollars.

One thing I should note about these sensors is prior to this, and what was going on mostly is using very expensive monitors that cost hundreds of thousands of dollars, but now these low-cost sensors can be used to supplement these more expensive reference-grade monitors. But it's crucial that we use AI with these sensors to interpret the massive amount of data that's being generated. In the next slide, you can see, for example, in Delhi, this is a sensor network with these low-cost sensors. They are co-mingled with these very expensive reference monitors and with the help of AI to interpret those results, we can now calibrate those sensors and get very accurate sensor networks, and saving a great deal of money, time, and effort, but AI is the key component.

If you look at the next slide, one thing I should note is we are also increasingly having satellite images and satellite information available. And now we can see these images. This is over Lucknow, India. And you can take a low-cost sensor network with these images, train those low-cost sensors to the images, and use AI to figure out what the concentrations are at very high resolution based on this chaining throughout a city. The critical part of this is that AI can now help us to understand not only the levels of pollutants, but the sources of pollutants in these cities. Crucial information needed when policymakers try to decide and figure out how to mitigate and lower the air pollution concentrations and lower the emissions from those key sources.

In summary, I would also say AI holds great promise for air quality modeling, for air quality forecasting, and to help us better understand how to take all of these information available to find where these sources and high concentrations of air pollutants are and make the air cleaner and better to improve human health.

[Genevieve Connors]
Thank you, Dr. Bergin. We're now going to move to our second lightning talk. We'll hear from Dr. Imran Hamid Sheikh. Dr. Imran is the Director General Environmental Protection Agency in Punjab, Pakistan. He should know— Punjab is a province of 130 million people and is one of South Asia's most pollution-stressed regions. Imran, the floor is yours.

[Imran Hamid Sheikh]
Thank you very much, and I would like to tell that the government of Punjab is taking an extreme step for the improvement of the air quality. And the first and foremost thing is basically the introduction of artificial intelligence in this whole regime of the air quality monitoring for better environment. And this artificial intelligence is being used at 5 different steps. The first of all is basically the data gathering through CCTV cameras. The second one is basically the air quality index forecasting system. The third one is the actual air quality monitoring when it goes above the Punjab Environmental Quality Standards. The fourth one is the inspection regimes which are corresponding to the air quality levels. And as the air quality worsens, then the enforcement regime triggers very high, and likewise it goes up and down. And the fifth and the last one is basically the artificial intelligence used in the utilization of the environmental impact assessment reports. Now the thing is that for the first, as an example, I would like to say that in Punjab we have more than 10,000 chimneys where we have installed CCTV cameras, and the live feed from those CCTV cameras basically pass through a video management system which is AI-based. And in that system, up to 40% opacity is admissible under Punjab Environmental Quality Standards. So, any smoke being emitted from any chimney out of these 10,000 chimneys which crosses the opacity of 40%, the artificial intelligence machine has been trained on it, and this machine immediately generates an alarm. This alarm goes not only to our enforcement regime but also to the management of that particular factory. So, the management is asked basically through a robocall from our helpline that your so-and-so chimney is emitting black smoke or more than the admissible opacity, so kindly take action. And if you don't take action, then our enforcement mechanism will be triggered. So, in this way, through a public-private partnership and through the government enforcement, and above all with the help of artificial intelligence, what we are doing, that we are trying to improve the Punjab's air quality as much as possible and with as much as possible. And let me tell you, this has basically decreased our effort. Previously, what we used to do, that our teams used to roam around from one area to another area in the search of these black smokes. And there was a lot of time consumption and the fuel consumption. And, you know, the effort was at times we thought that that is not up to the mark. But now the operations have become targeted. They have become AQI-based, scientific-based. And I think that the artificial intelligence is playing a great role in it. Thank you.

[Genevieve Connors]
Thank you so much, Dr. Imran. That was fascinating. These were great examples of what innovation in this space looks like in practice. And I particularly liked this very clear example in Punjab of the use of AI in all parts of the process, from monitoring the sources, the production of pollution, the forecasting the continuous monitoring, the compliance enforcement regulation. That was a really fantastic example of its full potential. Now I'm going to turn to our panelists. We have the remainder of the session, the 40 minutes we have left. I'd like to bring in our panelists for what I know will be a rich and wide-ranging discussion. We have 3 guests joining us today, each of whom brings a very different vantage point to this topic. I'm going to introduce them first, and then we will get to questions. First, I'd like to introduce Dr. Arunabha Ghosh, who's the founder and CEO of the Council on Energy, Environment and Water (CEEW), one of Asia's leading policy research institutions and one of the world's top climate think tanks. I'd like to introduce Sarah Vogel, who is Senior Vice President for Healthy Communities, Environmental Defense Fund, or what you may know as EDF, and Shakriya Pandey, who is the founder and CEO of BOTS Industries and the founder of VayuDrishti, a climate intelligence company developing affordable air quality monitoring, forecasting, and protection solutions for underserved communities across South Asia. So, a short housekeeping note before we dive into the panel discussion, I just want to remind our online audience that we have two World Bank Group experts on standby to answer your questions in the chat. So, please do ask them your questions, drop them in the chat. We would really like this to be as interactive as possible. Now let's dive in. So, let me start with you, Dr. Ghosh. You've worked across energy, climate, and development policy globally and are currently a member of India's Commission for Air Quality Management. Let me ask you, how do you see AI as a potentially transformative tool for air quality management? And what do you think are some of the biggest challenges of integrating AI-driven solutions into this AQM? Over to you.

[Arunabha Ghosh]
Genevieve, thank you so much for having me on this very exciting and important dialogue. I love the two fire starters that you had preceding this panel. As the previous speakers also said, AI is clearly emerging as a very key enabler of effective and cost-efficient air quality management. We know that in many geographies there are multiple bodies responsible for dealing with air quality within government, outside government. So, how can we use AI to bring it all together? Let me just illustrate this with a couple of examples. The first is how can AI enable the decision support systems to actually hone in on where the problem is. Air quality data is coming from multiple different sources, from high-cost monitoring stations, from low-cost sensors, from models and forecasts, from emissions inventories, surveys, etc. Now, gathering all this information on one platform is one challenge. Making sense of it and making it salient enough to act is a different problem altogether. AI-enabled decision support systems can bring air quality managers closer to the action by eliminating the need for this complex analysis. Simple natural language queries could help them get the insight from this huge variety of information that is being collected. So, for example, my organization CEEW and IIT Gandhinagar are developing systems that can help them scale. We've, for instance, developed something called Vayu Chat, which means air chat, which combines data from the Central Pollution Control Board, the air quality stations, the state-level demographics, as well related to the National Clean Air Program, all on a single platform. but where the user can ask the questions in simple plain English. For more advanced users, it will then expose the code behind the analysis. Our Clean Air team is also developing an AI-powered scenario tool that can then model the different pathways required to get to a desired air quality level. The traditional way, which is physics and chemistry-based models, are much more resource-intensive. So, you can make this more cost effective. One more example, and then I'll pass it on to the other panelists. AI-powered air quality forecasting is going to be extremely important to save lives, reduce the vulnerability that you were referring to, Genevieve, in your opening comments. So, once the models are trained, they can then be deployed on modest computational resources for day-to-day operations. Again, for instance, we at CEEW and IIT Gandhinagar are working on making weather and air quality forecasts orders of magnitude faster compared to traditional methods. The deep learning emulators that we are developing could replace traditional weather forecasting model while being 3,000 times faster than the traditional model. And then you can then detect forest fires, fugitive emissions, crop residue burning. You start pulling all of this together, you're doing it faster, you're doing it cheaper. The problem, however, is that like any technology, these are not 100% foolproof. Just consider this: if you have a tool that can detect stubble burning, and it has just a 1% false positive rate, if you then monitor 10,000 farmers or 10,000 construction sites, that error rate gives you 100 wrong answers. And then you've gone and deployed your enforcement officials in the wrong places. So, just because we're leveraging AI doesn't mean you can just leave it to operate autonomously without the human checks as well, especially on sensitive matters like enforcement and penalties and shutting down economic activities, etc. When you put in those guardrails, that's how the confidence, not just in the technology, but in the procedures associated with the technology, will increase the confidence that regulators have, but also consumers and citizens have.

[Genevieve Connors]
Thank you, Dr. Ghosh. That was fascinating. I hope we can come back to some of this later. There's recurring themes I'm sure we're going to hear throughout the day on faster, more data, data density, ease of use. But that last point you made, I thought was really fascinating, which is that the people don't go away, neither the people who are impacted by pollution nor the people, the human dimension in the way that you described. I think that's fantastic.

[Arunabha Ghosh]
But those people have to get smarter at using the tool.

[Genevieve Connors]
So, we all do. And this is the risk of AI, is that there is so much information and so much data that it's like, forgive the analogy, but like a smokescreen for covering up from action. I mean, we've known for a long time what to do about air pollution, but that hasn't— even with less data, but that doesn't necessarily mean that we've actually gone and enforced it and done it. So, thank you for that thought-provoking intervention. And we do have time for the panelists to feel free to come in. So, Sarah, let me turn to you. EDF is working to advance and implement solutions that protect people from toxic substances, including air pollution. And what cutting-edge innovations and AI-driven solutions do you see in your work? And I'd like to particularly ask you to tell us a little bit about how you think we can go from pilots to scale to system-level impact. Over to you.

[Sarah Vogel]
Yeah, great. Thank you so much for having me on such an exciting and timely conversation. By way of background, EDF, those who are not familiar, we are a science-based global NGO working to advance solutions that deliver healthier people and a planet. We're working to stabilize the climate, enable resilient ecosystems, and of course, reduce health-harming pollution. So, I head up a program that we call our Global Clean Air Initiative. We've been working with cities and states and national governments, community groups, and researchers to address the problem of air pollution. And really, for over a decade, we have been looking at how do we leverage some of these new innovations that we heard Mike speak about in technologies and air quality from, those low-cost sensors from satellite, from new data analytics to deal with all this large data sets and of course AI with the goal really of strengthening what we know about air pollution, its impacts, its sources, but also expanding usability, understanding those disparities in impact and expanding the inclusivity of access to that information. And I raise that because that's the kind of lens that we've come at AI with. So, as you noted in your opening remarks and many others, there's such an opportunity now to bridge what we sometimes refer to as the last mile delivery of data. How do you get information that's usable into the hands of people that need to make those decisions in a timely fashion? So, we have been focusing in this huge world of AI, which, you know, there's generative AI and agentic AI. We've actually focused more on some of the tools that have been around for a while, machine learning tools that can help around prediction, classification, optimization, these lower-cost model emulators where we think we can be driving more efficiency in the work to understand air pollution. And I'll point to 3 examples that we're engaged in and how we think about that issue of how do we go from pilot to systems level. We've been developing a platform called Air Insights, and it's designed to be a global open-source analysis suite that takes these big complex datasets and turns them into localized purpose-driven insights for targeted action. And one of those analytical tools within Air Insights is called Air Tracker, and it's a tool that helps you identify the source area of pollution in real time. And we heard examples of that in the second talk. We've been working to integrate it into existing dashboards for cities in different parts of the world, but we started in the United States. And the reason I bring this up is because Air Tracker relies on a weather forecasting model called the STILT model. And that model was trained in the United States, where you have a data-rich environment. So, in order to make a tool like this usable anywhere in the world, we then worked with partners to develop an emulator that now allows that Air Tracker to be used anywhere. So, you've got a less computational burdensome model. It's more efficient and allows for easy-to-collect, cheap data, basically surface temperature, wind direction data to be used. So, the idea here, right, is how do you expand accessibility? You go from piloting a simple model and taking it very quickly and making it accessible to others. And it's not a standalone. It plugs into your existing dashboards. Another area where we've been focused in on is using AI to increase productivity and, again, accessibility around health impact assessments. So again, these are models that have been developed and trained using in data-rich environments. They can be technically challenging to do, difficult. And so we've been developing a tool that has an AI assist function that would allow anyone in the world, particularly in low-resourced, lower-capacity settings, to be able to do a health impact assessment. So, you're looking at what might be the improvements in health outcomes for a given action. Or what's the current impact that you're seeing now? So, that's again, looking to scale. How do we get into the hands where you've got lower capacity, lower data space? And I'll just say one last thing, which is we're pretty excited about the ability to quickly communicate results. We've seen a lot of promise on that data visualization of information. But I'll say, and I know we're going to get into a lot of the limitations, and one of them is you cannot take people out of the decision-making. But the ability to communicate effectively is something that is so locally relevant and important to not give over to AI. And so, we've really emphasized that you just cannot offload those human decision-making and value judgments and not outsource that expertise, but use these tools to build expertise and capacity.

[Genevieve Connors]
Thank you, Sarah. Again, wow, I'd like to dig into so much of that. You raised some really interesting points and excellent examples of platforms and models that EDF has been working on and supporting, including here in the United States. I liked in the beginning your point on last mile. You know, those of us who've worked in development for decades know that for a long time the last mile literature and thinking was really about agricultural extension, and then later it became about sort of mobile access and mobile penetration, and now really thinking about last mile in terms of access to data and use of data and smart ability to use it in a smart way, it's like a whole different paradigm. So, thank you very much. And I like that last point about what we give to AI and what we don't. I think that's also goes a little bit along the lines of what Dr. Ghosh ended his remarks on as well. Okay, I'm going to turn to our next speaker, Shakriya Pandey, who is the founder of VayuDrishti, which integrates satellite data and ground sensors to improve the accuracy of air pollution forecasts in Nepal. And he's joining us from Nepal today. Shakriya, can you share how this AI-based solution is directly mitigating air pollution and what challenges you face in making advanced AI technologies affordable and accessible to communities? Over to you.

[Shakriya Pandey]
Thank you for having me for this conversation, and thank you for the question. Let me start with how VayuDrishti system actually work and then how it's creating real change on the ground. At the core, VayuDrishti is an AI-powered data fusion engine that pulls data from 3 layers. First, satellite data. We ingest multiple satellites' data, giving us the broad spatial coverage even in the areas with no ground infrastructures. Second, a network of low-cost IoT-based sensors feeding real-time air quality readings. Third, meteorological inputs— weather patterns, temperature inversion, humidity— because pollution doesn't just come from the source, it moves and accumulates based on the atmospheric condition. Each of these data sources has limitation on its own. Satellites has a coverage gap and are affected by the cloud cover. Low-cost sensors drift over time and are sensitive to humidity. So, we use machine learning calibration models to cross-validate and correct sensor readings, current reference grid stations, and satellite retrievals, significantly improving accuracy without the cost of traditional monitoring structures. On top of this, we use spatial-temporal forecasting to understand the historical pollution patterns, emission inventories, and weather data. And we can predict air quality up to 72 hours ahead of at the hyperlocal resolution. But the technology is only meaningful if it's changed behavior and outcomes. So, what has actually changed? In Kathmandu, our forecast has been used by multiple municipal authorities identifying real pollution hotspots tied to the traffic corridors and the brick kilns, information that previously doesn't exist at this resolution. Schools now receive daily advisories before the pollution peak, allowing them to move children outdoor activity, indoor, and hospitals in the high pollution zones are using our data to anticipate spikes in the respiratory admissions. And at that individual level, hundreds of thousands of people are accessing daily air quality forecast through our platform to make decision about when to commute, exercise, or keep their children home. The impact is just not about awareness, it's feeding into policy. Our data has been used in the different institutions, for planning and discussion in the Nepal itself and is informing emission reduction strategy in partner cities across South Asia. And this is how AI is mitigating pollution by not just clearing the air entirely, but by giving decision makers the precise and timely intelligence that they need to act. And the second part of the question, and honestly, the harder challenge, the making this accessible is not just about lowering the price of the sensors or monitoring. The barriers run much deeper. The first barrier is data infrastructure. So, as we mentioned earlier, many cities in the South Asian lack the continuous power supply itself and the stable internet connection itself, and the local server infrastructure needed to run AI system reliably does not exist. So, we have to design our platform to operate on the deemed constraint. So, we have to build lightweight models, offline data buffering, and the low bandwidth transmission protocols. The second barrier is the technical capacity. Even when the government want to use AI-based environmental tools, they often lack the in-house expertise to interpret, validate, or act on the outputs. Our technology capacity without building is just a dashboard nobody use. So, we invest heavily in the training local officials, environmental agencies, and the community to monitor their air quality. And the third barrier was trust. So, government and communities have been burned before by the tools that promised to just and deliver nothing. So, building credibility takes time. It means bringing transparent about model uncertainty, publishing validated results, and showing up consistently. A traditional reference station costs thousands of dollars. So, our approach brings that cost down dramatically by treating low-cost sensor not just a standalone instrument but as a node in the AI quality networks. This intelligence compensate for individual sensor limitation. So, we are especially doing with the software, what actually required the expensive hardware. Yes, thank you.

[Genevieve Connors]
Thank you very much. And I was thinking when you started, my goodness, all these municipalities in Nepal are using your data. How are they interpreting it all? So first of all, incredible that they're using it for action in the way you described, but also, your almost last point about training is equally important. Otherwise, data is data and is not usable. But thank you very much. I mean, really, I would love to hear your story one day of how and why you founded VayuDrishti. Maybe we'll have time to get to that later. I'm going to do a second round now, and in boring fashion, I'm going to stick to the same order. But if we have time for a third order, I promise to mix it up. So, Dr. Ghosh, coming back to you. I want to turn to the topic of partnerships, and I want to hear a little bit about your thoughts on what kinds of partnerships between governments, think tanks, or research institutes and startups might be most effective in accelerating innovation. Thanks.

[Arunabha Ghosh]
Well, I mean, in a way, the conversation we've had so far is already indicating that this massive potential that we've all identified is completely impossible to achieve without partnerships. So, sometimes we use the word partnerships as a nice to have, we all got to say we want to partner because, you know, who will say no, I don't want to partner. But if you actually look at it in the most practical way, it's not possible to do this without partnerships. So again, let me illustrate this with a few ideas. One is, for instance, at CEEW, we have developed an AI-powered climate risk tool. It's the first of its kind from anywhere in the developing world. We looked at 279 indicators over the past 40 years of data, projected out to another 50 years. It's called CRAVIS—Climate Resilient Analytics and Visualization Intelligence System. Now this hyperlocal data can then be layered with all sorts of sectoral data. It could be construction activities, data centers, power infrastructure, etc. Now this tool, CRAVIS AI, right now does not have the air quality information, but you can immediately see the kind of work, say, Shakriya is doing. You know, we've built it as an open layer, open data protocol platform. The kind of work that Shakriya is doing can then be brought into this platform like this. But you're then integrating not just air quality information, you're linking it with extreme heat, you're linking it to emissions, etc. Because we know all of this is intersecting and interacting. A second thing is actually the use cases. Many of us highlighted the use cases, but ultimately to develop a technology properly, you need the final users. Say, a regulator, air quality enforcer, to give you the feedback whether that technology is really fit for purpose. Otherwise, we are all tech bros kind of plugging our own models. So, it's very important that startups and research institutions like ours here get the prototypes into the hands of the government stakeholders early on and get the feedback for the fine-tuning. For instance, the Mumbai Municipal Corporation is working with IIT Kanpur on an AI-enabled decision support system to draw real-time insights from 75 different sensors, not too dissimilar to what Shakriya was describing. But it's a good example of the collaboration of the end user and the developer. Again, we are collaborating with another city, Thane Municipal Corporation, on an AI-enabled construction site monitoring system, and we are taking the feedback from the Municipal Corporation to develop this. So, this constant back and forth iteration is extremely important. And the last bit I will say is that ultimately governments do have a very important role to fund the projects that really matter. Let's leave aside the really big balance sheet tech companies that are on a sort of muscle-flexing match about who has the most compute power. Let's leave them aside for a second. For most normal human beings like us, we'll have to make some hard decisions of what are the AI applications that matter to our economies, that matter to our people, that matter to our societies. India's AI mission, for instance, has only— well, I will say a lot of allocation, a billion dollars— but compared to the capital expenditure of $500 billion, it's nothing. Prioritizing the AI mission for what is the kind of application that is needed— is it weather forecasting? is it healthcare and agriculture? Because all of this has air quality implications as well. So maybe a platform that shows the maximum intersections— So, what is the application that is the maximum implication for most issues of interest, say agriculture and healthcare and emissions and energy systems and transport. If you can build something like that, it'll be that much more appealing to a range of stakeholders. So, this is not just about technology. This is actually about intention. And what is the social contract around a technological tool? That's something worth digging our heels into.

[Genevieve Connors]
Thank you, Dr. Ghosh. If I can just ask you a quick follow-up, you mentioned two municipal examples. I think it was Thane and Mumbai. How did those come about? Where's the push and pull? Where's the demand? There's so many municipalities in India and an infinite in the world.

[Arunabha Ghosh]
Absolutely. I mean, again, Shakriya was talking about Kathmandu and other cities. We've worked in Bihar, in Uttar Pradesh, in Punjab. Again, technology does not guarantee you trust. Credibility guarantees you trust. I would hazard a guess that many of us represented on this webinar, and many perhaps who are listening, have had actually a long history of working with different kinds of subnational government, city governments, central governments, etc. And we built up that credibility that comes from driving an evidence-based approach towards dealing with a problem. Once you have built up that trust, then you say, look, we can make your life easier if we started also leveraging a newer technological tool that wasn't on the shelf a few years ago. So that's number one. The number two thing is, and I would again hazard a guess that many of us would agree with this, is the how. Which is why my first example to your earlier question was about how do you make this information salient enough to act. And because coming and telling officials who are already fiscally constrained and worse time constrained that here are the 25,000 reasons why you need to act is no good. You know, you're actually creating more frustration for them. Instead, as our colleague from Pakistan, for instance, was highlighting, the time that you save them by helping them go directly to identifying the specific sites that are polluting rather than, you know, go around in the blind trying to identify them, that then gives you a lot more the pull factor for your technological push. So, I would say these two things are necessary and should be filters. Have we got the trust with our stakeholder and what else needs to be there? And whatever it is we are offering, is it cheaper and faster and more inclusive than whatever else is on the table? If it isn't, that's fine. AI isn't the only solution. It is one in your quiver of arrows.

[Genevieve Connors]
I love the quiver of arrows. Speaking of social contract, Sarah, I'm going to turn to you because part of a social contract is, of course, thinking about prioritization and maximum benefits for society, but also serving the underserved, and being inclusive, and ensuring the poorest are not left behind. So, my question to you is, could you tell us a little bit your thoughts on how countries can ensure that the economic benefits of AI-driven environmental innovation actually reach smaller cities and underserved communities and not just large cities or major tech hubs?

[Sarah Vogel]
Yes. I just want to emphasize the point that Dr. Ghosh just made about the importance of the trust and the partnership. And it applies to how I want to answer this question. For example, the work that we do, you're not going to have actionable information if you're not actually co-creating what that information needs to be with the people that are going to use it. So, when we're developing the example I gave, Air Insights, it is deeply with partners sitting in cities doing that work or community groups doing that work. So, it's very much human-driven and it's being deliberately inclusive. And I think that's an important part. The question about the sort of distribution of benefits, I think, points to two important risks that I have been brought up and I just want to put some finer points to. First is obviously this significant risk of just an unequal adoption where the AI gets used. So, the example that I gave about a lot of tools being developed in data-rich environments, and that's between countries, within communities, and then applied to lower data environments without local capacity building. There's a big risk there of it being misapplied, deepening disparities. This plays into these issues that both the other panelists were talking about in terms of lack of trust, being burned in the past about different tools and technologies coming in and not delivering any benefit. That's huge. And then of course the second risk is just the risk of poor-quality information, and I would add expensive outputs because those who are trying to get access are being locked out of having that information. Because I think you could imagine a world in the future where you could have just dramatic proliferation of these kind of AI outputs without any standardization, you could have a situation in 10 years where we have a problem of too many products producing these inconsistent results. And then trust starts to fall apart. I think it's important to recognize if we're going to think about how do we get at smaller communities, underserved communities, is the— I think Dr. Ghosh used the word social contract. What are we doing that's different? Because if we look at the air quality management system today, whether it's in my country, in the United States, in Southeast Asia, there already are some significant inequities, right? So, the question becomes, let's be explicit of what we're trying to build. And what our goal is. If our goal is to actually deliver cleaner air with a focus also on those who are most heavily impacted by that air quality. We need to think about then what's the AI air quality management governance architecture that we need. I think to Dr. Ghosh's point, what are we going to prioritize in terms of the tools that we're using and developing, how they're being used, how we're bringing people in to influence those decision-making. So, there's a couple of areas that have been touched on that countries can think about investing in communities. One being capacity building, right? We don't want to leave smaller and underserved communities out of the ability to actually use tools, and be part of the human decision-making process. The other would be this issue of affordability and access. If there are poor agreements built into these systems or there's lack of transparency, you're not going to have the same kind of access for lower wealth or underserved communities. Another big area is reducing those important data gaps. So again, this is about building inclusivity from the beginning of the design. There's really interesting examples of this in the use of like AI judicial systems where more minority communities where their laws, systems, and languages have not been included in the training data. You have real risk of things going awry. And I think that applies here too. Having people involved in making these models and the systems better by participating in contributing that data and training the data on a more local level is really important. There's issues of consent, but I think broadly what could be really interesting in this space is to think about does there need to be a framework for responsible use of AI and air quality management? That is based on collaborative governance agreements, fair access and use, so that we're moving towards the goal, which clean air is a public good, right? It's not just a private market space. So, I think there could be some interesting potential energy in this space moving forward.

[Genevieve Connors]
Thank you so much, Sarah. And you raise an interesting point. My concern is AI waits for no man. And as we wrestle with the governance architecture, as you said, or the judicial implications, I mean, even and especially here in the United States, we are scrambling to get ahead of something that we are frankly currently behind. When I say we, I mean the public sector and the governance and regulation. But it is a very important topic. I want to turn to Shakriya now because you've started this incredible climate and pollution intelligence venture in your country, in Nepal, but this has also had an impact on local job creation and can have a huge potential when replicated in other job markets. We have an explosion of young people, millions and millions of people coming into the job market every year throughout South Asia. I think in about 10 years, South Asia will have close to 300 million people just in the age bracket of about 15 to 25. So, maybe you could tell us a little bit about your thoughts on local job creation and also what skills young professionals could develop if they want to work at the intersection of AI, climate pollution, and public health in the future. Thank you.

[Shakriya Pandey]
Let me start with when I was hiring my first team in Kathmandu. I realized very quickly that the person I needed actually didn't exist. At least not in a single job description. I needed someone who understood atmospheric science well enough to know why sensors reading was wrong, but also who could write the code to fix it. And I needed a person who could sit with the municipal office in the morning and could develop the machine learning pipeline in the afternoon. And that combination simply wasn't taught anywhere. So, we built it ourselves and the process taught us something important about the climate intelligence is actually doing to the workforce in this region. It is not just about creating the jobs, it's creating the entirely new category of the jobs. At the biodiversity, our team spans from AI engineers, environmental scientists, to IoT engineers, to different policy translators. Many of these people come from the different backgrounds that seems totally unrelated. But the climate intelligence is putting talent from unexpected places because the problem it solves are deeply local and deeply human. And that is what's happening across South Asia at this scale. As the government is beginning to mandate air quality monitoring, as cities seeking climate resilience plans and international financing flow into the green infrastructure, a demand for people like this are going to sit and ensure intersection of the data environment and the community, and it's growing faster than the universities are even producing them. And in South Asia, the significant share of jobs related to youth is going to be so environment engineering creating a role that even does not exist a decade ago. For the youth, I will say that the first skill they have to invest on is to have about the data literacy, not just about understanding about statistics or coding, but the full pipeline of understanding how data is actually collected in the messy real-world conditions and how it's cleaned, how uncertainty is communicated, and how it ultimately reaches a decision maker. In my context, how to have a deeper understanding of how does the pollution sensors read differently on a different humid day. And the second union is the domain depth. I have met many young technologists who can build a beautiful AI dashboard but have no idea about the things that they are actually doing and doesn't know what does the pollution actually does to the lungs tissue or why the temperature inversion actually trapped the pollution over the valley. So, AI without the domain knowledge produces tools that look impressive but does not give the right answer at the critical moment. So, they have to go deep in the environmental science, public health, and the atmospheric chemistry or the urban planning and bring the AI to it, not having AI go around other way. So, the third and the most important and most harder to teach is the ability to work across the institutional boundary. In this field, you will have to constantly be sitting between the satellite agencies, city government, community health workers, and the software team, and none of them speak the same language. The people who can create the most impact are the one who can translate across all these words without losing the trust of any of them. So, climate intelligence is not just a career opportunity, it's a chance to be a part of building infrastructure that will define public health in the region for the next generations.

[Genevieve Connors]
Thank you so much. And I love the story of how you started your company looking for that unicorn and then slowly built a group of next-generation employees who could cover those various skills, a sort of terms of reference for which nobody could meet the requirements in the beginning. And I think that says a lot about what the things we could do with research institutions and with universities and with secondary schools to start building the skills in a way that you just described. I did want to mention that we have a very lively chat going on. We have many, many, many hundreds of questions coming in and lots of amazing response to the speakers. So, I want to thank you. You may not see it now, but you're generating a lot of interest in the live chat. We have time for one last rapid-fire round. I'm going to ask each speaker to stick to 2 minutes. And I want to turn to something that I see, and I'm going to change the order, that I see as a bit of a paradox. Because we often talk about AI, as we are today, as a solution for air pollution challenges, climate challenges, but AI infrastructure also consumes a huge amount of energy and water at scale. So, I would just be curious on any reflections from you on how we can ensure that AI for environmental management is actually genuinely sustainable from production to end use. Sarah, I'm going to start with you.

[Sarah Vogel]
Yes, thank you. This is an issue that my organization has thought a lot about and is working on. And obviously, the explosion of data centers globally has caused a lot of consternation in communities and it's a significant political issue. There's a number of points here. First is coming back to what's the fit for purpose for AI in air quality management. And I mentioned at the beginning that we have really focused on where we can drive efficiency. And there's the conversational large language model and agentic AI that is driving a lot of this current demand in sort of those hyperscalers. And there is more efficient AI uses, the emulators. And I think we need to start differentiating out that space. That's on the AI side. The other part I'll say is that we're very much focused on making sure that rising demand in energy use is actually catalyzing a clean energy transformation, so that we're actually not using AI to control air pollution, and then putting on a lot more fossil fuel. That's really key. Being transparent about energy use, water use, and the pollution impact is going to be really essential I think, for the social contract for these data centers to continue to expand. But we can be smarter about how we're using it and really sending that signal that rising energy demand has to be tied to a clean energy source.

[Genevieve Connors]
Okay, thank you. Shakriya, I'm going to turn to you next. Two minutes, please.

[Shakriya Pandey]
Okay, so first of all, we have to actually differentiate what is the actual need and what's the objective that we actually wanted to solve. So, rather than just having a different objective and the different institution working on different sides of the air pollution mitigations we should be more specific and work collaboratively so that the technology and the resources that we are using are properly shared and properly utilized, so it's better for the humankind, and not just wasting our resources that everyone is just doing their work on the one and not collaborating more often and just destroying the resources and all.

[Genevieve Connors]
Thank you. I'm struck that from my backyard in Virginia, we heard from the Secretariat and Presidency of the UNCCD COP in Mongolia. The issue of environmental management and sustainability in data centers is everywhere, and it is a very, very hot topic. Dr. Ghosh, I'll give you the last word. Two minutes, please, and then I'll—

[Arunabha Ghosh]
I promise your viewers that this is not a setup question, but thanks for asking the question. I just published a column on this this morning called Redesign Power Markets for a Sustainable AI Future. Now, what you just mentioned, Genevieve, yes, there is a lot of conversation around the power demand that's coming, the water demand that's coming, which Sarah was mentioning for AI. But hang on, that's not the only part of the problem. We are also having power demand because of an extreme heat crisis here in Europe, in South Asia, etc. So let us not look at the AI problem in a silo. I call it the trilemma between dealing with AI resource demand, grid fragility and energy security, and dealing with climate extremes. To deal with this trilemma, a few ideas, and the rest of you can go and read my column. One, we've got to start designing power markets that reward flexibility. If you don't do that, you're not going to tap into the clean energy resources when it's most abundant, and you're not going to be able to bring in virtual power plants to drive this data center growth. Second, the mantra that we are increasingly advocating: site with foresight. Where you build a data center is more important than the choice of whether you build a data center. And having detailed hyperlocal assessments of those climate risks, of heat stress, of water stress, can help you design that data center infrastructure accordingly. Third, we need to have much more transparent disclosure, not just of emissions, but also of energy use, of water use, and link that. Just as we had, RECs for renewable energy, we should think about creating a ratcheted-up process of increasing resource efficiency of the data center infrastructure. So, the point is not an either-or. Yes, it's easy to say it's not an either-or, but the intentionality is missing in many cases, but it's not going to get solved just by the hyperscalers saying, oh yeah, we figured out our energy source. No, we are on one planet and we are over 7 billion people. You've got to respond to everyone's needs in a more inclusive way. So, let's be intentional about this.

[Genevieve Connors]
Thank you. Thank you so much. And a well-timed publication. I think everyone will now go check it out. So, thank you very much. All right, we have just a few minutes left, so I'm going to close us out. I did want to say that we have over 500 people connected. So, we will also record and share the recording of today's event. We'll be live on social media with various links and postings. Thank you very much to everyone who stayed connected and asked questions and answered questions. But before I thank our speakers, I just want to say we really did cover a lot of ground. We looked at hyperpollution, hyperlocal pollution sensors in Nepal, to community-level data advocacy. We talked about this governance architecture and the social contract, and then also some very practical questions of how we build these platforms and who uses the data and who benefits. What strikes me maybe most from this conversation is that, well, technology is moving very, very fast. As I said, AI waits for no man, but the real work is still fundamentally about people. At the end of the day, we are, yes, we are living on one planet and we are all human beings. And these people include the researchers pushing the boundaries of what data can tell us, the entrepreneurs making sophisticated tools and designing platforms that are for cities and for policymakers, the policymakers themselves willing to take chances on new approaches and reskilling often later in age and life. The young professionals building careers and reskilling themselves, perhaps for the first time, at this intersection of AI and climate and public health. And of course, the underserved, the ones who bear the highest costs of pollution. I just want to thank very much our speakers, Mike and Imran, who joined us for the lightning talks, Arunabha, Sarah, Shakriya, for the generosity of your time and the depth of what you've shared. Thanks again to everyone for joining us and engaging so actively in this chat. We hope it's a first of many more to come and the beginning of a conversation. We look forward to continuing to work with you. And everyone, have a good day, have a good evening. Thank you.

[Arunabha Ghosh]
Thank you.