How AI Drives Innovation and Economic Growth
20 Feb 2026 15:00h - 16:00h
How AI Drives Innovation and Economic Growth
Summary
The panel convened to examine how artificial intelligence can either narrow or widen development gaps in emerging economies [60-65]. Johannes Zutt highlighted AI’s capacity to boost productivity in sectors such as agriculture, health care and finance, noting that 15-16 % of South Asian jobs show strong complementarity with AI [12-20]. He also warned that AI may displace entry-level, knowledge-based jobs and that many low-income countries lack basic infrastructure-reliable electricity, broadband, and literacy-to deploy it effectively [21-31]. To address these constraints, the World Bank promotes “small AI”: affordable, locally relevant applications that operate with limited connectivity and data, citing India’s digital identity system and farmer-focused phone tools as exemplars [34-40][45-46].
Ufuk Akcigit argued that while the application layer of AI lowers entry barriers and encourages creative destruction, the foundational layer remains compute-, data- and talent-intensive, creating concentration risks that could spill over to downstream markets [86-98][99-101]. He stressed that without improving the business environment-such as reducing reliance on family size for firm growth-AI alone will not generate entrepreneurship in developing economies [111-115]. Anu Bradford emphasized the need for AI sovereignty and a rights-based regulatory approach, noting that Europe’s AI Act illustrates how regulation can protect public interests while still fostering innovation, and that India must adapt such frameworks to its own priorities [167-176][181-190].
Michael Kremer argued that targeted public-good AI, such as AI-generated weather forecasts for 38 million Indian farmers and AI-assisted traffic safety tools, can substantially reduce poverty if supported by evidence-based innovation funds and rigorous impact evaluation [133-158][263-292]. Iqbal Dhaliwal provided a concrete small-AI case in Indian schools where AI automated spelling checks, freeing teachers to focus on higher-order learning, and warned that hype must be separated from reality because technology often fails without complementary system changes [236-247][294-318].
Across the discussion, participants agreed that AI’s promise hinges on coordinated public-private effort, robust infrastructure, and policies that mitigate job displacement and market concentration [34-38][86-98][263-270]. They also concurred that neglecting governance-whether through weak regulation, insufficient public-sector adoption, or power dynamics that block scaling of successful pilots-poses a major risk to equitable outcomes [414-416][321-328]. The panel concluded that while AI can drive transformative gains in health, education and agriculture, realizing these benefits requires proactive policy, inclusive regulation, and safeguards against concentration and labor-market shocks [382-388][394-398]. Overall, the discussion underscored AI’s dual potential to accelerate development and exacerbate inequality, urging immediate action to shape its trajectory responsibly [414-416].
Keypoints
Major discussion points
– AI as a catalyst for development in emerging markets - Johannes Zutt emphasizes “small AI” that works with limited connectivity, data and skills, citing examples such as pest-identification for farmers, AI-assisted nursing, and credit scoring [10-21]. He highlights India’s digital identity and payment infrastructure as foundations for scaling such tools [39-42]. Michael Kremer adds concrete public-good cases, notably AI-driven weather forecasts that reached 38 million Indian farmers and improved planting decisions [136-151].
– Structural challenges and concentration risks - The panel notes that AI can displace entry-level, knowledge-based jobs and that the World Bank itself sees fewer such positions advertised [22-32]. Ufuk Akcigit distinguishes a high-barrier “foundational layer” (compute, data, talent) that tends toward concentration, warning that this may spill over to the application layer [85-100]. Anu Bradford and later speakers point to the global concentration of large-model development in the US and China, and to rising market concentration that could lock-in incumbents [185-190][342-347].
– Policy, regulatory and governance imperatives - Both speakers stress the need for AI sovereignty and rights-based regulation. Anu Bradford describes the EU’s AI Act as a “rights-driven” approach and suggests India can adapt such lessons while preserving local priorities [166-176]. Michael Kremer argues that governments and multilateral development banks must fill gaps where private profit motives fall short, e.g., funding AI for public goods like digital IDs and weather forecasts [128-133]. He also proposes evidence-based innovation funds to accelerate responsible deployment [266-277].
– Evidence-based implementation and evaluation - Iqbal Dhaliwal shares a school-AI pilot that freed teachers from routine grading, allowing them to focus on higher-order learning, illustrating the demand-driven benefits of “small AI” [236-247]. Michael Kremer outlines a four-stage evaluation framework-model performance, user impact, scalability, and continuous improvement-to ensure AI interventions deliver real outcomes [284-291]. The discussion repeatedly stresses the need to adapt institutional systems (e.g., teacher training, regulatory processes) to realize technology’s promise [308-315].
Overall purpose / goal of the discussion
The panel was convened to examine whether artificial intelligence will narrow or widen development gaps and to identify practical policy levers for emerging economies. Participants shared concrete use-cases, highlighted systemic risks, and debated how multilateral institutions, national governments, and the private sector can jointly shape an AI ecosystem that delivers inclusive growth.
Tone of the discussion
– Opening (0-10 min): Optimistic and celebratory, emphasizing AI’s transformative potential.
– Middle (10-30 min): Becomes more measured as speakers acknowledge infrastructure deficits, job displacement, and concentration of power, shifting to a cautious, problem-solving tone.
– Later (30-48 min): Pragmatic and solution-focused, with concrete examples, policy recommendations, and calls for evidence-based pilots.
– Closing (48-51 min): Balanced rapid-fire reflections, acknowledging both “big wins” (health, education) and “big risks” (market concentration, regulatory lag), ending on a sober yet hopeful note.
Overall, the conversation moves from enthusiastic optimism to critical realism, ending with a constructive, forward-looking outlook.
Speakers
– Jeanette Rodrigues – Moderator/host of the panel discussion [S1]
– Michael Kremer – Economist, Nobel laureate (mentioned in the transcript)
– Johannes Zutt – World Bank representative (referred to as “John” in the discussion) [S8]
– Iqbal Dhaliwal – Global Director of J-PAL at MIT [S9]
– Ufuk Akcigit – Macroeconomist (provides analysis on creative destruction and AI’s impact on economies) [S11]
– Anu Bradford – Expert on AI governance and regulation (contributes perspectives on AI sovereignty and regulatory approaches)
Additional speakers:
– None (all speakers in the transcript are covered by the list above).
Jeanette Rodrigues: The session opened with Jeanette thanking the participants and stating the panel’s aim: to explore whether artificial intelligence (AI) will narrow or widen development gaps in emerging economies and to identify the policy levers that should guide real-world implementation [2][3][60-65][71]. She noted that this was the fourth AI summit (the first having been held in the UK) and that participants repeatedly described the first session as “full of fear” about AI, framing the debate as a balance between hope and fear [2-3].
Johannes Zutt: Johannes described AI as a structural transformation already reshaping economies worldwide [6-7]. He clarified that the World Bank does not develop AI applications itself; its comparative advantage lies in advisory work-ensuring data reliability and helping governments create “AI sandbox” environments for experimentation [45-46]. He highlighted basic constraints in many low-income countries-unreliable electricity, weak broadband, low literacy, and reliance on very simple devices [26-31]. To address these gaps he introduced the Bank’s “small AI” agenda: affordable, locally relevant tools that function with limited connectivity, data and skills [34-36], and noted that the Bank is assisting governments in setting up AI sandboxes for pilots [49-52]. Examples included India’s digital identity programme and farmer-focused phone applications, which illustrate how small AI can be deployed at scale when supported by government standards and private-sector innovation [39-42][45-46][49-52]. He also pointed to AI-enabled services in education and health that can fill skill gaps for teachers and frontline workers [21-33].
Michael Kremer: Michael presented concrete public-good AI interventions. He explained that the Indian government’s AI-generated weather forecasts are a non-rival, public-good resource, justifying public investment [263-267]. The forecasts correctly predicted an early monsoon in Kerala and a later-than-expected progression elsewhere, becoming the only source of information for millions of farmers [263-267]. He also described AI-enabled traffic-safety tools-automated traffic cameras and the HAB (AI-based driver-licence testing) program-which have reduced unsafe driving by 20-30 % in pilot sites [263-267][268-277]. Emphasising the role of multilateral development banks, he argued that market failures leave critical public-good AI under-invested and proposed evidence-based innovation funds that follow a four-step evaluation framework: 1) model performance; 2) user impact; 3) scalability; 4) continuous improvement [284-291].
Ufuk Akcigit: Ufuk offered a macro-economic perspective on AI-driven creative destruction. He distinguished a “foundational layer” (compute-, data- and talent-intensive) with high entry barriers that tends to concentrate power, from an “application layer” where low barriers enable small firms to compete with incumbents [85-93][94-98]. He warned that concentration at the foundational level can spill over into downstream markets, limiting inclusive benefits [99-101][324-340]. He also questioned why entrepreneurship has historically been weak in emerging economies-citing family size and gendered labour dynamics as key determinants of firm growth-and argued that without reforms to the business environment, AI alone will not spark the desired dynamism [111-115][84-85].
Anu Bradford: Anu focused on governance and AI sovereignty. She described the European Union’s AI Act as a “rights-driven” framework that seeks to protect fundamental rights while distributing AI benefits more broadly [173-176]. She argued that the Global South must develop its own regulatory sovereignty, adapting lessons from the EU without merely copying them, to ensure AI serves local public-interest goals [167-176][181-190]. She also warned of the geopolitical concentration of AI capabilities in the United States and China, noting that supply-chain choke points in semiconductors and raw materials create strategic vulnerabilities for developing nations [357-371].
Iqbal Dhaliwal: Iqbal illustrated the impact of “small AI” on the ground. In a pilot in Indian public schools, AI automated routine spelling checks, freeing teachers to focus on higher-order learning and thereby improving educational outcomes [236-247]. He stressed that the success was demand-driven-teachers, students and districts all asked for the tool-and that similar time-saving AI could benefit health-frontline workers [248-250]. He identified two recurring patterns: (a) trust-highly accurate AI diagnostics can fail in practice if users lack trust or proper training [294-318]; and (b) institutional adaptation-the GST-fraud detection model was not scaled because it threatened existing discretionary power, highlighting the need to adapt processes alongside technology [309-322].
Points of Consensus:
– All speakers concurred that AI’s transformative potential is contingent on basic infrastructure (electricity, connectivity, literacy) [6-7][61-71].
– The panel uniformly endorsed the “small AI” approach as a pragmatic pathway for low-resource settings [34-36][236-247].
– There was consensus that robust, rights-based yet locally adaptable regulation is essential to prevent misuse and manage risks such as job displacement and market concentration [21-33][173-176].
– Participants agreed that public-sector investment and evidence-based innovation funds are needed to develop AI public goods that the private market will not provide on its own [133-158][263-292][266-271].
Key Disagreements:
– Ufuk warned that the compute-heavy foundational layer will entrench concentration, whereas Johannes’s emphasis on deploying small AI did not directly address this structural bottleneck [94-98][34-36].
– On regulatory sovereignty, Anu advocated for a rights-driven, locally tailored framework, while Jeanette highlighted the dominance of US and Chinese AI developers and questioned whether true sovereignty is achievable [162-166][167-176].
– Johannes identified job losses as a challenge, whereas Ufuk called for a deliberate slowdown of AI adoption to give workers time to adjust [22-24][405-412].
– Johannes stressed formal governance mechanisms, whereas Iqbal argued that trust, training and system-level adaptation are equally critical for successful deployment [21-33][309-322].
Key Takeaways:
1. AI can be a powerful catalyst for productivity in agriculture, health, finance and education, but its impact is limited by infrastructural deficits [6-7][84-85][61-71].
2. The World Bank’s “small AI” strategy-affordable, offline-capable tools co-designed with governments and private innovators-offers a viable model for low-connectivity contexts [34-36][236-247].
3. High entry barriers of the foundational AI layer risk increasing market concentration and talent migration from academia to incumbents, threatening inclusive growth [94-98][324-340][342-347].
4. AI sovereignty requires rights-based regulation that can be customised to national priorities while learning from the EU’s approach [167-176][173-176].
5. Rigorous evaluation-covering model performance, user impact, scalability and continuous improvement-should guide AI pilots, as outlined by Michael [284-291].
Concrete Actions:
– The World Bank pledged to expand small-AI sandboxes across South Asian states, collaborating with governments to ensure interoperability and offline functionality [45-46][49-52].
– Multilateral development banks were urged to scale evidence-based innovation funds such as Development Innovation Ventures, providing tiered financing for pilots, rigorous testing and eventual scale-up [266-271].
– India’s digital identity and payment infrastructure were highlighted as foundational assets that other developing nations could emulate [39-42].
– Private-sector developers were encouraged to create demand-driven applications that free frontline workers’ time, while regulators were asked to adopt a rights-driven framework that balances innovation with safeguards [173-176][236-247].
Unresolved Issues:
– How can policy prevent the foundational AI layer from cementing incumbent dominance?
– What mechanisms will align AI talent pipelines with local ecosystems to avoid excessive brain-drain?
– How should finance ministers balance AI sovereignty with geopolitical dependencies in the semiconductor supply chain?
– How can public-sector procurement be reformed to avoid monopsonistic lock-in while ensuring rapid adoption of proven tools?
These questions point to a need for further research on competition-friendly policies, talent development programmes and procurement reforms [342-347][357-371][397-398].
Rapid-fire Closing:
– Iqbal warned that unchecked market concentration could become a regrettable legacy [382-384][324-340].
– Anu cautioned that over-reliance on generative AI might make humanity “dumber” if critical thinking is outsourced [387-392].
– Michael echoed the risk that public-sector inertia could deny the poor access to AI-driven services [394-398].
– Ufuk highlighted the labour-market risk of rapid AI adoption outpacing job creation [405-412].
Overall Assessment: The discussion moved from an initial optimism about AI’s transformative power to a nuanced, evidence-based appraisal of the structural, regulatory and societal challenges that must be addressed. The consensus calls for immediate, collaborative action to build the enabling environment, fund public-good AI, and design governance that safeguards inclusive growth while preserving the innovative dynamism essential for emerging economies to thrive in the AI era [414-416][382-388].
all around the Bharat Mandapam. So once again, thank you very much for your time this afternoon and for choosing us to have a conversation with. To start off, I would like to introduce John, who will make some opening comments for the World Bank.
So thank you very much, Jeanette. It’s a great pleasure to be here speaking to all of you this afternoon. Over the past week, we’ve heard from a lot of world leaders, tech leaders, experts from across many, many countries about how AI is fundamentally reshaping our world, presenting not just a technological shift but a structural transformation with profound implications for economies and societies everywhere. For emerging markets and developing economies, as for all economies, AI could be a game changer. So sorry, that probably helps. I thought the mics were on. So, you know, for all countries, but especially for emerging markets and developing economies, AI can be a game changer, a unique opportunity to leapfrog longstanding development challenges.
It offers clear opportunities to enhance growth and productivity. We recently did some work in South Asia at the World Bank Group to see what sort of impact AI was having on jobs in the region, and we found that approximately 15 or 16 percent of jobs here have strong complementarity with AI. AI enables people in those jobs to expand their skills and their effectiveness in delivering the products and services that they are trying to provide. It also helps, you know, very, very diverse groups of people in many, many different sectors of the economy. It helps farmers to identify pests on their crops. It helps farmers to identify pests on their crops, diseases in their crops, and also how to address them.
It helps farmers to identify pests on their crops, diseases in their crops, and also how to address them. It helps nurses to identify the ailments and illnesses that their patients may be suffering, particularly the ones that they’re not very familiar with, but that they can research using appropriate AI applications. It helps financial institutions to understand better the ability of borrowers to take on loans, which, of course, expands the ability of the borrower to expand his or her business. So there’s clearly enormous potential for AI to fill skill gaps in the areas that I mentioned, also in education, in health care services, to detect patterns, to generate forecasts, to guide the allocation of public resources, and so on.
Of course, at the same time, on the flip side, AI also creates a number of challenges. One of them is there will be some job losses, particularly sort of entry -level jobs that are very much knowledge or document -based, performing relatively rote work that can be taken over by automation. And we’re actually seeing this in the World Bank Group. We went and looked at the number – the types of jobs that we are advertising these days compared to a couple of years ago, and what we found is that that layer, sort of at the bottom of the professional classes inside the bank group, there’s just fewer of those types of jobs being advertised in the World Bank Group today than there were a few years ago.
At the same time, you know, particularly for developing economies and emerging markets, many of them are going to struggle to harness the potential that AI offers because of very basic issues around the foundations for effective AI use. They may not have reliable electricity. We can start with that very basic one. They may not have an internet backbone that’s sufficiently strong. People in these countries may not have very, very basic skills of literacy and numeracy that enable them to work effectively with higher end devices. They may need to use very, very basic devices, not even smartphones, and rely on voice communication, asking a question and hearing a response. So there may be struggles of that kind in developing countries and emerging markets.
And I’m not even talking about all the governance and regulatory safeguards that can also come into play. So the question, of course, is how can emerging economies, developing markets, harness the potential of AI and avoid the pitfalls? And for us in the World Bank group, we’ve been very, very focused on focused recently on basically small AI. Small AI meaning practical, affordable, locally relevant AI that addresses specific problems and also works where connectivity, data, skills, infrastructure are fairly limited. And this is extremely important in countries like India where all of those conditions can apply. And yet there’s tremendous potential for people to expand their, to grow their productivity if they have timely access to information of the right kind in their local language tailored to their specific circumstances.
So that’s what we are trying to do in South Asia today and across the globe actually. And this is really about some of the examples that I mentioned earlier, having bespoke… applications that help farmers to do very basic investigation of the types of issues that they’re facing using their phone to analyze what’s going on to identify it to find out how to address it even to find out who within their local area in their market space can help them by providing the tools or the products that are necessary to address whatever they’re running into so India of course is a very strong example of what’s possible India has been a leading country in digital innovation for quite some time after the United States and China it has the largest if you like digital universe you in the in the world today it’s got some very good foundations there’s the the digital identity program as well as the digital payment platform that currently exists.
There are lots of Indian firms that are innovating in AI, including in the small AI applications that I’ve been talking about. And the governments of India have an objective of ensuring that there is AI for all. So they are very, very aware of the challenges that need to be overcome to make AI accessible to a very, very broad spectrum of the population and not just the very rich that, to some extent, need assistance the least, right? It’s the poorer parts of the country that benefit the most because they will be leveraging a tool that they are not very familiar with and have not been using that much in the past. So we’re working in India.
We’re working in a lot of different states, Uttar Pradesh, Maharashtra, Kerala, Haryana, Telgana. these different aspects working with governments to work on the foundational elements, interoperability, making sure that the accessibility is possible, that programs can run offline as it were so that people who aren’t able to get online all the time can benefit and so on. And then we’re also working with private sector investors who are developing apps. I mean we’re not actually developing many apps ourselves. That’s not really in our comparative advantage. Our comparative advantage as the World Bank Group is to do the more advisory work, make sure that the backbone information that’s embedded in the application is reliable and trustworthy because of course that’s critical for ensuring successful uptake.
But we are helping governments to create. We are helping governments to create the space that enables experimentation in AI sandbox to develop the different applications that people in this incredibly creative country are coming up with to help people get on with their work and become more productive. So I think it’s important to recognize that if we’re going to make effective use of this tool, we need both a public -facing effort to address the standards and the other issues, the interoperability and so on that I mentioned before, but also a private -sector -facing effort because it’s the private sector that’s actually generating, creating most of these applications that are working, particularly in the small AI area.
We’re doing a little bit on bigger AI. There’s obviously a connection between the two. Big AI can, through computational power, generate new knowledge that can help us to do things that we haven’t done so well in the past much, much better. But for… There are countries like India translating that. into small AI will also be very, very important for uptake. So I’m looking forward to hearing from all the distinguished speakers in this panel about their thoughts on what’s happening today in this sector. So thank you very much.
Thank you very much, John. John spoke about, of course, the use cases for AI, and on the other side of the spectrum we have the large language models, we have the foundational AI. But no matter where you sit on the spectrum, no matter where your interests lie, AI, innovation never disperses and never diffuses equally. Today on this panel, I hope to unpack what determines whether AI narrows the development gap or whether it widens the development gap. Especially we are looking to talk about the real world. What should policymakers in the real world think about and keep at the top of their mind as they go ahead preparing policies considering AI? Before I start, just setting the stage.
To a man, to a woman, everybody I spoke with who’s attended the first AI summit to today, this is, I think, the fourth AI summit being held. The first one was held in the UK. And without exception, all of them made it a point to tell me how the first session was full of fear. It was, oh, my God, AI is this terrible technology which is going to steal all our jobs, make us redundant. And when they come to India, they see the hope that technology and AI brings. And that’s the spirit of the discussion this afternoon, to figure out how can we balance both of those extremes, hope and concern, and go ahead in a pragmatic, policy -first way to prepare for the real world.
So if I could start with you, Ufuk, how do you think about AI? And especially, where do you see areas of creative destruction? To foster the innovation that we need.
Thank you very much. And so, of course, creative destruction is an important driver of economic growth in the long run. So that’s why, you know, it’s an interesting question how AI will affect creative destruction in general. Of course, we are at a very early phase of AI, and it’s a GPT. And typically, you know, when GPTs are emerging, there’s a huge surge of new businesses. And this should not be misleading. I think the main question we should be asking ourselves is what will happen to the creative destruction in the future? How does the future look like in terms of creative destruction? And I’m a macroeconomist, so that’s why I like to look at this with a, you know, bird’s eye view.
And I would like to, you know, separate advanced economies from emerging or developing economies. So when it comes to advanced economies, there, again, we need to split the issue into two layers. One, the foundational layer. and the other one is the application layer. When we look at the application layer, it’s great. You know, the entry barriers are low. Small businesses can do what only large businesses could do in the past, and, you know, they can do their accounting, marketing. You know, there are so many opportunities now. The entry barrier is low. As a result, this suggests that, you know, this is going to be more, you know, friendly for creative destruction on the application. But then there’s also the foundation layer, and I think that’s exactly where the bottleneck is.
When we look at the foundation layer, the entry barrier is really, really high, and, you know, the compute is very compute -heavy. It’s very data -heavy. It’s very talent -heavy. So as a result, you know, this market, at least this layer, is very concentration -prone. Of course, it’s very early. But, you know, normally we have to be concerned about the foundational layer and how things will pan out because this is the upstream to the application layer, which is downstream to foundation layer. So that’s why whatever will happen at the foundational layer will potentially spill over to application layer two. So that’s why I think we need to look at early indicators. But, you know, in the interest of time, I don’t want to go into the empirical evidence yet.
Maybe we can come back in the second layer. When we look at the developing countries, so I think, you know, I agree with Johannes. You know, I think AI is creating fantastic opportunities. So that’s why I think it’s really important to understand the opportunities as well as the risks for developing countries. And together with the World Bank, we are working on the world development. Report 2026, which is going to be on AI and development. And these are exactly the issues that we are focusing on. But I think before we go into those details, we should ask ourselves one major question. Why was there no entrepreneurship and dynamism before the AI revolution in emerging economies? Why was, you know, when we looked at the firm’s life cycle, for instance, why was it not up or out?
Why was it not, you know, very competition friendly? Why did the best predictor of firm size in emerging economies or developing economies was the size of the family and or the number of male children? These are still lingering issues and AI is not, you know, will not bring magic unless we understand and fix the business environment in these economies. You know, AI will just create new tools. But at the end of the day, we need to make sure that the business friendly environment is there for entrepreneurs to come and exercise their ideas
Ufuk, that’s a very interesting leaping of point, the real world. And the intention of this panel is to get exactly there. So if I may turn to you, quite literally turn to you, Michael, and ask you about the real world. You’re obviously doing a lot of work on the ground. Where do you see the potential for AI to spur gains? And are there any really transformative breakthrough areas that you’re looking at right now?
Yes. Thank you. Thanks very much. You know, I don’t want to minimize the existence of forces that may widen gaps. I think that if policymakers, primarily at the national level, but also in multilateral development banks, take appropriate actions and make appropriate investments, then I think AI has the potential to substantially narrow some of the gaps. And, you know, I think the… which policy actions to take can be informed by thinking through relevant market failures and relevant government failures. Let me give a concrete example or two. So private firms have incentives to develop and improve applications of AI that can generate profits. But there are some very important applications of AI for public goods, for example, that will not attract commercial investment to measure it with their needs.
And that’s an area where I think governments and multilateral development banks can play an important role. And I think some of this very much echoes what you were saying about small models, but also I’ll mention the link between the two. So an obvious example where I think India has been a leader for the world is in the development of digital identity. You know, this is… will enable, as Ufuk was saying, this enables a lot of work by individual entrepreneurs, a lot of other applications. So that’s a huge success, and I think multilateral development banks together with India can help bring that to many other countries. Let me take another example, one that’s not as well -known, but picks up on your comment about farmers.
So one thing that’s critical for farmers, they have to make a bunch of decisions that are weather -dependent. You know, when do you plant, for example? What varieties do you use? A drought -resistant variety, another variety. That, most farmers don’t have access to state -of -the -art weather forecasts around the world. I’m not talking about one country. In low – and middle -income countries, they don’t have access to that. Now, there’s a huge advance. We tend to think of large language models, but obviously AI is pushing science forward, and that includes in weather forecasting. There’s really a revolution driven by AI. But weather forecasts are non -rival. They’re largely non -excludable. They’re the classic definition of a public good.
So there’s a strong rationale for national governments, in some cases supported by multilateral development banks, to make investments in producing and disseminating AI weather forecasts. Again here, India is a leader. So if you, India in particular, in particular, India’s, the Indian government distributed forecasts to AI weather forecasts to 38 million farmers last year. And the evidence suggests that farmers, both from India, from this particular case, that in areas, I’ll say a little bit about last year’s monsoon, it came early in Kerala and southern India, but then there was an unexpected delay in the progression. The AI forecasts got that right, that was the only source of information that reached farmers with that. In the areas, we did a survey above that line, and farmers are responding, and they transplant more, they use hybrid seeds more.
Evidence from around the world is consistent with this. Farmers respond to these AI weather forecasts. So I think that’s one example, but many others, and happy to discuss them in education and traffic enforcement and elsewhere.
Michael, your answer should be read the book. Okay. We’ve spoken about the use cases of India, but setting up digital IDs, of course, is a sovereign decision. It’s something India could do unilaterally. When it comes to the large language models, that’s not reality. The large language models are concentrated in the US, in China now with DeepSeek. Anu, in a world where you largely have the rules being set by the two large powers, the US and China, arguably, there’s of course the EU as well, and you’ve done a lot of work on that. Who sets the AI rules for the Global South? Is there even the possibility for the Global South to talk about sovereignty?
So I think the Global South has the same kind of incentive for their own AI sovereignty, including then regulatory sovereignty, to design the rules that better work for their economies, for their societies, for what the public interest in these jurisdictions calls for. But regulating AI is really difficult even for very established bureaucracies. You need to be able to make sure that it is an innovation -friendly, and yet you at the same time need to be careful in managing the risks for individuals and societies. So even very established regulators like the European Union have found it one of the most challenging tasks to come up with the AI Act. So there’s probably something to be learned from these jurisdictions that have gone ahead and done the kind of thinking that had then resulted into some of those regulatory frameworks that we have now in place.
So if you think about the choices that India has when it looks around, one of them is to think about, okay, how does the EU go about this? The EU follows what I would call a rights -driven approach to regulation. So what is really characterizing this, the first horizontal binding, so economy -wide regulation that the Europeans enacted, it is a regulation that seeks to protect the fundamental rights of individuals, the democratic structures of the society, and that also seeks to ensure a greater distribution of the benefits from AI revolution. So the European approach is very conscious that it wants to also share some of the benefits so they don’t all go to the large developers of these models, but individual use as society at large.
smaller companies benefit from AI as well. So there’s something I think the Europeans can teach in terms of that regulatory approach in addition to maybe then some details of how that regulation in the end was constructed. But just one word, India is a formidable economy that doesn’t need to take a template and plug it into the economy as such. I think India is in a very good position to take the lessons that serves its needs yet make the kind of local modification and variations that are more reflecting the distinct priorities of this country.
Anu, before I turn to Iqbal, a quick follow -up question to you. As India makes its own rules, where does the trade -off lie between regulation and innovation?
So this is very interesting because often I am based in the U .S., but I’m initially from Europe, and these two jurisdictions are described as the U .S. develops technologies and the Europeans regulate those technologies. many ways does India want the innovation path or the regulation path? And I think there are many votes who would go for innovation. But I really would like to debunk this myth that to me it’s a false choice to say that the reason we don’t see these large language models being developed in Europe is not because there’s a GDPR, the General Data Protection Regulation. It’s not because there is AI Act. So the reason there is a perceived innovation gap between the United States and Europe is, I think, four things.
So first, there is no digital single market in Europe. It’s very hard for these AI companies to scale across 27 distinct markets. Second, there’s no deep, robust capital markets union. 5 % of the global venture capital is in Europe, over 50 % in the United States. That explains why the U .S. has been able to take much greater steps in developing AI technologies. Third, there are legal frameworks and cultural attitudes to risk -taking. You are not alone. You are not alone. You are not alone. You are not alone. You are not alone. You are not alone. You are not alone. You are not alone. You are not alone. You are not alone. You are not alone. You are not alone.
You are not alone. You are not alone. You are not alone. You are not alone. You are not alone. You are not alone. You are not alone. You are not alone. You are not alone. You are not alone. You are not alone. You are not alone. You are not alone. You are not alone. You are not alone. You are not alone. You are not alone. You are not alone. You are not alone. You are not alone. You are not alone. You are not alone. You are not alone. I wouldn’t encourage you to replicate that because it’s very hard to innovate on the frontier of technological innovation because sometimes you fail. But you need to be then given the second chance.
And the fourth, I think, the sort of foundational pillar of the robust U .S. tech ecosystem is that the U .S. has been spectacularly successful in harnessing the global talent that has chosen to come to the U .S., including many Indian data scientists, engineers, who think that U .S. is the place where they can start their companies, scale their companies, fund their companies, U .S. universities can attract them. So the idea that choosing to follow… Or imitate aspects of the European rights protective regulation would come at the cost of innovation, we need to understand better what drives the technological innovation and whether regulation should
Thank you, Anu. Iqbal, turning to you. You’re working in an area of the world, South Asia, where what is regulation? What is enforcement? At the risk of sounding like a provocateur, it’s the Wild West a little bit. And therefore, we talk a lot in our part of the world about small AI, about targeted AI. My question to you is that what should policymakers keep in mind when designing AI -enabled interventions, especially when it comes to small AI and the targeted use cases?
vulnerable public schools all the way from 11th to becoming the second best performing state in just a matter of two or three years. Phenomenal results, right? But then you start saying, let’s unpack this. What was this thing doing? The first thing that they find out is that a lot of people are like, oh, does this mean that I don’t need teachers anymore? No, you still need the teachers. What it replaces is the road task of the teacher having to correct spelling mistakes, calling you to the room and saying, hey, you forgot your comma, you forgot to capitalize. Instead, AI takes care of all of that. And now the teacher can sit with you in the free time and say, how did you set up the structure of this essay?
Did you think about this analytically or not? And that’s the first insight that comes from evaluation. It frees up the teacher time. Everything that we do in the field ends up adding to teacher’s time, adding to the nurse’s time, adding to the Anganwadi worker’s time. Very few teachers do that. Free up time. So if your AI application can free up the time of the health frontline workers, first of all, that’s a winner. The second thing that is really important here was that this is a demand -driven thing, right? Like there was a demand by the kids to improve their essays. There was a demand by the teachers to free up their time. But most importantly, there was a demand by the school districts to show progress.
So I think those is kind of a great example of how everything comes together if you think about it ahead of time.
Ladies and gentlemen, a topper of India’s notoriously difficult civil services exam. So take Iqbal more seriously than you would as just a normal.
Thank you. I thought that was history now.
It’s never history in India, Iqbal. Michael, turning to you, almost as equal in accomplishment by winning a Nobel. What risks should multilaterals like the World Bank keep in mind? Or let me rephrase that actually. Is there a risk that multilaterals are moving too slowly relative to the technology?
I think there certainly is. As I noted before, there are certain areas where the private sector is going to move, but there are other areas where they’re not going to move quickly, and it’s going to be very important for governments and for multilateral development banks and for philanthropy to move. I think there are a number of approaches to this. One way is by encouraging innovation by setting up institutions like innovation funds, particularly evidence -based, to echo Iqbal, I think evidence -based innovation funds. So I’ll give you one example of something that I’m involved in. Development Innovation Ventures, that was initially set up in the U .S. government, but it’s now been relaunched independently. It has tiered funding, so there’s initially very small…
grants to pilot new ideas. Then there’s somewhat larger grants to rigorously test them as Iqbal emphasized and then for those that are most successful there’s funds to help transition them to scale up. I think why is that important? Well that’s important because if we’re thinking about the services that public services and there are other sectors where this is needed but there’s probably going to be insufficient competition. Private developers are going to come up with innovations but then there if they have to sell them to the government they’re facing a monopsonistic buyer. They’re not going to probably not going to get rich doing that. So some support to generate more in that market, generate more entrance in that market, well I think is very important.
It’ll also mean that prices will go down and quality will go up when the government does that thing. Does that. Let me, I’ll just again let me give a example of the potential of how you know we we tend to focus on certain examples time after time here let me give another another example that is you know something that I doubt many people here are thinking of when they think of AI you know one of the things that you know traffic safety and we’ve all been exposed to traffic in the past few days you know traffic is a real problem interfering with urbanization which may drive growth there are a lot of deaths from from traffic a lot of citizens around the world have very difficult and painful experiences with traffic enforcement well you know you can have automated traffic cameras that have the opportunity to improve improve traffic outcomes but also improve people’s perception of fairness in government India’s moving in this let me mention another thing that within traffic safety that’s being done Microsoft Research India developed a program called the India Research Program and it’s a program that’s been developed by the government and it’s a program called HAB that is for driver’s licenses and that it automatically uses AI to test are that are the drivers until they actually pass in their exams they when this was introduced it’s been introduced I believe in 56 sites across India hundreds of thousands of people have taken tests this way we took a leaf from a false book we followed up the we’ve got information from Ola on ratings on and the number of drivers who were rated as driving unsafely that went down 20 to 30 percent where hams had been installed so you know that’s something that was developed not by Microsoft’s main business but by Microsoft research we can just create some support for more ideas like that to be developed to be rigorously tested that can benefit India can benefit the whole world we are we are running out of time probably this is this is one place in in India where time is really respected and we have to end in time.
So I had a list of wonderful questions, but if I could now move to a space where we are really giving shorter answers and quick answers and the deeply, deeply interesting ones about who’s winning and who’s losing. Michael, if I could start with you, actually. We’ve seen many promising technologies fail to live up to their promise. How should we think when we are evaluating AI interventions? How should we think about it? What should be the metrics that we use? Okay. First, model evaluation. So AI companies typically do that part. How good is the model output for specific tasks? You know, forecasting the weather. Does it do a good job? Does it match your local language well?
Second, user impact. Here, I think there’s a role both for sort of initial pilots akin to a medical efficacy trial. If you put the work into trying it, does it lead to improvements and outcomes for the users? Second… scalability and usage at scale that’s more like an effectiveness trial in medicine that it’s important to think not just about the tech but also about the human systems are the teachers actually going to use the product I think is it is an example how can you get the teachers to use the product and then the fourth area is continuous improvement you want a system that improves the underlying models so I think in procurement we might want to think about requiring continuous a B test publicity about what the what the impact usages and impact is and perhaps even thinking about requiring open access as part of the procurement package
thank you Michael. Iqbal, I want to flip that question to you where do you see where do you see hype in the promises of AI that you don’t think will play out
I think hype is natural because the technology is exciting. It’s a general -purpose technology. It’s evolving so quickly. The marginal cost of deployment for the next users is very low. It’s multimodal. Today you are doing it in text. Tomorrow you’re doing it in video. Day after tomorrow you’re doing it on audio. Everybody who has a smartphone has it. So I can understand the hype, right, like where it is coming from. But I think what we really need to do is separate the hype from the reality on the ground. And the reality on the ground is that many of these technologies are not having the final impact that we are having. And I see kind of two, you know, like once again my job at J -PAL always, you know, sitting at the top is like to say not worry about one professor’s evaluation or one researcher’s evaluation, but say when I connect all these dots, what am I seeing?
And I’m seeing two patterns. One is about trust in technology, and the second part is about the reality of the policy world. Let me elaborate quickly on both. Trust in technology. There are studies which found that even if you give doctors and frontline health care workers access to diagnostic tools, including radiology, tools, using AI, AI enabled prediction of the diseases, oftentimes it doesn’t lead to an improvement in results. And when you try and unpack that, even though this technology worked even better than the human intervention in the lab, right? So some of these diagnostic things can work, have better predictability in the lab, but in the field, they end up decreasing, not only is their efficiency lower, but it lowers the efficiency of the doctors, because we have not trained them enough important.
And the second thing is the enabling mechanism, the world around us. We just assume that just because the technology works, even if it works in the field, the rest of the system will adapt to it. No, you have to adapt the system to the rest of the world. So this example quickly comes from India, where, you know, we have a with one particular state government, we try to improve the collection of value added taxes, it’s called GST in India, there is a whole worry about bogus firms that are created to get these GST or value added tax thing. The machine learning algorithm is able to increase the probability of predicting a bogus firm from 38 % to 55 % in one shot at a very, very low cost.
When it came time to scale up this program by the government, they refused to scale it up because you think about it, you have taken away the discretion of the human to decide whether they should raid Michael’s firm or they should raid Iqbal’s firm. That is power. And if you haven’t thought through that point, what is the point of technology?
I won’t terrify anyone in the room by asking why they didn’t want to scale up this tech. But talking about weeding out the bad actors, talking about firm -level decisions, moving on to UFOOC, does the firm -level evidence show productivity gains diffusing evenly across?
So just going back quickly to the question of the firm. In the earlier model that I highlighted, I think it’s important to understand what’s happening at the upstream. so that we can then understand where things will be going in the future. And the evidence there, the early signs, is a bit worrying. So first of all, when we look at, for instance, the dynamism or market concentration in the U .S., market concentration has been increasing since 1980 but in an accelerating way after 2000. So that’s the first set of evidence. The second set of evidence comes from how innovative resources are allocated across firms. And when we look at the inventors who are creating the creative destruction and technologies, there’s a massive shift towards market incumbents.
And when I say incumbents, those firms that have more than 1 ,000 employees. In around 2000, 50 % of employees used to work for incumbent firms in just 10 years. That shifted. To more than 60%. A massive reallocation of innovative resources. And the final piece of evidence, and we are going to release this study next week, we looked at the universities, how AI is impacting universities, and we look at the AI publishing scientists. And AI publishing scientists in academia, the top 1%, used to make around $300 ,000 in 2000. It went up to $390 ,000 over two decades. Similar people in industry used to make around $550 ,000. Now it went up to $2 million. And there has been two breakpoints. One of them was in 2012. The other one was in 2017.
Of course, image processing and then the foundational model revolution in 2017. The more worrying part about this, which brings me back to the foundational model side of things, is that this created a massive out -migration from academia to industry.
And after 2017 especially, B2B. When the compute and infrastructure became so important. And then we saw the rise of AI. The target or the destination is large incumbent information companies, which again highlights where things are going in terms of the concentration. And the worrying part also is that when people are moving to industry from academia, their publication record goes down by 50%. They start patenting by 600 % more after they move, which means that we are moving from open science to more protected science. Now, spillover is extremely important for creative destruction, for the future of innovation. So that’s why, and if we will keep the foundational layer contestable, I think that the fundamental players there will be universities.
And keeping universities in a healthy way is extremely important, but there is very little discussion on this, which I think before it gets too late. Because once you start buttoning the wrong button, and then the rest will follow wrong as well. So that’s why I think we have to have this frank conversation early on in the game, otherwise it might… too late.
Ufuk, what you spoke about boils down to something Iqbal mentioned as well, power. Because power still makes decisions in this world today. So Anu, before I move to the final section of this panel, if I could ask you if the finance minister of a developing country let’s say India, comes to you and asks you, Anu, how should I think? What would you tell her?
So today if you think about how much political power but also geopolitical power is shaping our conversations around AI it is something where I think each country is now pushed towards greater techno -nationalism, techno -protectionism AI sovereignty has become almost a sort of uniformly goal for everyone. But I would remind even when encountering players like the United States and China that nobody in today’s world will be completely sovereign when it comes to AI space. If I just take one layer of the AI stack as an example. What is now driving a lot of the global AI race is this idea that we want to do frontier AI we want to have these powerful foundation models.
That means you need to have a lot of computers. You can’t have a lot of compute unless you have access to the high -end semiconductors. The U .S. is well positioned there. It is hosting companies like NVIDIA. The U .S. leads in the design of semiconductors. But who is manufacturing them? We really need to think about the role of Taiwan there. But then the Europeans have ASML in the Netherlands that leads in the high -end manufacturing with the equipment needed for manufacturing. But that is dependent on chemicals where Japan is leading. And the entire supply chain relies on raw materials from China. So ultimately, all these choke points can in principle be weaponized, but that is not ultimately a sustainable strategy.
Even President Trump had to walk back some of the export controls to China because Chinese were saying, okay, then the raw materials are not coming your way. So there are the potential ways to weaponize these interdependencies that ultimately make us all poorer. So as a finance minister of India, when approaching other middle powers, the great powers,
Easily said than done. Our final, final section is, of course, the rapid fire round. We all love this in this room. In one sentence, in one sentence, if I could ask all of you, and Johannes, you’re not getting away easily, you’re going to answer this as well. So in one. if I could ask you, we’re sitting in New Delhi 2035. Could you predict one development outcome that will have dramatically improved with the use of AI and one risk we’ll regret not addressing now? I guess you already know my second answer.
I think the concentration, the future of market concentration is something that we should be concerned about and we might regret not having discussed this sufficiently in 10 years. On what will change in a positive direction, clearly health care and education, I think. It’s a no -brainer.
Anu?
So first of all, it’s so inspiring to hear all the use case examples, whether we talk about traffic or agriculture or education, because I often talk about the risks and the downsides, so it’s a really good reminder. I’m personally very excited, especially what happens in the education space but also in the health space. In terms of the risks, I think one thing that we are not paying attention to, and what I would even call a systemic risk, is the idea that many worry about AI getting almost too smart. But I am more worried about us getting dumber as a humanity. There is a temptation to start skipping steps, outsourcing your thinking and your creativity to these models.
And as an educator, when I think about how I will teach my students to use generative AI to enhance but not substitute their capabilities, we will just make a tremendous mistake if we just forewent that hard work, that beautiful moment of thinking hard problems and creating and investing in our own capabilities. And all that just cannot be so outsourced, because otherwise we don’t even know what kind of questions we should be asking the AI going forward.
Michael.
I agree that there is huge potential in health. and education. I think we’ll see big improvements in that, but the risk is that the public sector won’t adopt these, and therefore the poor won’t have access to them. And that’s because the public sector, as Iqbal indicated, the government systems and the government workers may not adapt to use these. There’s also risks of copycat regulation that are over -focused on certain problems that other countries may be worrying about, but might not be relevant for emerging economies. And then final risk is that the procurement systems are just set up in such a way that we don’t get sufficient competition, we get lock -in, and then we just don’t wind up with good quality.
Thank you, Michael. The buzzer’s down, but I’ll take a risk and quickly run through the other.
Yes. I think I am much more optimistic about the government actually adopting this thing. Whether it is when you call 100, your call is going to get answered very quickly. The PCR van is going to be at your house much faster. The hospitals are able to be able to link your health record. So I think the government sector productivity is going to improve leapfrogs. The biggest risk, I think, is definitely the labor market. If there was a dial where I could slow down the adaptation and give time to the labor market to catch up, that’s my biggest worry. You talked about entry -level jobs. An entry -level coding job might be an entry -level job in the United States.
It’s the aspirational job that created Gurgaon’s and Noida’s and Mohali’s of this country. And those people are going to be running out of jobs very, quickly. And I think the labor market, whether it is ESI, Provident Fund, Gratuity, we are piling on and making it harder and harder to hire labor. when, on the other hand, capital is not taxed. We are giving incentives to people to use AI, and we are taxing them through provident fund and labor market regulations to hire labor. And I think that, for me, is the biggest risk, actually.
So I think that for the first time in human history, we may actually have the tools available to enable us to target poverty reduction, poverty elimination initiatives on individuals. And that could be tremendously transforming. But at the same time, I do worry that we will not get the governance right or we won’t be able to make that governance sufficiently robust to prevent abuses.
Thank you very much to all of our panelists and to you for your time and attention once again. I had the very rare fortune of being able to peek into Michael’s screen while he was speaking, and I saw all the messy human notes. Our panelists are definitely not outsourcing their thinking anytime soon, and thank God for that. Thank you, ladies and gentlemen
“This was the fourth AI summit (the first having been held in the UK) and participants repeatedly described the first session as “full of fear” about AI.”
The transcript excerpt S12 explicitly states that this is the fourth AI summit, the first was in the UK, and every previous session was described as “full of fear.”
“Jeanette Rodrigues is the moderator/host of the panel discussion.”
Source S1 lists Jeanette Rodrigues as the moderator/host of the discussion.
“Many low‑income countries face unreliable electricity, weak broadband, low literacy, and rely on very simple devices.”
Infrastructure challenges such as unreliable electricity and limited internet access in low‑income settings are documented in S93 and S92.
“The World Bank helps governments create AI sandbox environments for experimentation.”
AI sandboxes are discussed as a mechanism for responsible innovation in developing countries in S90, though the source does not specifically name the World Bank.
“AI‑enabled services in education and health can fill skill gaps for teachers and frontline workers.”
S24 describes AI decision‑support tools for frontline health workers and assessment tools for teachers, confirming the claim.
“India’s digital identity programme and farmer‑focused phone applications illustrate how small AI can be deployed at scale when supported by government standards and private‑sector innovation.”
S95 references India’s digital public infrastructure—including identity, payments, and UPI—showing government‑led platforms that enable large‑scale digital services, providing context for the claim.
The panel shows strong convergence on four main themes: (1) AI’s transformative potential is contingent on basic infrastructure and a supportive business‑environment; (2) “small AI” solutions that are affordable and locally relevant are vital; (3) robust, rights‑based governance and regulatory sovereignty are needed to manage risks and prevent concentration; (4) public‑sector investment and evidence‑based funding mechanisms are essential to deliver AI public goods and avoid lock‑in. Concerns about market concentration and talent migration are also widely shared.
High consensus across speakers, indicating a shared understanding that policy, infrastructure, and governance must accompany technological advances to ensure AI narrows rather than widens development gaps. This consensus suggests that future initiatives should prioritize coordinated public‑private funding, rights‑focused regulation, and capacity‑building to harness AI for inclusive development.
The panel shows broad consensus that AI can help narrow development gaps, but disagreements arise around how to manage structural concentration, the balance between rights‑based regulation and sovereign policy, the primary mechanisms for scaling AI (public‑private collaboration vs evidence‑based funding), and the most effective way to protect labor markets and build trust. These divergences reflect differing priorities among economists, development practitioners, and policy experts.
Moderate to high – while there is shared optimism, the participants differ substantially on the pathways and institutional levers needed, implying that coordinated policy design will require reconciling these perspectives to avoid fragmented or counter‑productive AI strategies.
The discussion was shaped by a series of pivotal insights that moved it from a broad, hopeful overview of AI’s potential to a nuanced examination of structural, institutional, and societal constraints. Johannes’s framing of “small AI” and the infrastructural gaps set the stage, while Ufuk’s two‑layer model introduced a structural lens that underpinned later concerns about concentration. Michael’s concrete public‑good example and Iqbal’s field‑level successes grounded the debate in real impact, whereas Anu’s deconstruction of the regulation‑innovation myth and her warning about cognitive atrophy broadened the policy conversation. The recurring theme of power—whether in the GST model or in market concentration—highlighted governance as a decisive factor. Collectively, these comments redirected the panel toward concrete policy levers (e.g., evidence‑based innovation funds, regulatory design, support for universities) and underscored the need to balance AI‑driven productivity gains with safeguards against inequality, concentration, and loss of human agency.
Disclaimer: This is not an official session record. DiploAI generates these resources from audiovisual recordings, and they are presented as-is, including potential errors. Due to logistical challenges, such as discrepancies in audio/video or transcripts, names may be misspelled. We strive for accuracy to the best of our ability.
Related event
