Indian Council of World Affairs

 

Unidentified Speaker: Good afternoon, distinguished guests. Thank you for joining us this afternoon. Before we proceed, may I kindly request you to switch off your phones or put them on silent mode. On behalf of Indian Council of World Affairs, it is my pleasure to welcome you all to today's panel discussion, 2026 India's Tryst with AI. Artificial intelligence is rapidly shaping economies, governance structures, and global power equations in real time. As India prepares to host the AI Impact Summit 2026, today's discussion aims to examine India's role in shaping responsible inclusive global AI governance, balancing innovation with ethics, data sovereignty with openness, and national priorities with the interest of the Global South.

 

This panel discussion aims to explore the strategic choices that will define India's AI trajectory and its contribution to the emerging global norms. Today's discussion will be chaired and moderated by Ambassador Rakesh Sood, who is a former Indian diplomat with extensive experience in diplomacy, security, and governance. Our panelists are Dr. Anulekha Nandi, an independent researcher and consultant. Her expertise lies in digital governance, AI, emerging tech, and digital infrastructure. Joanne D'Cunha, Program Manager with Center for Communication Governance at National Law University, Delhi.

 

Our third panelist is Mihir Kulkarni. He'll be joining us online. He's an AI scientist and team lead with Vadhvani AI. He's also working as a machine learning consultant with India AI Mission. Dr. Nikhil Agarwal, he'll again join us online. He's a Managing Director of the Foundation for Innovation and Technology Transfer IIT Delhi, a renowned expert in startups, innovation, and policy. The discussion will be followed by a brief Q&A session, which will be moderated by the chair.

 

With this, may I now request Ambassador Sood to make his opening remarks and conduct the proceedings. Thank you.

 

Rakesh Sood: Thank you. At the outset, let me thank ICWA and compliment them on taking this initiative for convening this panel. As everybody in this room probably knows, AI is certainly the flavour of the month, with India's AI Summit scheduled to take place later in February. Now, India has been relatively a newcomer into the field of, into the domain of AI, because our first policy paper came out only in 2018. But since then, there have been a number of think tanks that have gotten engaged, there have been a couple of other policy papers, there were guidelines that were issued last year, and so on.

 

We also bring a new kind of a direction into the domain of AI, because this is based on our journey into the world of digitalization, as it were. That, I would say, is something that has created or given us the framework for a digital public infrastructure. And if we look at the elements of that, we find that it is basically, consists of… we have the Aadhaar card, the identity, we have the UPI, which is the payments, we have a consent system, which we use to clear everything, and then we have the Digilocker, in which we can have certain documentation like passports and give access to people who need identity, or who need to verify identity from us. So EKYC and things like that.

 

And these things constitute a digital payment infrastructure. And they found widespread acceptance. So this is the experience that we would like to build upon when we enter the domain of AI, and if I look at the AI stack, then it will obviously have to be based on many of these apps, which will get modified to introduce the elements of AI into it. So one sort of element of this would be the apps, which we'll obviously have to develop, because our own compulsions with these apps will remain the same, that have powered our journey into the digital world, which is cost efficiencies, which is scale, which is to contribute to our economic development, to bring about inclusivity, after all these were the fundamental drivers behind the push towards the four pillars of the digital public infrastructure.

 

And these will remain the guiding pillars or the drivers as we adapt to the AI models. The apps, today the flavour is the LLMs. Now whether LLMs will define AI, there is a whole jury that is out there, a number of people who feel that LLMs is just one thing. From language models, we are already moving towards reasoning models. But is reasoning going to be the final goal that will give us AI? I would tend to think perhaps not, even with reasoning, because after all, a lot of human cognition, human decision making is not necessarily based on reasoning models, in our own minds, in our own heads.

 

We have reached where we are by virtue of evolution and evolution adaptiveness, or there is something known in the field of evolution known as complex adaptive mechanisms. That is something that has given us, which is why we don't consume the kind of data or we don't need that kind of data or that energy to reach our decisions as today's models do. So is that going to be sustainable? We don't know yet, but that's at least in today's context, that is one of the factors.

 

Then we also have other elements in terms of the compute, where we are trying to develop on the principle of inclusivity, develop a compute capability which would be more easily accessible to users and cheaply so, which brings in the cost element. Data, now we have data, but we need to ensure that it is quality data. And finally, the kind of energy resources, now we all know we are in the middle of a whole industrial electrification program and that has to increase hugely if we have to meet the goals of Viksit Bharat. Our electricity consumption in per capita terms, is just about 20% of our total per capita energy consumption, which is very, very low compared to the developed world or even with regard to China. So with some of these, we will need to work our own way.

 

And now the last point that was mentioned, which we will be talking about in some of the questions, would be the role of the Global South and whether India can assume a role of leadership in this. Here, if you look at so far, multilateralism has not taken root. There have been efforts at multilateralism. Our summit is proof of that. And our summit has been preceded by three other summits. So if you look at the first one, that focused much more on safety. And at that time, the idea was looking much more on commercial AI adaptation and also its impact on the workplace.

 

Then came the second summit in Seoul. That kept the safety focus and added to it innovation and inclusion. Then last year, we had the Paris Summit, which President Macron and Prime Minister Modi co-chaired. And there, what we added was an element of operational frameworks and a focus on humanist values. And we tried to bring in values in that sense. We've called the Delhi Summit as an Impact Summit, in a sense that we want to highlight or we want to be able to demonstrate our focus on AI, which means how the measurable kind of an impact that AI can have in a developing society like India's, in terms of tangible outcomes, in areas which affects the common people, whether it is education, whether it is health, whether it is agriculture, whether it is climate change, sustainability, good governance, and so on.

 

What we want to be able to show is that, look, we are trying to make use of this to achieve economic development, to push broader scientific and technological development, while keeping intact our concerns on sovereignty, and while ensuring public safety. Now, it's a big call. It's a tall ask. So whether we can do it or not is something that we will have to wait and see. And what I propose to do in the next one hour or so that we have is to ask our panelists to dwell on this and ask them to make their opening remarks.

 

The first person is Dr. Nandi.

 

Anulekha Nandi: And thank you very much, a very good afternoon to all of you, and thank you very much to ICWA for having me here. And I think this is a very interesting month to be talking about AI, particularly with the AI Summit coming up. And I think one of the questions that confront India, particularly, is how do we define our strategic autonomy in this place? Particularly, with strategic technologies like AI with immense economic value across different sectors of the economy, different segments of the industry, how do we define our comparative advantage, dimensions of our strategic autonomy in this space, which is characterized by intense competitiveness between the U.S. and China, which dominate segments of the AI stack already?

 

And I mean, the elements of the competitiveness is quite well known. The U.S. dominates across all three, compute data and models, China on data and models, and particularly, the economic dimension is very important, because most of these models are currently being used in application developments within India as well. So within this context, how do we define the road ahead? This is important, because both the U.S. and China, while competitive, are taking two very divergent approaches, the U.S. as a pro-innovation and deregulatory approach, and while China is intensely doubling down on global governance of AI, particularly through its influence in standard setting bodies and so on.

 

So defining diffusion and definition of norms and standards globally. And the implications for this are cascading, because what standards are globally accepted determine market access and market penetration going forward. So it also kind of paves the way for diffusion of AI technologies, particularly diffusion of Chinese technologies, if those standards become the norm going forward. So while India is developing its innovation capabilities and data compute models as well, sovereign LLMs and so on, how do we define the way forward in terms of having a strategy for global governance of AI, not just in terms of principles and frameworks, but what are the concrete norms, standards, tools, and solutions we offer that make sense, both in light of our own domestic contexts as well as in terms of the global context as well as contextual realities of the Global South.

 

As we grapple with technologies that increasingly define what we can or cannot do, that are increasingly defining economic opportunities for countries that do not possess the entirety of the stack. Now, what I have looked at in my previous work is also not just developing the technological and economic capabilities within the technology, but the articulation of a governance framework that India proffers.

 

And this is, again, highlighting the reason it is important is because if you look at the Politburo study session in April 2025, China focused on development assistance, technological assistance to the Global South, building on its experience with the Digital Silk Road. And this becomes the basis for not just developing infrastructural capabilities in the Global South, but also modes and mechanisms of how they are governed. So it's both the technology, as well as the standards and norms on which they are based, kind of looking at two planks of adoption and market penetration.

 

So going forward is, how does India define its comparative advantage? What opportunities and platform does the India AI Impact Summit offer? In my previous work, I've looked at three areas of defining our comparative advantage in the near term. So one is focusing on complementary innovation. While we are developing foundational innovation capabilities, it is also important to understand where we can contribute now meaningfully in terms of what the Global South needs. So on what should development cooperation or AI cooperation be based on going forward?

 

There is a product that IIT Madras developed through its foundation, IIT Madras Pravartak Foundation. And it's called Compact AI. Basically, the AI helps large language models run on local systems like CPUs. So this becomes important because if we are to push last mile AI innovation, last mile AI adoption, convert AI potential to socioeconomic benefits that then have upstream economic effects, we need to look at how do we make AI innovation frugal? How do we make AI innovation accessible and adoptable at scale?

 

So complementary innovation like this help position India's comparative advantage in terms of what it can readily offer. It is still catching up when it comes to compute data and models. Why I say it's trying to catch up with models? We have a huge amount of data. Sorry, why it's trying to catch up with data? We have a huge amount of data, but the data is currently fragmented. So we need to have systems in place to make that data amenable and usable for uptake in AI innovation. So complementary innovation is probably one of the areas where we can suggest as a plank for outlining a comparative advantage.

 

The second is widening the ambit of trusted partnerships. We have looked at how dependency on these two AI technopoles can increase risk and vulnerability interfaces. So which countries can we partner with and in what and in which manner? We have our own comparative advantage, let's say, in complementary innovation. We have countries like the UAE where there's sovereign wealth funds. They have a dedicated AI ministry. We have countries like Switzerland who have launched a public LLM, as well as countries like Singapore and Brazil who share our experiences in building digital public infrastructures at scale.

 

So how do we leverage these complementary relationships? How do you widen the ambit of trusted partnerships? Also countries like Japan and South Korea have a strong R&D ecosystem as well as hardware manufacturing capabilities. So how do we define what directions and what modes and modalities of those trusted partnerships are? Which countries do we include within them? But in essence, kind of widening that ambit, widening that network and looking at exploring relationships with countries that can leverage our strengths and enhance cooperation going forward.

 

The third is kind of building on what sir said, our experience with digital public infrastructures. What India offers in this domain is a model of values-based leadership. While we have China coming in with development assistance, infrastructural development, but increased economic dependency. Increased debt traps. What India offers is a solution that is easily adaptable to each country's sovereign and social realities, as well as economic realities. So what we offer is a model of a values-based leadership in terms of AI cooperation going forward.

 

So how do we replicate that model in AI? What kind of solutions for governance, what kind of solutions, tools, frameworks, norms for governance do we offer? Do we put forward? That makes sense both within our contextual realities as well as resonates with the wider Global South. So the AI Impact Summit, in essence, offers these opportunities to forge these trusted partnerships, to have those conversations, to identify areas of alignment and cooperation. And it also offers the opportunity to think through with countries what interoperable AI governance frameworks look like.

 

On what foundations do we go forward and build trusted partnerships? What are the reciprocal advantages that both countries get in this space? And these three are the kind of planks or areas where I think as India moves forward in crafting its AI journey, these offer the areas through which it can determine its comparative advantage within the AI space, define the modes of governance that it wishes to pursue, and also kind of reduce or define a framework of strategic autonomy and AI cooperation and technological cooperation that works both for itself and the wider community of Global South countries and provides a model of AI cooperation, both for itself and countries in the Global South going forward. Thank you.

 

Rakesh Sood: Thank you.

 

Nikhil Agarwal: Ambassador Sood, can I make a small request, please, with your permission?

 

Rakesh Sood: Yes.

 

Nikhil Agarwal: So I'm really sorry. I apologize. I am not able to come there personally. The reason is because I've been asked by the Chief Minister of Uttarakhand to meet him tomorrow morning, so I am in the way. So if you can kindly allow me to speak first, because I've just taken a break at a nearby cafe just to attend this. So I know it's out of turn, but my apologies. So I'm requesting your permission.

 

Rakesh Sood: All right. But then will you be available for the Q&A or no?

 

Nikhil Agarwal: I will log in again. I will log in again because I'm on the road. So somebody can message me, then I'll be. All right.

 

Rakesh Sood: Why don't you speak now then? Thank you, sir. Thank you very much.

 

Nikhil Agarwal: So the chairman and all the dignitaries. I think it's a great honor to be here at speaking at ICWA. So it's a very important conversation and in fact I've been involved with AI for many, many years. I was with IIT Kanpur previously for a few years and now with IIT Delhi. One of my responsibilities to promote AI based companies. We have close to 200 plus startups working today at IIT Delhi. It is one of the most prestigious incubators. 50% of the uniforms come from our campus. So that provides us a great insight on how these technological advancement can provide a better impact and value to the people.

 

The question today is about the innovations with Global South and I would like to approach this not from the technological perspective but from the policy and human impact perspective. Let me start with a very simple observation. For decades much of the Global South has consumed technology that were designed in different contexts. For example, it was built for high income markets, for stable infrastructure, for uniform user devices. But the challenges faced by developing nations are very different. They are not understood by the developed nations. They have limited resources, they have diversified language, and uneven access to services.

 

On top of it, there is a very large undeserved population which is increasing day by day. So that's the reason that India's experience has become extremely relevant because India has somewhat navigated some of those challenges and we have learned to innovate under constraints. There are examples that exist. There are dozens of examples exits how we have innovated from Mangalyaan to building up the AI models to creating a larger impact, there are dozens of that. We have built solutions that work with low bandwidth, low-cost devices, multiple languages, massive population scale.

 

And when an Indian startup creates an AI solution for rural healthcare or crop advisory or financial inclusion or even education access, it is not just solving the Indian problem, it is solving that exists across Africa, across South Asia and parts of Latin America. So let's take an example here. Think of an AI crop based advisory platform built for a smart farming in Bihar. It works on a basic smartphone, we have seen that. It gives recommendations in local language and uses a very limited amount of data.

 

The same model can be adopted for a farmer in Kenya or Vietnam. Or consider AI tools that help doctors in X-rays or detect early signs of diseases in an area where specialists are scared. These solutions are scalable, affordable, and designed for a resource-constrained environment. And I think this is what the Global South needs. So opportunity is very real. But the question is, how do we take these innovations beyond India? Something that we are also contemplating.

 

From a policy perspective, there are three important pathways. The first is credibility through deployment. One of the UPI great credibility is that we are able to deploy in UAE, Singapore, and other places. So Indian startups must prove their solution at scale with India. When a technology is already worked across millions of users in complex, real-world conditions, it builds that trust internationally. In many ways, India is a living laboratory. If something works here, most likely it will work in other developing nations.

 

The second is partnerships and diplomacy. Technology does not travel alone. India-ASEAN partnership, India-Israel technology relationship, India-Rwanda, there are a number of such things which I have seen. It travels through relationships, technology and relationships hand-in-hand. Government-to-government collaboration. Recently, we opened up a campus in Abu Dhabi and thanks to the effort of the Indian government, the campus of IIT has been opened in UAE. That's one of the great examples.

 

Development of partnerships and institutional linkages, that's my own experience. Institutional linkages can open doors for startups. We have already seen how India's digital public infrastructure has started inspiring other countries, not only the developing countries but also developed countries. AI can follow the same path if it is supported by the right policy and framework.

 

The last important part that I want to talk about is the support for global expansion. Many Indian startups are strong in innovation but they are weak in global market access. They face challenges in understanding the foreign regulation, localised language model, building distribution network, policy support, whether through the pilot or international testbeds, or a collaborative funding model, because we require a serious amount of money to do that, can reduce this risk and help to take these startups to the new geographies.

 

Now, certainly we have to acknowledge certain hurdles. The biggest barrier is a trust. The Indian technology is never considered to be the world class. So how do you create that trust, the other sensitivity about the data sovereignty, data security, and the dependence on foreign technology. The other challenge, what we have seen is the adaption. The AI model, which are trained in India, they need to be retrained for the local language in Global South, or the crop healthcare condition, or governance system in the other countries. That requires patient capital and long term development, which nobody venture capital does not give that kind of leeway.

 

But beyond all this, there is a larger strategic opportunity. If we look at the global landscape, some countries are leading in building the large foundational model. Others are dominating hardware, Taiwan and all China are dominating hardware. But India's natural strength is not about the large language models or dominating hardware. Something which we have seen is deployment in large scale, complex, diversified, resource scale, and constraint environment. And this is where we can lead. The Global South does not need the most sophisticated AI in the world, but it needs AI that works reliably in villages, small towns, public hospitals, government schools, and informal economy. We have dozens of such examples from IIT Delhi.

 

In many ways, this is a defining moment. Years ago, we have seen that India is globally known for IT services. Today, we have an opportunity to be known for something even more meaningful, which is creating AI solutions that are inclusive, accessible, and built for real impact. Sir, if we get it right, we may not be just exporting the technology, we will be contributing to better healthcare, better agriculture, better education, and better livelihood for many parts of the developing world. And that is what the fundamental of the trust.

 

And I believe it's a true promise what we hold for the Global South. And that's something which we have seen. I have been very deeply involved with India AI Summit. Myself as a jury member, as a participant, as a core team member. And this is something, a message which is going across that India should look at a testbed or creating technologies for the rest of the world. The idea is emerging from the West, but the implementation is happening in India. So I'll rest my case here. Thank you very much.

 

Rakesh Sood: Thank you. Let me now turn to our next panelist. Joanne, you have the floor.

 

Joanne D'Cunha: Thank you, sir. I hope I'm audible. Thank you to the Council for having this conversation. I think it's quite timely, like my panelist said, given we're in the throes of preparation for the AI Summit. I feel like we've come full circle with some of the discourse that we're having currently on data. Some of this thinking and strategizing for how India's data governance approaches should play out was happening back in 2019, while the regulatory landscape was still being determined. And now we're here with our frameworks, and it continues to be an extremely relevant conversation to be having, given the role that data plays in the AI ecosystem that we're trying to develop.

 

I'd like to quickly begin by highlighting three points on India's approach to data governance, how it's differed from some of the other jurisdictions, and maybe some key characteristics of its approach. First, India has had a very long journey, an evolving journey, with variations of its data protection frameworks, its data governance landscape itself. We've moved from different levels of stringency, different levels of compliance from where we started.

 

We've had different governance frameworks on data as well, and we've identified a current positioning that has taken some experimenting, both based off of internal understanding of what suits our context, but also learnings from other jurisdictions. And common and widely known, India's approach is often said to be this middle ground that we've taken in terms of our approach to data governance, compared to more popular models of governance that we're seeing in the EU, the U.S., sometimes China. We've sort of tried to balance our approach with heavy compliance, market-driven approaches, but also having entirely maybe sectoral approaches.

 

The second point in this journey has been a key characteristic, which is how India has been thinking of incorporating techno-legal approaches to governance within its framework, and that's by trying to bring in regulatory principles to technological architectures and infrastructure. The data empowerment protection architecture is an example of this, but also a similar approach that has been articulated more recently is in a white paper on an AI governance framework with techno-legal approaches that were released maybe about 10 days ago in sort of the lead up to the summit itself.

 

And the third point I want to make is, for a long time, the way that we've been envisioning data governance, and specifically data sovereignty within this thinking, has been sort of synonymous to conversations that have been happening on data localization. And this sort of tussle has played a huge role in the way that India's approach to the intersection of data governance and data sovereignty has happened. And as many of you know, the concept of data localization is often used in how we talk about what kind of data gets limited within the country, what kinds of flows you want to have between different kinds of countries.

 

And like I was mentioning, India's position on a lot of this has shifted over the years. And we've moved from different kinds of restrictions while trying to understand what is best suited for our thinking on what it means to balance having that kind of compliance and sovereignty, but also allowing innovation to foster. And the reason I highlight some of this is to set the context a little bit for how I think India has been navigating some of its data sovereignty.

 

First is, of course, while some of the thinking that has happened in our approach to data sovereignty lies in resistance to data colonialism, the use of users and our data being consumers of technologies and things that have been developed in the Global North, it's also in terms of attempting to shift consolidation of access, control, dependency on infrastructures that happen within the Global North. A good portion of the way that navigating data sovereignty is in trying to promote our own resilience and self-reliance being that underlying objective.

 

So India has sought to establish this kind of sovereignty at various levels, and not just at a regulatory sort of landscape, but also at infrastructural levels. And some of this has been mentioned in the context of building digital public infrastructure, having local data and local infrastructure, local data sets, indigenous tech. And this also is a good point to highlight how India's approach to data governance and sovereignty is so intrinsically linked right now to how India is thinking of its AI vision and its AI ambitions.

 

The second point on how I think India's been navigating some of this data sovereignty is this intersection with AI and the significance that data has in the AI ecosystem, it's shaping the way that we are even thinking of what it means to have data sovereignty and what it means to sort of operationalize our thinking on data sovereignty, while also assessing a balance with innovation. I think that it's displayed a little bit in how the most recent data protection rules have sort of not been prescriptive enough, not prescriptive at all, but allows space for India to exert control when required.

 

However, while India is trying to tow this line of various avenues for data governance, I think, to me, a critical question that needs to guide whatever approach we sort of toy with is who is benefiting from the approach. We need to keep that in mind when we're developing data governance strategies, data sovereignty strategies. And it's important to also note that when we're trying to address concentration of access and concentration of control from one set of actors, it shouldn't move to another set of actors.

 

This brings me to my final set of comments on what I'd like to speak about, just in terms of what kind of insights can we get from other jurisdictions. That's often what we do. We look to how maybe advanced economies or other jurisdictions in their journeys sort of deal with some of these issues. And first, I'd like to point out that India's data protection efforts have sort of already incorporated some of these learnings that have come from other jurisdictions. From other regulatory instruments, we've commonly adopted principles that come from, say, the EU and the landscape that operates there. We've also tried to build a balance between restrictions on data flows, learning from experiences that have come from compliance experiences across different jurisdictions.

 

Second, I think that while it's important to look at how different jurisdictions experiment and learn from the way that they operationalize their own strategies, we need to continuously also assess and monitor what it means for our own efforts as we implement our various approaches and various frameworks within the data governance landscape. We need to ensure that we are understanding what is working from what we are developing, what kinds of challenges that are cropping up, and make sure that in our response and as we continue to develop our data governance approaches that we respond with evidence-based approaches to some of the things that we're seeing.

 

And I say this because it's important for us to respond and develop these approaches based off of our context because otherwise we run the risk of importing mechanisms, importing thinking that come from other jurisdictions that are not necessarily suited to us. And my final thought on this is that when it comes to data governance and data sovereignty, data is part of a very complex socio-technical system and a significant anchor with data is the individual. And how data is governed and how it is used for other kinds of benefits, how it's used in other forms of technologies, how it fuels the AI ecosystem indirectly and directly impacts the individuals and different communities of individuals.

 

And so, to me, a key aspect as we navigate balancing data governance and data sovereignty is that we ensure that we are adequately protecting the individual, we're adequately protecting our data both within our country and even in our engagements outside with other countries. And I think that this emphasizes our commitment to data sovereignty and our data sovereignty and it ensures that we're also not just meeting standards that are set outside but also setting standards ourselves for other contexts. Thank you.

 

Rakesh Sood: Can we move to the last panelist? He's Mr. Mihir. Mr. Mihir Kulkarni.

 

Mihir Kulkarni: Hello. Hi. Am I audible? So just give me a minute. Hi. So, first of all, it's an honor to be invited for this forum with all the dignitaries and we really appreciate the opportunity here. So I was asked to focus on ethics and technical safeguards for India as we kind of move ahead on our AI journey, especially with the summit coming so close. Some background on myself, I work at this company called Vadhani AI that implements large-scale solutions for the public sector in India and I've been working on developing ML models and deploying these at scale, affecting millions of patients, primarily in tuberculosis and maternal health. I've been doing this for five years.

 

Before that, I was an undergrad at IIT Bombay and then I went to Princeton University to train in machine learning, where I also got to work with Arvind Narayanan to understand kind of the Western lens on ethics and responsible AI and kind of try and see how that translates to the Indian context. So, I think when you talk about ethics and AI, I think before we talk about technical solutions, the first thing you need to understand is kind of the social and the political landscape and the economic landscape, really, that we're in.

 

So, essentially, I think we need to understand how can we build AI for public good. And what that means is any solution you deploy will have to be beneficial for the end users or the people who are benefiting. For instance, in health, we want to make sure your AI solution makes the lives of ASHA workers, say, like ground-level TB officers, better, as well as improve patient outcomes. And oftentimes, the hype is conflated because a lot of the officials or the senior officials see it as a tool to make their efficiency kind of higher. And given all the hype, it's very attractive to kind of deploy this. But it's really important, I think, to foundationally and scientifically understand what AI is and what it does.

 

So, I think before I get into maybe the specific technical or like ethical aspects, one thing I would like to maybe set out as a foundation is some basic maybe concepts that are often, I think, ignored in the discourse around AI and ethics. The first thing is that any AI solution, any predictive model, needs to be backed by some scientific evidence. What that means is that you could have a dataset and some labels that correlate with each other, but that does not necessarily mean that this is something that you can deploy, especially in the public sector and affect lives.

 

For instance, we see lots of tools or like we see several tools promoting, for instance, AI for astrology or AI for like sort of dodgy concepts like personality detection from faces that are kind of known to be kind of pseudoscientific and not based in scientific reality. And in this hype, I think it's important to first understand what problems AI can and can't solve. So, to build any predictive model, you need to be able to have some scientific basis to know that the data can predict the label. And so, if you ever see an AI solution, and I would also hope that this also makes it to kind of the policy discourse sometime, how do you ensure that there is some kind of regulation around certain applications that are pseudoscientific?

 

I think even if the applications can kind of be activated, it's important to kind of think about as a society what we're doing. For instance, there are lots of use for AI in surveillance, a kind of automated weaponry, and it's important for us to think about whether these use cases are something we want to kind of promote as a society.

 

Then I think the next thing to keep in mind is the fact that these are fundamentally probabilistic tools. And so, you can never have 100% accuracy with an AI tool. Even say you have the best safeguards, or you have any sort of responsible AI guardrails, the fact that it's probabilistic always means that there is some probability of failure. That's often okay because it's still better than what currently exists, but the fundamental probabilistic nature of AI is something we have to keep in mind while developing the solution.

 

And finally, I think it's important to understand how these tools actually work on the ground. And my experience is mostly in public health, so I can talk about that. If you look at the maternal health situation, the ASHA worker has I think like 10 plus apps these days to do data entry and to kind of log everything. And there are all sorts of AI tools that are proposed to help their lives, but it's important to kind of see how your tool fits into the broader ecosystem.

 

So as a technologist working in the field, one thing that I've seen time and again, there's a lot of good intent, both from administrators and from kind of other people on the ground to use technology and the potential is very clear. We have so much data, we have limited resources, we can use the data to make smart decisions, but actually actioning that on the ground means you to keep a lot of these like basic concepts in mind to make sure that they're actually helping people, both the workers and the patients, and not just kind of having something imposed from the top.

 

And finally, I would say it's important to have your sovereign AI, as important to kind of build things at home and from scratch and kind of own the technology you build. But it's, I would say, more important to make sure that it's used for actually improving service delivery, rather than kind of just doing something that you impose with hype. I would say in terms of the recent kind of technical, or kind of ethical ideas that have been in the works, we know that the material is the AI governance guideline, the number 2025. I think that it's quite clear that the idea is more like innovation with safeguards rather than being super precautionary.

 

So the governance guidelines wants to promote innovation while ensuring that we have basic safeguards in place. And then in my opinion, this puts a heavier burden on the technologists and the developers to understand the situation, understand the context therein, to be able to develop safe and responsible AI. I'm going to describe like a few maybe technical solutions that we think about when we think about ethical AI. There are also some resources, I know that NASSCOM has published some responsible AI kits. Ikegai Law recently released a handbook for AI developers that we've also contributed to, where they've talked about a bunch of these things. And I'll try to describe a few of these in some details.

 

I think one thing before we get into the models, we can talk about data. So DPDP, of course, has been fairly comprehensive in trying to describe how we should govern data and AI developers. Because myself often ran into issues with that, for instance, how do you implement notice and consent while having AI models? Like, is it possible to take consent from a bunch of people who've had your health records over like years and years? It's hard to kind of comply with that. And for instance, there are proposed solutions. For instance, you could have like a stewardship model where there's a data steward that manages the consent for a group of people.

 

There are other models as well. But essentially ensuring that you'd bring data as well as like you're able to ensure the protection of individuals while ensuring that innovation can reach them is something that we need to think about. It's also important to ensure that we have the appropriate data lineage and provenance when we are creating datasets. So any dataset used for building models, it's important to understand where they come from, or the history of the data, and kind of the quality in which it's been collected.

 

Indian datasets in particular are known to have certain issues. For instance, it's known that a certain class of people, like for instance, like women, or like certain underprivileged communities, appear less in the data. They have less representation. The quality of the data collection, for instance, in poorer regions is lower quality. And therefore, any models you build on them are not going to be as good. And so it's very important to fix your data pipeline. And like that also necessitates creating that infrastructure to collect data, to be able to build models on top of them. I understand this is perhaps a very public sector, a very kind of heavy underprivileged kind of view. But I think it applies to any kind of solution of AI you build in India.

 

Getting into the models, I think there are a few things to keep in mind. There's the idea of like algorithmic fairness, to make sure that your models do well on the cohorts you care about. Within fairness, I think it's important to first understand the concepts of fairness in India, and also to understand what cohorts matter. Most fairness frameworks we have these days come from the West, but we have certain Western ideas of fairness. For instance, there's equity of opportunity. There are like different kinds of ways you can define fairness. And the Indian ideas are not always aligned with them.

 

So for instance, like one maybe more legal example is like the U.S. has affirmative action, we have reservation. Similarly, when you have like an algorithmic model, how do you sort of define the right fairness criteria to deploy this in the field or differ? We also have different axis of fairness. So in the U.S., you have maybe race and gender and things like this. In India, we also have religion, caste and so on. And so there are other kinds of axis to keep in mind when you're defining fairness for Indian cohorts. And oftentimes they're not like directly available in the data. So it's important to think about how you're representing these cohorts in your data by your training models.

 

There are technical solutions to ensure fairness, but sometimes you just have to collect more data and do the work to make sure that you're able to collect these appropriate data sets. So more data scarcity, as well as bias in the data itself are kind of important. So if you don't have the data, it's a problem, but also if the data is not as accurate for certain communities, you're also going to build models that are biased against them. So it's important to kind of ensure that you have a protocol in place to take care of this. There are human loop solutions you can have for this. You can do some surveys, other ways to kind of fix this issue. But it's important to, I think, define what fairness means and what's causing issues in the models being unfair before you kind of deploy them.

 

The other thing I want to talk about within the models is interpretability. So a lot of these models, especially in the public sector, are important. Essentially, they have consequences on life, right? So, for instance, choosing a model to determine which person gets a health intervention, which person doesn't, this has consequences on people's health outcomes. And therefore it's important to both understand the models and also have accountability. So if the model makes a mistake, who's at fault? Recourse mechanism.

 

So, for instance, in public delivery, we have like a grievance officer who handle these grievances and there are ways to hold people accountable. But when it's a model, what is the solution? How do you hold the model accountable? Like, who is at fault if the model is wrong? So kind of figuring out this issue, and also understanding why certain decision was made. Machine learning interpretability as a science is not as advanced as we'd like, especially for the most advanced model, like neural networks, they're not as interpretable as we'd like them to be.

 

And so it's important especially when you have public consequence, it's important to make sure that you have the right safeguards in place to technically evaluate what happened. We also have other technical solutions, the idea of confidential computing that lets you run completely in encrypted world. So you can basically run models without the data ever being decrypted. And so you can achieve privacy by training your models in these kind of worlds, the idea of differential privacy, where you can kind of trade off accuracy with privacy. So you kind of use less private data by slightly trading of accuracy to be able to maintain patient confidentiality and still achieve good performance.

 

In terms of regularity guardrails, technically, I like the idea of regularity sandboxes. And so there's like maybe a really accepted benchmark, other sandboxes with like benchmarks within it, where you can deploy these models. And the results of this kind of regulatory sandbox can then be made public. And so this helps you test it effectively before you actually deploy in a way that the public can also access. And it will also be helpful, I think, to have like natural frameworks, for instance, for benchmarking, ensuring that you have even independent data sets to be able to make sure that any model deployed for public use is kind of available to the public in terms of its performance.

 

And finally, I would say, in terms of governance, I think it's important to also have participation from the people who are using the model. So there's some work, for instance, by Vidhi NIIT Madras, the paper, they talk about participatory governance in the Indian context. But there are already kind of models for this in India, in the digital space before AI, which I'm sure people here are aware of.

 

I'd like to finally conclude with the final part of the deployment, which is maintenance. So when you deploy a model, there is a maintenance phase that you have, you kind of track metrics, just to make sure you understand what's going on. Oftentimes, in India, when you deploy in the public sector, things change on the ground very fast, the data isn't absorbed as fast enough to detect the change. So how do you do data detection? How do you make sure that you're tracking the right metrics? How do you make sure that you actually understand something on the ground, and the developers also are aware of that, to be able to change the model or make changes if something is going off, especially in places where data is slow to come, it's really important to kind of think this through. Because once you deploy, it's still kind of live.

 

And so it's important to be able to act quickly, and to set up pipelines that help you do that. One way to do this is to have a human in the loop. So, for instance, you can have a human make the final decision and not let the model finally decide. Because of the model flag, maybe doubtful cases to the human to make sure that they're able to handle the edge cases. For instance, for a Kisan chatbot there are LLMs that like help do farmer advisories. Oftentimes, like what we've done is that we also…

 

Rakesh Sood: I think you'll need to wind up now very quickly because we're already running out of time.

 

Mihir Kulkarni: Sure, I'm done. So finally, I would say, maybe the final kind of points that I have to make is we like to understand the landscape in India. So the government guidelines and the handbooks that we have, make sure you understand the data and the situation really well, and make sure the user has some way to kind of give feedback and kind of understand what the model is doing to be able to act and improve on it. So I think we have the opportunity to show the world that.

 

Rakesh Sood: Thank you. I think in the interest of time, it's best that we open up for any questions. I assume that all of you have now learned a lot about the Indian situation of AI and are much better informed. While we still have any questions for the panelists, please feel free to ask your questions. Keep them short, because we really have very little time. Yeah.

 

Sachin Yadav: Good evening to all the panelists. My name is Sachin Yadav. I'm a PhD scholar at Jamia and a research intern at ICWA. My question is to Joanne, ma'am. Ma'am, last year in September, I presented a paper in Jamia Millia Islamia. And the main arguments of my paper was that the major and main data centers for AI are situated in the U.S., and there is no proper legal framework for AI still there. Now, what I want to ask is, how much India has progressed in terms of legal frameworks for AI and digital?

 

Zeeshan Ali: Thank you so much to the panelists. My name is Ishan. I'm a research intern at the council. My question is to Anulekha, ma'am. How can India bridge the AI gap, given that AI is projected to contribute around $16 trillion to our global GDP by 2030, yet approximately 85% of its gains are expected to go mostly to Europe, U.S., or China, and India faces low internet penetration, 51% AI talent gap, and GPU shortages?

 

Azra Shahab: Good evening, everyone. My name is Azra. I'm a research intern at ICWA. My question is directed to Joe and ma'am. Data forms the backbone of AI, and we have been witnessing a unidirectional flow of massive data from Global South to Global North. For example, Facebook has most users in India, but the data centers are located in America and Europe, as Sachin said. How can India secure its digital sovereignty and ensure its data advances equitable development and governance? Thank you.

 

Joanne D'Cunha: Hi, and thanks for the question. Your question was how can India bridge the AI gap, right, in light of its talent and GPU shortages. So on talent, it's a little more nuanced. India has a lot of talent when it comes to low-end AI skills, but when it comes to advanced AI skills in terms of developed mathematical statistical models, like actually developing the models, that's where the actual talent gap lies. So it has a lot of AI talent in terms of developing AI applications, but kind of doing the core AI innovation is where perhaps this 51% gap you're talking about lies.

 

In terms of the GPU shortage, India is one of many in the world, because obviously U.S. controls the upper end of the value chain in terms of semiconductors, as well as influences a lot of the semiconductor policies. And given its dominance of the upper ends of the value chain, can then implement choke point strategies, which it has, particularly in the case with China. So obviously India has a semiconductor policy, but it's still a long way to go. And I think in scenarios like GPUs, which are a scarce economic resource, it's not like models that you can develop. It does benefit to kind of understand which countries do we need to partner with.

 

We already have an inflow of private investment, particularly from companies like Micron. We have the fabrication plants coming up, and so on. So how do we kind of enhance our equipment manufacturing? Because that is a critical ecosystem input that we need when it comes to building up the compute ecosystem. So I think GPU shortages is something that India is grappling with in tandem with most of the world. And while, again, similar to its AI strategy, while it is building up its kind of production capabilities, its core sovereign capabilities, in the short to medium term, it does help to think about how it can partner, what the modalities of such partnership would be. The Tata PSMC plant is a licensing partnership.

 

So how it can kind of license, import, take advantage of the technologies, how does it kind of partner with the countries that have them. So I think in the short to medium term, that is the strategy. I think the long-term strategy is obviously developing the sovereign capabilities, which it's doing through the semiconductor mission. I think I just have one question.

 

Anulekha Nandi: Thank you for your questions. I'm going to keep it brief, because I know that we're running short of time. One was a question on progress in terms of how we're thinking of the AI regulatory landscape. And if I'm getting it right, it's about equitable development of data, and how do we sort of bridge the gap to the fact that the Global South has been a huge producer of data, and then it being situated in the Global North. So just to respond very quickly to the regulatory framework and the landscape that we are situated in right now, India has been, for a while now, toying with the idea of whether to regulate AI or not.

 

And I think it's fairly evident with the fact that it has been more inclined to put out strategies and guidelines that help industry and help different stakeholders grapple with how to develop their AI. But we're still not concretizing any sort of legislative framework. And this has come from a little bit of learning with the data protection law, and how long that sort of journey has also taken. And also learnings from other jurisdictions that have already established their AI frameworks. With the rapid development of AI, how do you develop a regulatory landscape that captures what is happening now, and then what captures the future as well.

 

So we have to be very careful about how we develop it. That doesn't mean that it's not required. Which brings me to the point of, while we're figuring out what it means for the AI regulatory landscape, it becomes all the more important to zero in and focus on the data protection landscape that we have in place. So that's where the focus should lie.

 

To respond to how the Global South needs to sort of meet the Global North at its number of data sets, and just the fact that the data is within the Global North. And there's a whole discipline that explores what it means for data to be, for the Global South to be producers and consumers of different kinds of technology, and is really situated in how do we navigate data colonialism. It's an active area of research. But like my co-panelists said, a lot of how we address those issues come from building our own infrastructures, from building our own data sets, and it's something that the Indian thinking and landscape has already started to do. Very briefly, this is how I can quickly respond to this, but thank you for your questions.

 

Rakesh Sood: We can take another round of questions, yes.

 

Unidentified Participant: Good evening esteemed panel. My name is Ziaul and my question is where does India stand realistically in shaping global AI governance and what step does it need to take to turn its AI leadership ambition into reality? Thank you.

 

Anulekha Nandi: I think India is still thinking through in terms of what its AI governance looks like in the country, and there are kind of two documents that outlines that vision. But like I said, going forward, when you think of global AI governance, how that stance translates to norms, policies, and principles that are applicable to countries and contextual realities globally is the legwork that we have to do.

 

In terms of standards, in terms of norms, and in terms of frameworks that are acceptable to countries, like we are proposing a techno-legal solution or technical modes of AI governance. But what do those technical governance look like? Do we have tools that we can offer, or is it a suggestion? There are also practical realities in terms of how do you do technical modes of governance. Do we need to get into a conversation about AI audits and compliances? How do we develop then those auditing standards?

 

So it's a long conversation that needs to begin, and I think in terms of understanding the areas of contribution, we need to understand how we can contribute. So like, for example, China is hugely influential in the standard-setting bodies, is that a way we would want to participate in the global governance regimes? So kind of identifying what our fora are is very important, and then kind of understanding what we prefer in terms of solutions, tools, and frameworks.

 

Rakesh Sood: Well, let me just conclude this session now. I'm afraid we don't really have a leadership position and I would say that this is very clear. We have a certain amount of data, but data by itself means nothing because it is the quality of data that counts. All the LLMs that have been made are not based on Indian data and they are working fine. So let us not get carried away with this data business. We do not have compute, we are trying to get some, we are trying to make it available to people. We will see how that process continues forward.

 

We are trying to develop our own models. These are much smaller models compared to the models that Google, that Gemini and ChatGPT, the OpenAI models that are coming up or even Anthropic is coming up with. So I think this is a far cry from where we can assume leadership, number one.

 

Number two, the current AI fascination has begun in 2022 November when OpenAI released its first ChatGPT-3 or whatever. Now we have seen that their approach is based on large language models, the more the data you can throw at it, which means that they can put in more, it consumes more power, it needs a bigger data centre, it needs more GPUs and things like that. Today itself, there is a huge question mark about whether LLMs are going to be the answer. We have already seen that in so far as the Western LLMs are concerned or the American LLMs are concerned, the Chinese when DeepSeek came out with R1, they showed that you could achieve a similar degree of integrity and fidelity with much less resources.

 

So tomorrow there is going to be a discussion about that. And that is why people are moving towards reasoning models, to be able to say that, look, this is how a model arrived at a particular solution, to reason it out, to be able to explain with conviction that it is, "intelligent." Now is that going to be enough or not? Because that itself is also consuming electricity and relying on GPUs. Meanwhile, what is happening is that there is an incestuous kind of a circular relationship that has grown up between NVIDIA, which dominates the GPU models and the GPU designs, and companies that use them.

 

And now this is being challenged because Google is coming out, I mean, if the new NVIDIA model which they have unveiled and which is only going to become available sometime in August or September, called the Vera Rubin, is going to cost something like about $40,000 a chip, Google is coming out with their TPUs, which is costing about $4,000. Similarly, Microsoft is working on its own chip, which doesn't necessarily, will not necessarily need the same kind of a graphic processing unit that NVIDIA is trying to market. I mean, NVIDIA, because it dominates or it has dominated the market for GPUs, has seen its valuations climb to whatever astronomical levels, but that may not remain the case.

 

Finally, if India has to go down the app route, then as one of the speakers mentioned, Dr. Agarwal, as he mentioned, you will have to look at more affordable models. He gave the example of something for a farmer in India. Now, the farmer in India, if he needs a model, he needs a model which serves his limited purpose. He does not need that model to help him plan his vacation and do his hotel bookings and airline bookings and all of that, and also organize his theater tickets while he is on vacation, and so on.

 

So what we are looking at is very different kind of requirements. And so that would mean much, much smaller models, and not necessarily the same route. So I think that unless we figure out how we want to proceed in this direction, it would be very difficult for us to put forward models for the rest of the world to assume leadership.

 

At most, as I see it, this particular summit that is going to take place in Delhi will basically, from what I have been able to understand and talk to people involved with it, is that what we are trying to do is we are trying to call it an impact summit, and that will focus on showing the potential for an impact in a developing society using affordable models, using small data sets, using economies, economically prudent and frugal approaches, higher efficiencies in terms of applicability, which will also lead to greater acceptance, and so on. I think if we can achieve that, I would say that we had a successful summit. So with that, let me once again thank ICWA for having organised this and bring this session to a close. Thank you.

 

Unidentified Speaker: I take this opportunity to once again extend my heartfelt thanks to our chair and esteemed panelists for their insightful remarks. We have all gained immensely from your perspectives. I would like to express my gratitude to my colleagues and a special thanks to Director Research Dr. Nivedita Ray. With that, I warmly invite you all to join us for high tea in the foyer. Thank you once again and have a wonderful evening.

*****

List of Participants

1. Ambassador Rakesh Sood, Former Indian Diplomat (Chair)

2. Dr. Anulekha Nandi, Independent Researcher and Consultant previously with ORF

3. Ms. Joanne D'Cunha, Programme Manager with the Technology and Society team at Centre for Communication Governance at NLU Delhi

4. Mr. Mihir Kulkarni, AI Scientist and team lead at Wadhwani AI (online)

5. Dr. Nikhil Agarwal, Managing Director of the Foundation for Innovation and Technology Transfer (FITT) at IIT Delhi