AI for Good Technology That Empowers People

20 Feb 2026 10:00h - 11:00h

Session at a glance

Summary

This discussion focused on Edge AI applications and their potential for development in the Global South, hosted as part of the ITU’s AI for Good initiative. Frederick Werner from ITU opened by emphasizing the transformative potential of AI and the need to ensure it serves humanity beneficially, introducing AI for Good as a year-round global movement organized around solutions, skills, and standards.

Professor Brijesh Lall from IIT Delhi presented research on haptic applications requiring edge computing due to their need for ultra-low latency and real-time processing. He explained how edge AI enables the convergence of communication, computing, and control, particularly important for tactile applications where delayed feedback can be problematic. His team’s work includes split control architectures and intent-based signal processing to improve interoperability between different haptic devices.

Dr. Ranjitha Prasad discussed federated learning as a key enabler for edge AI in telecommunications networks, addressing privacy concerns while enabling distributed intelligence. She explained how federated learning allows AI training to occur at the edge while preserving user privacy, supporting applications like traffic prediction and vehicle-to-everything (V2X) communications. This approach enables personalization of AI models in real-time while maintaining data sovereignty.

The panel discussion featured practical examples of edge AI deployment across different regions. Mala Kumar described XR applications for cultural heritage and medical emergency response using both public and private 5G networks. Alagan Mahalingam shared experiences developing AI solutions for farmers in Portugal and Sri Lanka, highlighting how edge computing enables AI functionality in areas with limited connectivity. Sakshi Gupta from Qualcomm emphasized the importance of on-device AI capabilities and distributed architectures for addressing Global South challenges.

The session concluded with remarks from UN ambassadors who co-chair the Global Dialogue on AI Governance, emphasizing the need for human-centered AI development, international cooperation, and practical solutions that can be implemented on the ground to bridge digital divides.

Keypoints

Major Discussion Points:

– AI for Good Initiative and Edge AI Focus: The discussion centered around ITU’s AI for Good summit and movement, emphasizing the shift from cloud-based AI to edge computing solutions that bring AI capabilities closer to where data is generated and decisions need to be made, particularly for applications requiring sub-10 millisecond latency.

– Real-world Edge AI Applications: Multiple speakers presented practical implementations including XR-assisted medical emergency care, haptic technology for tactile applications, federated learning for telecom networks, agricultural AI solutions for farmers in Portugal and Sri Lanka, and on-device healthcare assistants that work offline.

– Technical Challenges and Solutions: The conversation explored key technical aspects such as federated learning for privacy-preserving distributed intelligence, the convergence of communication-compute-control systems, AI-native network architectures, and the need for context-specific AI models rather than universal foundation models.

– Global South Development Strategy: Panelists emphasized the importance of edge AI for bridging digital divides in developing regions, addressing connectivity limitations, and enabling AI deployment in areas with unreliable internet infrastructure through solutions like mobile computing units and offline-capable devices.

– Standards and Governance Framework: The discussion highlighted ongoing standardization efforts through ITU, TSDSI, and other organizations, concluding with information about the upcoming UN Global Dialogue on AI Governance aimed at creating inclusive, practical approaches to AI development and deployment.

Overall Purpose:

The discussion aimed to showcase practical edge AI implementations in the Global South, demonstrate how edge computing can make AI more accessible and effective for underserved communities, and promote collaboration between researchers, industry, and international organizations to advance AI for good initiatives.

Overall Tone:

The tone was consistently optimistic and collaborative throughout, with speakers demonstrating genuine enthusiasm for solving real-world problems through edge AI. The atmosphere was professional yet accessible, with technical presentations balanced by practical examples and success stories. The discussion maintained a solution-oriented focus, emphasizing cooperation and knowledge sharing rather than competitive positioning, and concluded on an encouraging note with calls for continued international collaboration.

Speakers

Speakers from the provided list:

– Vishnu Ram OV – Session moderator/host

– Frederick Werner – Chief of Strategic Engagement Department at ITU (International Telecommunication Union)

– Professor Brijesh Lall – Former Bharti School chairman, currently focusing on edge AI research at IIT Delhi, involved in ITU activities including AI for Good Challenges, WTSA Hackathon judge, and Kaleidoscope

– Ranjitha Prasad – PhD from ISE, researcher specializing in causal inference, survival analysis, Bayesian neural networks, federated learning, and other learning paradigms; PI of Intellicom Lab at IIIT Delhi

– Mala Kumar – Technologist at Center of Excellence Wired and Wireless Technologies at Art Park, former postdoctoral researcher at Teikian Group at Technical University Berlin, involved in 6G initiatives for AI RAN and millimeter wave communications

– Alagan Mahalingam – Founder, CEO, and Chief Software Architect of RootCode; ICT Entrepreneur of the Year (2021), Young Entrepreneur of the Year (2024), envoy for Estonia e-residency program

– Qualcomm Member (identified as Sakshi Gupta in introduction) – Global Government Affairs responsible for Qualcomm, tech policy professional specializing in AI and emerging technology policy analysis, market research, and stakeholder engagement

– Egriselda López – Ambassador, Permanent Representative of the Republic of El Salvador to United Nations Office and other international organizations; co-chair of the Global Dialogue on AI Governance

– Reintam Saar – Ambassador and co-chair of the Global Dialogue on AI Governance

Additional speakers:

None identified beyond the provided speakers names list.

Full session report

This comprehensive discussion on Edge AI applications and their transformative potential for development in the Global South was hosted as part of the ITU’s AI for Good initiative, bringing together researchers, industry leaders, and UN officials to explore practical implementations and governance frameworks for distributed artificial intelligence systems.

Opening Framework and AI for Good Initiative

Frederick Werner from ITU opened the session with a profound philosophical question that set the tone for the entire discussion: “What if the last thing that humans ever invent is invention itself?” This existential framing, inspired by AI safety expert Roman Yampolsky’s observation that AI might be the final human invention, established the stakes for ensuring AI development serves humanity beneficially. Werner emphasised the wordplay that AI must be “for good” in both senses – beneficial and permanent.

The AI for Good initiative, launched in 2017, has evolved from a concept-focused summit addressing the “fear, promise, and hype” of AI into a practical year-round global movement. Werner explained that whilst early iterations featured presentations with limited substance, the initiative now encompasses real-world applications spanning from generative AI in 2023 to AI agents in 2024. The initiative operates through three core pillars: solutions, skills, and standards development.

When technical difficulties arose with the video presentation, Werner humorously remarked, “AI is easy, AV is difficult,” highlighting the irony of discussing advanced AI while struggling with basic audiovisual technology.

Technical Foundations and Haptic Applications

Professor Brijesh Lall from IIT Delhi presented research on haptic applications that exemplify why edge computing is essential for certain AI implementations. His work focuses on the convergence of communication, computing, and control systems, particularly for applications requiring sub-10 millisecond response times where delays can be catastrophic.

Lall’s research encompasses two main areas of haptic computing: kinesthetic feedback (related to movement and position) and tactile feedback (related to touch and texture). He explained that these applications require robust edge capabilities because cloud-based processing introduces unacceptable latency for real-time haptic interactions.

The professor emphasised that whilst foundation models attempting to solve universal problems receive significant attention, context-specific solutions are increasingly important, particularly for Global South applications. He argued that edge capabilities enable better leverage of local context, making smaller, targeted solutions potentially more valuable than large universal models for addressing real-world challenges in resource-constrained environments.

Federated Learning and Telecommunications Integration

Dr. Ranjitha Prasad from the Intellicom Lab at IIIT Delhi provided insights into federated learning as a key enabler for edge AI in telecommunications networks. Her research, funded by METI and the India AI Initiative, addresses the critical challenge of data explosion whilst preserving privacy.

Ranjitha explained how federated learning enables AI training across distributed devices without centralising sensitive data. This approach is particularly valuable for telecommunications applications where privacy concerns and bandwidth limitations make traditional centralised training impractical. Her work demonstrates how local devices can contribute to model training while keeping data localised, enabling AI capabilities even in areas with limited connectivity.

The integration of federated learning with edge computing creates opportunities for developing countries to participate in AI development whilst maintaining data sovereignty and addressing privacy concerns that might otherwise limit AI adoption.

Real-World Implementation and Global Deployment

The panel discussion revealed diverse practical applications across different geographical contexts. Mala Kumar presented XR (Extended Reality) applications demonstrating the convergence of edge computing with immersive technologies. Her work includes cultural heritage experiences that provide museum visitors with immersive content in regional languages, making cultural artifacts more accessible and engaging for diverse audiences.

Perhaps more significantly, Mala described XR-assisted medical emergency care systems for cardiac arrest situations. When emergencies occur, first responders arrive equipped with XR glasses and automated external defibrillator (AED) kits. During CPR administration, the XR system provides real-time guidance and vital sign monitoring, whilst medical experts can provide remote assistance through the immersive interface. This system demonstrates how edge AI can enable life-saving interventions through real-time decision support whilst maintaining the ultra-low latency requirements critical for emergency response.

Alagan Mahalingam shared extensive experience developing AI solutions across multiple countries, serving millions of users through his company RootCode. His agricultural AI project initially succeeded in Portugal, providing soil nutrition analysis and plant disease detection through mobile applications. However, when adapted for Sri Lankan mountain villages, the solution faced significant challenges due to unreliable connectivity.

This experience led Mahalingam to develop innovative edge computing solutions using optimised models and local processing capabilities. His approach emphasises working backwards from specific agricultural tasks rather than starting with general-purpose models, arguing that agricultural AI assistants should focus exclusively on plants and agriculture rather than possessing broad general knowledge. He developed model optimisation techniques to create smaller, task-specific versions that maintain effectiveness whilst operating within edge device constraints.

Industry Perspectives and Device Integration

Sakshi Gupta from Qualcomm provided industry insights into the expanding availability of edge AI capabilities across consumer and commercial devices. She emphasised that edge AI deployment requires thinking about distributed architectures rather than simple cloud-versus-edge dichotomies.

Qualcomm’s developments demonstrate the rapid expansion of on-device AI capabilities across smartphones, automotive applications, and IoT devices including smart glasses. These developments address key Global South requirements including reduced power consumption, lower costs, enhanced privacy, and decreased dependency on reliable internet connectivity.

Sakshi briefly mentioned Qualcomm’s Tech for Good programme, which partners with startups globally to develop edge AI solutions for underserved markets, though specific details about partnerships and implementations were not elaborated during the session.

Standards Development and International Cooperation

Werner outlined ITU’s approach to AI standardisation, with standards development covering future networks, 5G/6G technologies, and AI-native network architectures. This work includes pre-standardisation efforts that will enable the seamless integration of artificial intelligence into telecommunications infrastructure.

Professor Lall highlighted collaborative efforts between TSDSI (India’s standards development organisation) and ITU on technical reports covering AI/ML applications for various use cases. These standards development efforts are crucial for ensuring interoperability and enabling global deployment of edge AI solutions.

The importance of open-source approaches emerged as a theme, with Mala Kumar advocating for making AI for Good solutions available through platforms like ITU to enable international community access and collaborative improvement of solutions across different contexts and regions.

Governance Framework and Global Dialogue

The session concluded with presentations from UN ambassadors co-chairing the Global Dialogue on AI Governance. Ambassador Egriselda López from El Salvador, based in New York, emphasised three key principles emerging from global consultations: people must remain at the centre of AI development; closing digital gaps requires decisive support rather than mere rhetoric; and avoiding fragmented approaches through coordinated international cooperation.

Ambassador Reintam Saar from Estonia outlined the structure and objectives of the first Global Dialogue on AI Governance, scheduled for July this year in Geneva. The dialogue aims to bring together governments and stakeholders for practical, results-oriented discussions that focus on capacity building, particularly for Global South countries. The approach emphasises inclusive participation, transparency, and human rights foundations whilst seeking actionable insights rather than theoretical discussions.

Key Insights and Future Directions

The discussion revealed that edge AI is not merely a fallback solution for areas with poor connectivity, but rather enables new categories of applications requiring real-time processing, privacy preservation, and contextual adaptation. The emphasis on task-specific optimisation over general-purpose models suggests that developing countries might achieve better outcomes by focusing on targeted solutions addressing specific local challenges.

The convergence of technical innovations with governance frameworks creates opportunities for coordinated global action that could accelerate edge AI adoption whilst ensuring equitable benefits distribution. The session’s emphasis on working backwards from specific use cases rather than starting with general-purpose models provides a practical philosophy for edge AI development.

Several challenges remain, including the need for specific metrics and evaluation frameworks for measuring edge AI deployment success, standardisation for interoperability between different systems, and systematic approaches to infrastructure development and financing for sustainable deployment at scale.

Conclusion

The session concluded with ceremonial elements, as Vishnu Ram OV mentioned arrangements for photos and mementos presentation, highlighting the collaborative spirit of the international gathering. The discussion demonstrated that edge AI represents a crucial technology for Global South development, enabling AI functionality in areas with limited connectivity whilst providing benefits in speed, cost, privacy, and personalisation.

The upcoming Global Dialogue on AI Governance provides an opportunity to integrate these technical insights with broader policy frameworks, potentially creating coordinated international approaches that could accelerate beneficial AI deployment whilst addressing governance challenges. The combination of technical innovation, practical implementation experience, and governance framework development suggests that edge AI could play a transformative role in global development if properly coordinated and implemented.

Session transcript

Vishnu Ram OV

Thank you. Thank you very much. We have very little time, so I want to first of all introduce Fred. Fred Werner is the Chief of Strategic Engagement Department at ITU Welcome Fred to give the opening remarks

Frederick Werner

Hello Let me start with a question What if the last thing that humans ever invent is invention itself? Now what do I mean by this? We had, if you’re familiar with Roman Yampolsky He’s a leading AI safety expert And I met him in New York at the UNGA last fall And he said, Fred, what is AI for good? I said, well, what do you mean? He said, well, is it for good or for good? Well, what do you mean? And he said, well, for good as in beneficial, as in good Or as in for good, forever I said, hmm, good point And he said, well, for good, for good And he said, what if AI is the last thing that humans ever invent?

Now, you might agree or disagree with that statement, but it’s not hard to imagine a future where most future inventions will either be invented by an AI or with the help of an AI. And if that is the case, then I think we do need to make sure that AI, if it’s going to be for good, is indeed for good. So my name’s Fred Werner from the ITU. It’s the United Nations Specialized Agency for Digital Technologies, and we’re also the organizers of AI for Good with 50 -plus UN sister agencies. Now, AI for Good was created in 2017. And if you think about that, that’s basically an eternity in terms of AI years, looking at how fast it’s been developing.

And back then, it was really all about the fear and the promise and the hype of AI. Most solutions existed in fancy PowerPoint slides, but there wasn’t a whole lot of substance. But that changed rather quickly. In 2023, we saw the advent of generative AI. Last year, the unofficial theme of the summit was the rise of the AI agents. And now we’re looking at a world where you’re basically entering a zero -click world where agents are not waiting for our prompts. They’re actually acting on our behalf. And in addition, you have the physical embodiment of AI in the form of robotics, embodied AI, brain -computer interfaces, and we’re even looking at space AI computing now.

Now, so I think we’re safe to say there’s no shortage of high -potential AI use cases that can be used to help solve global challenges. Anything from affordable health care to education for all, food security, disaster response, the use cases are definitely there. So what is the goal of AI for Good? Well, simply put, it’s to unlock AI’s potential to serve humanity. And how do we do this? Well, first of all, we can’t do this alone. Nobody can. That’s why we have AI. We have 50 UN sister agencies as partners of AI for Good, contributing knowledge, sharing expertise, helping to drive our standards work. building cooperation around AI governance. And we’re very privileged to have here the two co -chairs and facilitators of the UN AI Global Dialogue who will be doing the closing remarks.

Now, I could talk about AI for Good for days, but to save us some time, I just want to show you a little video so you can actually see AI for Good in action from our last summit. If we could please play the video. I have a joke that I always say for these occasions. AI is easy. AV is difficult. Actually, we don’t need to see the video. Oh, ah. Is it going to happen? Yes. But now we need sound. Since there’s no sound, that’s lovely, Geneva. Aha. We are more than the AI generation. We are the generation that is determined, ladies and gentlemen, determined to shape AI for good. So no matter how fast technology moves, let us never stop putting AI at the service of all people and our planet.

If you want an AI literate society, meaning resilient and ready for the future, we need to integrate these new tools into schools, curricula. Let’s build a future where AI advances progress for all humanity. A shared digital future that is again inclusive, equitable, prosperous and sustainable for all. It is no coincidence that this era of profound innovation has prompted many to reflect on what it means to be human and on humanity’s role in the world. AI must help bring us closer, not to divide us apart. That’s one of the foundational promises of AI for good. We all now have, I think, a much greater level of awareness around AI, and we all need to shift into that as fast as possible because this technology is moving so fast.

Ladies and gentlemen, this was a real… fast -track operation that we did, what we call the International AI Standards Exchange Database. standards in your domain or industry that require this type of trigger. And we have just started the last step right from the general division. Let’s go! I think it’s fair to say that AI for Good is indeed more than a summit. It’s a movement, it’s a global community, and it would be nothing without you, the participants. Three, two, one. Thanks for watching. I’m not sure who that last guy was. Now I think one of our… I think people often misunderstand that AI for Good, it’s known as a summit that takes place each year in Geneva.

But it’s actually a year -long activity. We have online events almost every day of the week, all year long. And we’re organized around three pillars. Solutions, skills, and standards. And if you look at the solutions pillar, we have machine learning challenges, we have startup pitching competitions, all types of activities to identify real practical applications of AI that you can use here and today. And on the topic of Edge AI, we had a build -a -thon on Edge AI just a few weeks ago here in India. And we also had machine learning challenges on tiny ML, tiny machine learning devices. And when we’re looking at skills, we launched the AI Skills Coalition. And a big piece of that is going to be creating basically machine learning environment sandboxes where we can do training and mentoring for governments to upskill their constituencies on the use of AI using the data from our machine learning challenges.

So it’s not hypothetical. It’s using real data for real solutions. And the last piece, of course, the bread and butter of ITU is standards. And we have over 400 AI standards published or in development covering a whole suite of topics. But more specifically related to the session, we have a standards work on future networks, basically 5G, 6G and beyond, and a pre -standardization effort on AI native networks. So basically, these are examples of AI for good in action. And the theme of this session is actually edge AI in action in the global south. And I’m very much looking forward to the discussion and thank you for your time and attention.

Vishnu Ram OV

Thank you so much, Fred. Now, we have the keynotes coming. Thank you. First of all, let me call Professor Lal. Brijesh is my great friend as well as colleague. He was the Bharti School chairman, but also right now he is currently looking at edge AI research. Our touch points with ITU are many, where he’s hosted AI for Good Challenges, WTSA Hackathon. He was a judge, as well as Kaleidoscope. He’s very active. Thank you very much, Vijay Singh, for coming, and over to you.

Professor Brijesh Lall

So it’s been a while. Thank you, Vishnuji, for having me. I’ve been participating in these AI for Good activities, so there’s been a lot happening, not just these talks that you have, but also something on the ground. Hackathon is an example of that, with participation from all over the globe. So today I’m going to talk about some of the work that’s happening here at IT Delhi, where we’re trying to leverage the edge. And the other thing that I’m going to run through very quickly, is TSI and its role in edge. So because we’re focusing here on accelerating development across the global south, so I’m going to pick up those two examples today. Right. So what we’re trying to say is that you have lots and lots of edge agents that will now act simultaneously and in coordination.

So the reason why edge is becoming more and more important is this converge of communication, compute and control. And this convergence is now quite real. And because this convergence is real, it is enabled at least in today’s technology only by a strong edge control specifically for tasks in the area of haptics. As I will show in the next slide, require you to not miss or make mistakes because some of them are catastrophic. And for that reason, a strong development in the area of edge is important. The other reason why looking at edge is important from the perspective of global south is that. While it might not be easy to have foundation models that solve all the problems of the world, at least.

to an extent context has become increasingly important in modern times. People want to provide solutions which are very, very specific to the task at hand and context can be best leveraged or used if there is a strong edge capability that is present. So in that light it is important that the global south focuses on building its strength in the area of edge. This slide here talks about some of the work that we are doing with respect to haptics. Haptics as you know is this sense of touch primarily consists of two aspects. One is kinesthetics which is the pressure that we feel and the second is tactile or texture which is the quality of surface that we, you know the fine grained texture of the surface that we are able to measure using our skin.

So the thing with this kind of a modality is that while it seems to be almost abstract it is quite pervasive. It is all around us, the temperature. including you know the hardness the softness or the way people meet each other greet each other you know all of that is very very important we just don’t you know it’s not overt but it’s important nonetheless so we sort of take it for granted however it is very very important and therefore it needs to be looked at a little carefully now the challenge with haptics is that while as we move from speech to video people did talk about bandwidth and they did talk about latency and there were quality of experience measures that evolved with haptics it goes to the next level because if you have unsynced and delayed haptic inputs or feedback then it becomes quite confounding and it confuses the person and it sometimes can be quite disconcerting so for this reason it is extremely important that the haptics data that you receive is accurate and received on time.

So for this it becomes extremely important that there is a strong capability that is present at the edge. Now here at IIT we are trying to implement it using two ways. One is what we term as split control where we have tried to move from having solutions deployed only in the cloud and the endpoint. We try to put in significant amount of capability on the edge itself. The other aspect that we are looking at carefully is trying to convert signals which are haptic informed to signals which give you the intent rather than actual measurements of pressure as what haptics is to machines. So these two things are primarily handled at the edge. The first one is quite clear.

Let me just say a few words about the second. So when we talk about intent in today’s world whenever you look at a haptic solution it is sort of locked in right from the operator to the endpoint where you have some kind of manipulation, dexterous manipulation of the environment around the device. However it’s very very hard for devices of different manufacturers to interoperate and this happens because it is very very tightly coupled to all the signals that are generated and the form factor of the devices. It’s not as simple as pick up any camera and the image that you get you can show it on any display. So for that reason the idea is to convert those signals into intent, send the intent to the other side and the edge on the other side makes sense of the intent and converts into a signal that the other far point can then use to do whatever works needed.

So these are the two things that we sort of look at with reasonable amount of interest at IIT Delhi and we continue to contribute to standards primarily in the area of MSE and quality of experience where multi -modality is involved. Right. Now this is the edge foundation network i’ll skip this in interest of time because i do have a couple of slides that i want to uh walk you through because there’s some work that’s also being done by tsdsi which is our sdo at here in india and they have in conjunction with itu doing quite a lot of interesting work which is edge centric so uh let me talk about few of those so there are a few technical reports that have come out of late there there’s the stock of dynamic ai ml models for self -sustainable v2x applications so v2x applications is being looked at carefully there’s also work in the area of security aspects and advanced and ai enhanced passive digital twinning initiatives so we have uh some technical reports that have happened in this area there’s also uh developing of standards work that’s happening there’s architectural support for tactile applications that i just spoke about there’s talk of 6g ai architecture for ran and also ai native scalable reference architectures I think maybe we’ll talk about quality of experience in the next slide but that’s another thing we’re looking at.

We’re also carrying out technical studies in all of these areas in interest of time. You’ll have the slides you can go through them when you find the time. This is the other thing that they wanted me to bring to light to this audience. Just a couple of minutes. So the global standard forums that are of interest to the audience here people who look at edge carefully. There’s ITUR IMT 2030 framework for included ubiquitous intelligence for overarching design and then there’s ITUT related standards there are CGPP standards and of course the M2M. So all of these standards are of interest to the audience here and people trying to do research in this area and besides this TSTSI has been trying to be inclusive by holding these flagship conferences annual ones so that more and more people get insight into what is happening.

With that I’ll close because we’re really short of time here. So, Vishnuji, back to you.

Vishnu Ram OV

Thank you, Bajesh. Thank you for bringing out the Indian research in the topic and bringing out the 8GI framework also. It’s very less time. Let me invite Ranjitha. Ranjitha obtained her PhD from ISE. Her current research involves causal inference, survival analysis, and Bayesian neural networks. Over to you, Ranjitha.

Ranjitha Prasad

Yeah, so something that he also missed. I actually do work in federated learning and many other learning paradigms. So let me just start. So mine is going to be a technical talk where I’ll tell you the motivation for using federated learning, especially the role of federated learning in telecom networks and why really are people discussing about this. The motivation is, of course, data explosion. There’s exponential growth in mobile data. There’s exponential growth in data traffic. And you have all these diverse services that are there in 6G, EMBB, URLC. I’m sure this audience is well aware of this. Then there are bottlenecks in these legacy networks which actually motivated moving more towards edge -centric architectures. The goal is, of course, I think this is something very important that most of the standards are looking at.

Predictive zero -touch automation, closed -loop wireless control, and this loop closure latency requirement about less than 10 milliseconds for mission -critical optimization. And this is exactly where federated learning comes in as a key enabler of privacy -preserving and distributed intelligence. So all of this is captured in the AI -native network concept, where now AI is no longer a peripheral layer, but it’s actually coming into the RAN. So this is enabled by what is called as this ORAN alliance, particularly the RIC or the RAN intelligence controller. And this is how the whole system, sub -10 millisecond latency requirement is fulfilled. But something that is not very clear here is… So why do you really require edge intelligence, right?

So to make it even faster and achieve the sub 10 milliseconds, you actually have to bring in inference and training to the edge rather than taking data to the cloud, right? So that’s where the whole paradigm shifted and this argument about edge intelligence or edge native intelligence came in. And especially something called as MEC or multi -axis edge computing also was introduced. So this brought in a huge architectural change. That is, now we have the core network talking to RAN and then RAN talking to the UEs. And this is where the whole, you know, the UEs basically now have the intelligence along with the MEC controllers. So federated learning. Upon all these things, one very important aspect.

that’s how we relate to AI for good is that of privacy right so think of the use case of traffic prediction where there is you know there’s a need for loads and loads of data but you know this data consists of raw user logs location history or I mean if you share it with the with the centralized controller it’s just privacy violation so the solution is to now bring code to the data and not take data to the code right so that’s the that’s where federated learning comes in the intelligence now or the training happens at the edge and only certain metadata is given to the cloud so what is this what is its implication in telecom so there’s impact on privacy that’s exactly where it’s supposed to make the impact and then of course there’s impact on latency and bandwidth so personalization of AI models is possible in real time large scale training can still happen in the core network but the personalization of smaller models for localized applications can happen actually the edge and there’s impact on bandwidth because I no longer need to send data to the server and of course there’s a huge impact on architecture because you saw there that it becomes a hierarchical style of an architecture where core network is at the top and user ues are at the bottom okay so I just wanted to introduce quickly introduce two use cases in fact left it’s in fact a use case from from France it was this is for a traffic prediction in fact predicting certain traffic spikes when they had a football match and this scenario is where you need to allocate dynamically resources for this particular stadium event so here what’s happening is each of these ues or base stations are picking up the traffic in their local area sharing it with the core network sharing it with a MEC controller and then the core network is able to say you know what’s happening here and then you can send data to the server and then you can send data how to really route the traffic so that you know there’s less congestion The other one is V2X.

So this is, again, for sharing road conditions or, you know, accident information and other things. It’s very easy to see why FL may be useful here. Each car can talk to its own edge server and then go to the cloud server where the global model is trained. So this sort of envisages how federated learning has become a very important technology. So last but not the least, I come from, I’m the PI of the lab, which is called as Intellicom Lab at IIIT Delhi. We have a collaboration with IIIT Delhi for this entire work on federated learning on systems. And this project is funded by METI and the India AI Initiative. This is actually more of a security use case.

I’ll not go too much into the use case. But we have built a very similar, I mean, it’s a prototype of the federated learning use case that I just showed you, where we have like a main server, we have some federated clients. and we are looking at certain security incidents that happen only at one client. The client picks it up and then now the entire network knows about the possible security issues that can come up in such scenarios. And we have some couple of publications on this. I’ll really not go into this. Thank you

Vishnu Ram OV

Thank you, Ranjitha, for the excellent talk. We had an introduction at least for federated running and also the framework that architecture that she explained is really interesting. Last time when ITU colleagues were here, we had visited the lab. If you haven’t done that, please talk to her. It’s a very exciting research which they do. And we also have great collaboration with BAPI and colleagues in IIIT Delhi. Thank you, Ranjitha, for coming. we have a panel now we have approximately 20 minutes maybe for the panel let’s kick off the panel can I invite Fred to moderate the panel and can I invite the panelists Mala Alagan and Sakshi to please take the seats Fred to kick off, thank you very much over to you Fred

Frederick Werner

thank you so I’m looking forward to this panel where we can aim to demystify Edge AI a little bit and explore the practical use cases and AI strategies but first I’ll introduce the panel so the first panelist her name is Mala she has a full name but she personally asked me to just call her Mala and I wish all panelists would do that it’s much easier that way so Mala is currently a technologist at the Center of Excellence Wired and Wireless Technologies at Art Park sorry, Art Park so Mala is currently a technologist Prior to this, she was a postdoctoral research at the Teikian Group at Technical University Berlin. She’s also involved in 6G initiatives such as AI RAN for efficient resource allocation and millimeter wave communications.

And she also has been a visiting researcher at UC Davis and TU Berlin. Lala, welcome. Our next panelist is Alagan Mahalingam, founder, CEO, chief software architect of RootCode. Alagan is the founder of RootCode, and in his early 20s, he worked as a researcher at international research organizations such as the Geoinformatics Center at the Asian Institute of Technology, Thailand, also the University of Tokyo, Japan, where he worked on satellite communications and solar panel optimization algorithms. Alagan was also a research associate at the University of Tokyo, Japan, and in his early 20s, he worked as a researcher at International Research Organization, such as the Geoinformatics Center at the Asian Institute of Technology, Thailand, also the University of Tokyo, Japan, where he worked on satellite communications and solar panel optimization algorithms.

Alagan was also awarded the special title of ICT Entrepreneur of the Year at the National ICT Awards in 2021. Alagan was also awarded the special title of ICT Entrepreneur of the Year at the National ICT Awards in 2021. and also the Young Entrepreneur of the Year in 2024. And he’s also the envoy for the government of Estonia e -residency. So I see a lot of Estonia connections here today. Last but not least, we have Shaxi Gupta. So she’s the Global Government Affairs responsible for Qualcomm. She’s a tech policy professional in AI and emerging technology policy analysis, market research, and stakeholder engagement. So if we could have a please warm welcome for the panel. So first question is for Mala.

Mala, as an AI -enabled XR applications, and they’re split between 5G and public, sorry, private 5G and on -premise public 5G, could you please give us some examples of XR applications in different scenarios, and what are the trends? Trade -offs in scalability, security, Thank you.

Mala Kumar

They get the immersive experience in their own preferred regional languages. And one other application that we have done is the XR -assisted medical emergency care. Here the focus is on the, to provide timely medical response to the patient who was suffering with a cardiac arrest and so on. And an SOS alert would be sent from the, by the bystanders from the life circuit exact to the first responders. And the medical experts and the ambulance through 5G connectivity. Once the first responder gets the alert, he arrives at the scene with XR glasses and IOT wearables. and also the AED kit. And while giving the CPR, the IOT patient vitals would be displayed, augmented onto the real -time video.

And the real -time video would also be sent to the medical expert. And the medical expert will guide whether to continue the CPR or it could be the AED and so on. So the timely response will save multiple lives. So in this case, we have used public 5G network. But for the XR -assisted facility tour, we have used private 5G network. So the private 5G network is mainly to have on -premise HCI applications. And this would bring the core next to the… the data generation. And then we can also… do real -time decision -making for industry 5 .0 applications. And going forward, we would like to have some of our applications to be in the open source and have it in the best place, like ITUs, AI for good, right?

So then the international community can access this open source AI models and they can fine -tune the models and they can do the rigorous testing before it is bringing it to a real -world deployment. That is what I’m looking forward for this.

Frederick Werner

Yeah, thanks so much, Mala. And I think this really is a good example of AI for good in action. And I think, to your point, these solutions don’t happen by magic. There’s a lot of difficult problems. There’s a lot of problems to solve. And by putting these solutions in the AI for good, good sandbox that might lead to future standards which could make them replicable and then you could have that adoption at scale. So I’ll just go to the next panelist, Alagan. Given your rich experience in developing AI solutions for partners in different geographies, can you please give us some examples of edge AI deployment in real world scenarios, their impacts, the nuances you see in AI strategies on edge AI in the different regions?

Because from your bio you’ve been involved in many different parts of the world. Thanks.

Alagan Mahalingam

I started RootCode 11 years back because I was in love with building AI solutions as a college student and then now 11 years later the technology that we have built is used by more than 92 million people across 27 countries including many European governments like the government of Estonia, Portugal and many others. We chose to build edge AI in many cases. One, the obvious one, to bring technology to under -connected spaces and also to increase speed in many cases and sovereignty. And the most interesting project that we have done recently, let me tell that story, a couple of years back, Portugal realized that their farmers, especially the small -scale farmers, didn’t get enough access to advisory and intelligence to grow their crops and things have been changing because climate change and unpredictability in growing crops, a lot of people were leaving farming.

And so we built a solution from a hardware, a software product and also an AI model. The hardware goes into the soil so you understand the soil nutrition and you take pictures with the mobile app and we can process the pictures to understand is there a problem with the plant, right? And we built and it worked out fantastically well. And then I tried to bring that to Sri Lanka. I grew up in Sri Lanka, and to date, a big part of our development team is in Colombo, in Sri Lanka, more than 120 people. And so we went into one of these villages in the middle of the mountains of Sri Lanka, Nubaralia. And I was super fascinated.

And when we tried to deploy this, we realized they don’t have reliable connectivity in some corners of the villages. And our solution was worthless. And that’s where we started bringing in Edge. So we brought in a new version. We had a Raspberry Pi, and we started testing models like GemR, and also we did our own convolutional networks like 2D, things to figure out, like, where do you optimize? You don’t want to use LLM for everything, right? And by the end of it, we managed to bring the same value that the software gave to connected users. And that’s how we got to where we are today. And that’s how we got to where we are today.

And that’s how we got to where we are today. to people who didn’t even have internet in some part of Sri Lanka. And that reminded me how much edge is needed, especially in the global south. And yesterday I was at a dinner talking to some of the development finance colleagues from DZ. And somebody was talking about why don’t we put computes on the wheels in a tuk -tuk? So imagine we can’t process too many things on a small device of Raspberry Pi. What if you get a tuk -tuk coming to your village every other day or once a week with a data center built in, with Wi -Fi LAN, so farmers can connect and do the processing.

Smaller banks, smaller institutions can do. And I was like, yeah. So this week has been super fascinating. And sometimes when we think about edge, we think it’s needed only in places that are not really connected, like rural parts. We have built this, we have built a beautiful solution that’s used in America. If you think America is well connected, you should take a road trip. When you go out of the city, you realize some parts are very disconnected. And we built, for one of our clients, we built a solution that helps rural patients who are at high risk with remote patient monitoring. And then, yeah, EDGE works all around the world, not just in the South.

If I, when I think about all my learnings here, because there are so many learnings building EDGE for multiple geographies, multiple customers, multiple communities. If I were to single out, I would single out the fact that when you are trying to do something in the EDGE, we shouldn’t try to think of the model and go find a solution. But instead, think of the task and then work backwards on how do you build and distill or fine tune a smaller model. And that runs on the EDGE because in the EDGE, you can’t do everything, right? if you are building an AI assistant for farmers, you don’t want the AI to be able to tell why two of the famous CEOs didn’t want to hold hands.

I mean, that doesn’t matter. You want it to answer about plants and agriculture. So the heavier the model is, it becomes impossible to deploy. So we work on multiple technologies to quantize or prune the models in a way that creates a smaller version that does exactly what’s supposed to happen. And I think the global south needs to grow with this AI transformation of the world because infrastructure takes decades, but the next few years is going to change the way we live. And that’s why we are here. So I’m excited

Frederick Werner

Yeah, thanks a lot. And I think what you’re saying here has almost been the theme of this week where you don’t need the biggest AI or the biggest large language model. And I think if you look at the example of India, where they’ve managed to enable… billions to have a digital ID to enable financial inclusion, financial payments with the public interest at heart and with relatively low -tech solutions, you can indeed bring AI to the edge in cases that make a lot of sense. So thanks for that. Sapsky, question for you. In your experience with Europe, the intersection of technology, innovation, and AI strategies, what do you think are the metrics to evaluate the usage of edge AI such as availability and capability of hardware at the edge and also the connectivity, privacy, and data issues that you see in your line of work?

Qualcomm Member

Thank you, Fred. And let me start by saying it’s an absolute pleasure to be sitting with fellow panelists and speakers who have preceded me who are deploying edge AI and are doing research on edge AI. At Qualcomm, we are very focused on edge AI and think that that’s going to be the future of how not just Global South, but globally, we’re going to be using AI. So we… Um… And I really relate to what you said about the way to think about deployment of AI is actually to think backwards about what is the use case that you’re trying to solve. And then you think about what is the best architecture that you want to use.

Is it just cloud? Is it on -prem? Is it an edge cloud? Or is it on -device AI? So we have to think about it from a distributed architecture point of view when we think about the use cases that we have here in the Global South. And I do want to mention one important distinction here, which was touched upon earlier also, is that when we think about AI, there’s a training part of it and then there’s an inferencing part of it. Inferencing is where you’re thinking processing is actually happening. So while training can continue to happen on the cloud, a lot of it, a lot of the inferencing, as we’re seeing, is moving towards the edge.

Now, in terms of availability, if I want to talk about it, I think we’re increasing. We’re increasingly seeing, and Qualcomm is deploying this at the edge of… So, you know, AI being available at the edge, not from, you know, the very basic thing that we all use every day is your smartphones. That we have on -device capabilities coming onto smartphones with 10 billion parameters models already running on device. So that means that you do not need to be connected. If you’re in flight mode, you do not need to be connected on internet and you can still use AI. So that’s amazing in my point of view. We also have it coming to actually cars. So Qualcomm has developed that technology where you can now use Edge AI onto the, or it’s actually in development.

We have demos at the Qualcomm booth, which I’ll come to later, but which you can, so Edge AI is coming to the cars as well. And it’s increasingly coming to IoT devices and your smart glasses as well. So in terms of availability, I think we are seeing increasingly that it’s coming to all types of devices. That are connected to internet now. Now, why is AGI relevant? And some of my panelists have already touched on it. I think latency, security, privacy, personalization, low cost, low power are all very important factors for why AGI becomes important for Global South. We may not have access to as much power. We may not have access to as much water as needed.

But with AGI, we don’t have to worry about that. Apart from that, I do want to touch on one thing. That is Qualcomm, one of the things that we have is a program called Tech for Good, wherein we partner and work with startups and small businesses around the world. We invest in them. We mentor them. They use Qualcomm hardware to develop solutions at the edge. In fact, I do want to encourage that in Hall 4 at our Qualcomm booth, we have some of these startups who are displaying this technology. One of the examples is actually from India. It’s called Raksa Health. They’ve actually built an on -device AI healthcare assistant where it’s for doctors and patients both, where the doctors can take down symptoms and provide solutions for their patients and for patients to actually look up their prescriptions and be able to access all their records offline and ask questions about it.

So, yeah, I think that’s how we’re seeing the transition happen. Thank you.

Frederick Werner

Thank you. Yeah, some amazing use cases. And I think this week’s coming out of Davos where the narrative was all about go, go, go, the insatiable demands for energy. There’s talk of putting data centers in space. But I think this panel also brings things a bit down to earth where, you know, you can have AI on the edge, and, of course, there’s a lot of things to solve there. when it comes to connectivity, when it comes to data compute. I think there’ll be a lot of standards development work that needs to emerge from this to make this work at scale. But I think your use cases and the way you’re approaching the problem, especially starting from the what are you trying to solve and work backwards from that, I think is very refreshing compared to all the headlines we’ve been seeing lately.

And I don’t see it either or. I see it as a big piece, a complementary piece of the puzzle. So with that, I really want to thank the panel. And if we could have a round of applause for them. Thank you.

Vishnu Ram OV

Thank you very much, Fred, for running the tight panel. Now we are coming to the closing. Thank you, panelists, insightful remarks. Yes. Okay. Can I ask a quick group photo of the panelists, please? Panelists. Yes. Thank you. Thank you very much. Thank you very much. Now, we are coming to the closing. There are excellent closing remarks coming. Can I please request Her Excellency, Ms. Lopez, Ambassador, Permanent Representative, Permanent Mission of the Republic of El Salvador to United Nations Office and other international organizations in Geneva to please give her closing remarks.

Egriselda López

I’m actually based in New York. Thank you. Well, good afternoon. I know that I don’t have much time, but I had just to say that this discussion was very enlightening. Thank you so much for sharing everything what you’re doing on the ground. And I guess that it was very clear to me that HAI means simply using an AI closer to where things happen. That means closer to people, closer to services, communities, rather than deepening only faraway systems. So amazing what you’re already doing. So this can be important for development because it can work better in places with limited connectivity, as we were hearing, and it can help with speed. And it can help with speed, cost, and privacy, since not everything has to be sent everywhere.

So I guess that I had to begin also with something. I am also the co -chair of the Global Dialogue on AI Governance. This is going to happen in July this year, and it’s going to be the first dialogue of its kind. So trying to also bring together what we have been hearing from member states and also other stakeholders in these months, I can tell you three specific things connecting with what we just heard today. First, that people must remain at the center. And we have heard with all these examples. And I guess that a common message that we have been hearing also in this week is that AI should be developed and used in a way that protects but also helps people.

Second, closing the gap is not a slogan. We are hearing this a lot. It requires decisive support. And I was very pleased, for instance, saying that you’ve been trying to replicate in some countries what it has in others, for instance. And I think that’s a good thing. This information sharing, this is critical if we’re talking about closing the gaps. And the third message, the final one, is that we should avoid a world of disconnected approaches. And this also is aligned with what I was just saying, that cooperation across different national but also regional approaches, it will help us to reduce fragmentation. So, with that, I just have to tell you that we’re very looking forward to see some of you in Geneva in July, so we can hear and learn more about what AI is.

So, it’s my pleasure to give the floor to my distinguished co -chair, Ambassador Reintam Saar. He’s going to explain to you very, very shortly what the global dialogue on AI governance is. And this is really important work that we are putting a lot of effort to it. Thank you so much again for the invitation. Thank you.

Reintam Saar

yes hello hello everyone frankly i really feel humbled among real experts not to say i feel helpless so please allow me then to do a little bit of awareness raising when it comes to the first global dialogue on ai governance and maybe this way i’ll try to fit into a discussion that we’ve heard here today so three points on my side first about tasking so the tasking was to put together a distinctive identifiable un global dialogue with all the elements that are prescribed in the mandate so bringing governments and stakeholders together to exchange best practices and of course to focus on cooperation and to execute it back to back with itu you you uh ai for good uh um summit in july in geneva produce co -chair summary.

So this is what we are going to do. So, so far we’ve engaged with member states, with stakeholders, multi -stakeholders, and from member states we’ve kind of covered three different approaches, I would say a little bit. Risks versus opportunities, state -centric approach versus multi -stakeholder approach, closing AI divide versus free market innovation, but we also were able to pick up three convergences, practical outcomes preferred over endless theoretical discussions, alignment with existing UN processes, avoiding duplication, clear timeline formats and thematic focus to produce actionable insights. And the unified element, I would say, in these discussions is that the dialogue needs to be inclusive. and capacity building was absolutely a crucial element that is, of course, one of the most important things to a global self.

So from multi -stakeholders, what we’ve heard, the key words, so to say, were trust, transparency, no duplication, interoperability, equal access and participation for everyone, rooting the dialogue in human rights and to be of a practical value and innovative in form. So what we are going to do, we will guide the discussions, but we will not predetermine the outcome. It’s for member states, it’s for you, for stakeholders. And, of course, we will engage also with international scientific panel that was also established through the same resolution. We will rely on member states and your wisdom, we would need to collect this wisdom somehow. and this is something that we are going to do because we would need this wisdom so that the dialogue would be really inclusive we would come up on certain point with a road map to Geneva where you would see building blocks towards dialogue and whatever opportunities to engage into dialogue and of course I very much hope that all these fantastic ideas and frankly I mean chapeau to the panel because you are already making or changing life on the ground and it’s absolutely fantastic we really need this also to inform our dialogue and so that the dialogue would be also result oriented on the ground.

Thank you very much.

Vishnu Ram OV

Thank you thank you thank you thank you and can we have the momentos for Brijeshji thank you very much Ranjitha Ranjitha please Moala can I request Nodal officer to please felicitate Fred yeah we have an event with him yeah yeah yeah yeah yeah yeah yeah yeah yeah yeah thank you very much for attending the session session is closed thank you thank you Thank you. Thank you.

Frederick Werner

Speech speed

150 words per minute

Speech length

2097 words

Speech time

833 seconds

AI for Good Vision

Explanation

The AI for Good initiative is designed to unlock AI’s potential to serve humanity by leveraging a partnership of more than 50 UN agencies and focusing on solutions, skills, and standards. It is positioned as a year‑long programme rather than a single annual summit.

Evidence

“Well, simply put, it’s to unlock AI’s potential to serve humanity.” [1]. “We have 50 UN sister agencies as partners of AI for good, contributing knowledge, sharing expertise, helping to drive our standards work.” [3]. “I think it’s fair to say that AI for Good is indeed more than a summit.” [12]. “But it’s actually a year -long activity.” [16].

Major discussion point

Vision and Goals of AI for Good Initiative

Topics

Artificial intelligence | The enabling environment for digital development

Autonomous AI Agents

Explanation

Future AI systems are moving toward a “zero‑click” world where agents act without waiting for human prompts, indicating a shift toward pervasive edge AI.

Evidence

“And now we’re looking at a world where you’re basically entering a zero -click world where agents are not waiting for our prompts.” [60].

Major discussion point

Importance and Technical Foundations of Edge AI

Topics

Artificial intelligence | Information and communication technologies for development

Professor Brijesh Lall

Speech speed

167 words per minute

Speech length

Standards Development and AI Governance

Topics

Artificial intelligence | Internet governance | The enabling environment for digital development

Ranjitha Prasad

Speech speed

171 words per minute

Speech length

978 words

Speech time

342 seconds

Federated Learning for Latency & Privacy

Explanation

Federated learning enables sub‑10 ms latency for mission‑critical telecom tasks by moving inference and training to the edge, while preserving privacy through distributed intelligence.

Evidence

“So to make it even faster and achieve the sub 10 milliseconds, you actually have to bring in inference and training to the edge rather than taking data to the cloud, right?” [56]. “And this is exactly where federated learning comes in as a key enabler of privacy -preserving and distributed intelligence.” [64]. “The motivation is, of course, data explosion.” [73]. “Predictive zero -touch automation, closed -loop wireless control, and this loop closure latency requirement about less than 10 milliseconds for mission -critical optimization.” [74].

Major discussion point

Federated Learning as Enabler for Edge AI and Privacy

Topics

Artificial intelligence | Data governance | Human rights and the ethical dimensions of the information society

Traffic Prediction & V2X Use Cases

Explanation

Federated learning is applied to traffic‑prediction for large events and V2X safety, allowing local data to stay at the edge, reducing bandwidth and preserving privacy.

Evidence

“this is a use case … for a traffic prediction … predicting certain traffic spikes when they had a football match …” [71]. “The other one is V2X.” [71]. “there’s impact on bandwidth because I no longer need to send data to the server” [71]. “that’s where federated learning comes in the intelligence now or the training happens at the edge and only certain metadata is given to the cloud” [71].

Major discussion point

Real‑World Use Cases and Challenges in the Global South

Topics

Artificial intelligence | Data governance | Social and economic development

Mala Kumar

Speech speed

108 words per minute

Speech length

301 words

Speech time

166 seconds

XR Edge AI with 5G

Explanation

XR‑assisted facility tours and medical‑emergency care are delivered over a mix of private 5G and public 5G networks, showcasing practical edge AI deployments.

Evidence

“But for the XR -assisted facility tour, we have used private 5G network.” [35]. “Mala, as an AI -enabled XR applications, and they’re split between 5G and public, sorry, private 5G and on -premise public 5G…” [36]. “And one other application that we have done is the XR -assisted medical emergency care.” [37]. “And the medical experts and the ambulance through 5G connectivity.” [38]. “So in this case, we have used public 5G network.” [40]. “So the private 5G network is mainly to have on -premise HCI applications.” [41].

Major discussion point

Importance and Technical Foundations of Edge AI

Topics

Artificial intelligence | Information and communication technologies for development | Closing all digital divides

Open‑Source XR Models via AI for Good

Explanation

Open‑source AI models for XR can be shared globally through AI for Good platforms, enabling the international community to fine‑tune and test them before real‑world deployment.

Evidence

“So then the international community can access this open source AI models and they can fine -tune the models and they can do the rigorous testing before it is bringing it to a real -world deployment.” [90]. “And going forward, we would like to have some of our applications to be in the open source and have it in the best place, like ITUs, AI for good, right?” [91].

Major discussion point

Real‑World Use Cases and Challenges in the Global South

Topics

Artificial intelligence | Data governance | The enabling environment for digital development

Alagan Mahalingam

Speech speed

166 words per minute

Speech length

854 words

Speech time

307 seconds

Task‑First Edge AI for Agriculture

Explanation

Edge AI for agriculture starts with the task, then works backwards to distill smaller models that can run with limited connectivity, as demonstrated in Portugal’s small‑scale farmer advisory project.

Evidence

“And sometimes when we think about edge, we think it’s needed only in places that are not really connected, like rural parts.” [46]. “But instead, think of the task and then work backwards on how do you build and distill or fine tune a smaller model.” [47]. “And the most interesting project that we have done recently… Portugal realized that their farmers, especially the small -scale farmers, didn’t get enough access to advisory and intelligence to grow their crops…” [48].

Major discussion point

Importance and Technical Foundations of Edge AI

Topics

Artificial intelligence | Social and economic development | Closing all digital divides

Remote Patient Monitoring

Explanation

Edge AI solutions are used to monitor high‑risk rural patients remotely, providing continuous health oversight without requiring constant connectivity.

Evidence

“And we built, for one of our clients, we built a solution that helps rural patients who are at high risk with remote patient monitoring.” [88].

Major discussion point

Real‑World Use Cases and Challenges in the Global South

Topics

Social and economic development | Health | Closing all digital divides

Qualcomm Member

Speech speed

172 words per minute

Speech length

Collaboration and Multi‑Stakeholder Engagement

Topics

The enabling environment for digital development | Capacity development

Vishnu Ram OV

Speech speed

68 words per minute

Speech length

469 words

Speech time

410 seconds

Multi‑Stakeholder Collaboration

Explanation

The panel demonstrates cooperation among UN agencies, academia, industry and civil society, co‑creating edge AI solutions and linking them to AI for Good activities.

Evidence

“Our touch points with ITU are many, where he’s hosted AI for Good Challenges, WTSA Hackathon.” [97]. “Panelists.” [113].

Major discussion point

Collaboration and Multi‑Stakeholder Engagement

Topics

The enabling environment for digital development | Internet governance | Capacity development

Egriselda López

Speech speed

151 words per minute

Speech length

458 words

Speech time

181 seconds

People‑Centered AI Governance

Explanation

AI should be developed and used in ways that protect people while also delivering benefits, emphasizing information sharing to close gaps and reduce fragmentation.

Evidence

“And I guess that a common message that we have been hearing also in this week is that AI should be developed and used in a way that protects but also helps people.” [109]. “This information sharing, this is critical if we’re talking about closing the gaps.” [110]. “And this also is aligned with what I was just saying, that cooperation across different national but also regional approaches, it will help us to reduce fragmentation.” [111].

Major discussion point

Standards Development and AI Governance

Topics

Artificial intelligence | Human rights and the ethical dimensions of the information society | Closing all digital divides

Reintam Saar

Speech speed

127 words per minute

Speech length

481 words

Speech time

225 seconds

Inclusive Global AI Governance Dialogue

Explanation

The Global AI Governance Dialogue aims for practical outcomes, alignment with existing UN processes, avoidance of duplication, and inclusive multi‑stakeholder participation.

Evidence

“Risks versus opportunities, state -centric approach versus multi -stakeholder approach, closing AI divide versus free market innovation, but we also were able to pick up three convergences, practical outcomes preferred over endless theoretical discussions, alignment with existing UN processes, avoiding duplication, clear timeline formats and thematic focus to produce actionable insights.” [102]. “So from multi -stakeholders, what we’ve heard, the key words, so to say, were trust, transparency, no duplication, interoperability, equal access and participation for everyone, rooting the dialogue in human rights and to be of a practical value and innovative in form.” [106].

Major discussion point

Standards Development and AI Governance

Topics

Artificial intelligence | Internet governance | Human rights and the ethical dimensions of the information society

Agreements

Agreement points

Edge AI is crucial for addressing connectivity challenges in rural and underserved areas

Speakers

– Frederick Werner
– Professor Brijesh Lall
– Alagan Mahalingam
– Egriselda López
– Qualcomm Member

Arguments

Context-specific solutions are more important than foundation models for Global South, requiring strong edge capabilities for localized applications

Agricultural AI solutions for farmers require edge deployment to work in areas with poor connectivity, using hardware like Raspberry Pi for local processing

Edge AI is crucial for Global South development as it works better in areas with limited connectivity and helps with speed, cost, and privacy

On-device AI capabilities are expanding to smartphones, cars, and IoT devices, enabling offline functionality and reducing dependency on internet connectivity

Summary

All speakers agreed that edge AI is essential for bringing AI capabilities to areas with limited connectivity, particularly in the Global South, enabling offline functionality and reducing dependency on cloud-based systems

Topics

Artificial intelligence | Closing all digital divides | Information and communication technologies for development

Task-specific optimization is more effective than general-purpose AI models for edge deployment

Speakers

– Professor Brijesh Lall
– Alagan Mahalingam
– Qualcomm Member

Arguments

Context-specific solutions are more important than foundation models for Global South, requiring strong edge capabilities for localized applications

Edge AI deployment requires working backwards from specific tasks to build smaller, optimized models rather than using heavy general-purpose models

On-device AI capabilities are expanding to smartphones, cars, and IoT devices, enabling offline functionality and reducing dependency on internet connectivity

Summary

Speakers consistently emphasized the importance of developing context-specific, optimized AI models for edge deployment rather than relying on large general-purpose models

Topics

Artificial intelligence | Information and communication technologies for development

International collaboration and standards development are essential for AI deployment

Speakers

– Frederick Werner
– Professor Brijesh Lall
– Mala Kumar
– Vishnu Ram OV

Arguments

AI for Good is a year-long movement with 50+ UN agencies focused on solutions, skills, and standards to unlock AI’s potential for humanity

TSDSI is developing standards for dynamic AI/ML models, security aspects, and architectural support for tactile applications in collaboration with ITU

Open source AI models and rigorous testing frameworks are needed for international community access and real-world deployment validation

ITU has strong collaborative relationships with Indian research institutions and actively engages in AI for Good initiatives including hackathons and challenges

Summary

All speakers emphasized the critical importance of international collaboration, standards development, and open-source approaches to ensure AI benefits are accessible globally

Topics

Artificial intelligence | The enabling environment for digital development | Capacity development

Privacy and security are fundamental considerations in AI deployment

Speakers

– Ranjitha Prasad
– Qualcomm Member
– Egriselda López

Arguments

Federated learning enables privacy-preserving distributed intelligence in telecom networks, allowing personalization while maintaining data privacy

On-device AI capabilities are expanding to smartphones, cars, and IoT devices, enabling offline functionality and reducing dependency on internet connectivity

Edge AI is crucial for Global South development as it works better in areas with limited connectivity and helps with speed, cost, and privacy

Summary

Speakers agreed that privacy preservation and security are key benefits of edge AI deployment, with federated learning and on-device processing reducing privacy risks

Topics

Artificial intelligence | Human rights and the ethical dimensions of the information society | Data governance

Similar viewpoints

Both speakers emphasized the technical requirements for edge AI, focusing on low-latency applications and the need for optimized, task-specific models rather than general-purpose solutions

Speakers

– Professor Brijesh Lall
– Alagan Mahalingam

Arguments

Edge AI enables convergence of communication, compute and control, essential for haptics applications requiring sub-10 millisecond latency

Edge AI deployment requires working backwards from specific tasks to build smaller, optimized models rather than using heavy general-purpose models

Topics

Artificial intelligence | Information and communication technologies for development

Both speakers provided concrete examples of AI applications serving local communities with specific cultural and practical needs, demonstrating the real-world impact of edge AI deployment

Speakers

– Mala Kumar
– Alagan Mahalingam

Arguments

XR applications include immersive cultural experiences in regional languages and medical emergency care with real-time expert guidance via 5G connectivity

Agricultural AI solutions for farmers require edge deployment to work in areas with poor connectivity, using hardware like Raspberry Pi for local processing

Topics

Social and economic development | Artificial intelligence | Information and communication technologies for development

All three speakers representing UN organizations emphasized the importance of inclusive, practical, and human-rights-centered approaches to AI governance and development

Speakers

– Frederick Werner
– Egriselda López
– Reintam Saar

Arguments

AI for Good is a year-long movement with 50+ UN agencies focused on solutions, skills, and standards to unlock AI’s potential for humanity

The Global Dialogue on AI Governance aims to bring governments and stakeholders together for inclusive cooperation and capacity building

UN global dialogue should focus on practical outcomes, avoid duplication, and be rooted in human rights with equal access for all participants

Topics

Artificial intelligence | Human rights and the ethical dimensions of the information society | Capacity development

Unexpected consensus

Arguments

Edge AI deployment requires working backwards from specific tasks to build smaller, optimized models rather than using heavy general-purpose models

Context-specific solutions are more important than foundation models for Global South, requiring strong edge capabilities for localized applications

Summary

While both speakers advocate for localized solutions, Mahalingam emphasizes starting with specific tasks and working backwards to create optimized models, while Lall focuses more on leveraging context through edge capabilities. Mahalingam is more explicit about avoiding general-purpose models entirely.

Topics

Artificial intelligence | Closing all digital divides

Infrastructure approach – fixed edge vs. mobile edge computing

Speakers

– Alagan Mahalingam
– Mala Kumar
– Professor Brijesh Lall

Arguments

Mobile edge computing and innovative deployment methods like compute-on-wheels can bring AI capabilities to disconnected rural communities

XR applications include immersive cultural experiences in regional languages and medical emergency care with real-time expert guidance via 5G connectivity

Edge AI enables convergence of communication, compute and control, essential for haptics applications requiring sub-10 millisecond latency

Summary

Mahalingam proposes innovative mobile solutions like compute-on-wheels for areas with poor connectivity, while Kumar and Lall focus on fixed infrastructure solutions using 5G networks and established edge computing architectures. This represents different philosophies about how to deliver edge AI to underserved areas.

Topics

Information and communication technologies for development | Closing all digital divides

Unexpected differences

Governance approach – technical standards vs. inclusive dialogue

Speakers

– Frederick Werner
– Egriselda López
– Reintam Saar

Arguments

ITU has over 400 AI standards published or in development, including work on future networks, 5G/6G, and AI-native networks

The Global Dialogue on AI Governance aims to bring governments and stakeholders together for inclusive cooperation and capacity building

UN global dialogue should focus on practical outcomes, avoid duplication, and be rooted in human rights with equal access for all participants

Explanation

While all speakers represent UN-related organizations working on AI governance, there’s an unexpected divergence in approach. Werner emphasizes technical standards development through ITU, while López and Saar focus on inclusive multi-stakeholder dialogue processes. This suggests potential institutional differences in AI governance philosophy within the UN system.

Topics

Artificial intelligence | Human rights and the ethical dimensions of the information society | Follow-up and review

Overall assessment

Summary

The discussion revealed relatively low levels of direct disagreement, with most speakers focusing on complementary aspects of edge AI rather than opposing viewpoints. The main areas of difference centered on technical approaches (task-specific vs. context-aware models), infrastructure deployment strategies (mobile vs. fixed), and governance methodologies (standards vs. dialogue).

Disagreement level

Low to moderate disagreement level with significant implications for implementation strategies. The differences suggest that while there’s broad consensus on the importance of edge AI for the Global South, there are multiple valid pathways for achieving these goals. This diversity of approaches could be beneficial for innovation but may require coordination mechanisms to ensure interoperability and avoid fragmentation.

Partial agreements

Both agree on the importance of standards and frameworks for AI deployment, but Werner focuses on formal ITU standards development while Kumar emphasizes open source approaches and community-driven testing

Speakers

– Frederick Werner
– Mala Kumar

Arguments

ITU has over 400 AI standards published or in development, including work on future networks, 5G/6G, and AI-native networks

Open source AI models and rigorous testing frameworks are needed for international community access and real-world deployment validation

Topics

Artificial intelligence | The enabling environment for digital development

Both advocate for distributed AI architectures that preserve privacy and reduce cloud dependency, but Prasad focuses on federated learning approaches while Qualcomm emphasizes on-device processing capabilities

Speakers

– Ranjitha Prasad
– Qualcomm Member

Arguments

Federated learning enables privacy-preserving distributed intelligence in telecom networks, allowing personalization while maintaining data privacy

On-device AI capabilities are expanding to smartphones, cars, and IoT devices, enabling offline functionality and reducing dependency on internet connectivity

Topics

Artificial intelligence | Data governance | Human rights and the ethical dimensions of the information society

Similar viewpoints

Both speakers emphasized the technical requirements for edge AI, focusing on low-latency applications and the need for optimized, task-specific models rather than general-purpose solutions

Speakers

– Professor Brijesh Lall
– Alagan Mahalingam

Arguments

Edge AI enables convergence of communication, compute and control, essential for haptics applications requiring sub-10 millisecond latency

Edge AI deployment requires working backwards from specific tasks to build smaller, optimized models rather than using heavy general-purpose models

Topics

Artificial intelligence | Information and communication technologies for development

Both speakers provided concrete examples of AI applications serving local communities with specific cultural and practical needs, demonstrating the real-world impact of edge AI deployment

Speakers

– Mala Kumar
– Alagan Mahalingam

Arguments

XR applications include immersive cultural experiences in regional languages and medical emergency care with real-time expert guidance via 5G connectivity

Agricultural AI solutions for farmers require edge deployment to work in areas with poor connectivity, using hardware like Raspberry Pi for local processing

Topics

Social and economic development | Artificial intelligence | Information and communication technologies for development

All three speakers representing UN organizations emphasized the importance of inclusive, practical, and human-rights-centered approaches to AI governance and development

Speakers

– Frederick Werner
– Egriselda López
– Reintam Saar

Arguments

AI for Good is a year-long movement with 50+ UN agencies focused on solutions, skills, and standards to unlock AI’s potential for humanity

The Global Dialogue on AI Governance aims to bring governments and stakeholders together for inclusive cooperation and capacity building

UN global dialogue should focus on practical outcomes, avoid duplication, and be rooted in human rights with equal access for all participants

Topics

Artificial intelligence | Human rights and the ethical dimensions of the information society | Capacity development

Takeaways

Key takeaways

Edge AI is essential for Global South development as it enables AI functionality in areas with limited connectivity while providing benefits in speed, cost, privacy, and personalization

AI for Good has evolved from a concept-focused initiative in 2017 to a practical year-long movement with real-world applications, organized around three pillars: solutions, skills, and standards

Federated learning enables privacy-preserving distributed intelligence in telecom networks, allowing AI training at the edge while keeping sensitive data local

Successful Edge AI deployment requires working backwards from specific use cases rather than trying to adapt general-purpose models, with emphasis on model optimization and quantization

Real-world applications demonstrate Edge AI’s versatility across sectors including healthcare (emergency response), agriculture (crop advisory), cultural preservation (XR experiences), and telecommunications (traffic prediction)

The convergence of communication, compute, and control at the edge is enabling new applications requiring sub-10 millisecond latency, particularly for haptics and mission-critical systems

International cooperation and standards development are crucial for scaling Edge AI solutions globally, with ITU leading efforts on AI-native networks and related standards

Resolutions and action items

ITU to continue developing AI standards with over 400 currently published or in development, focusing on future networks and AI-native architectures

TSDSI to advance technical reports and standards work on dynamic AI/ML models, security aspects, and architectural support for tactile applications

Global Dialogue on AI Governance scheduled for July in Geneva to bring together governments and stakeholders for inclusive cooperation

Promotion of open source AI models through platforms like AI for Good to enable international community access and rigorous testing

Encouragement for startups and researchers to participate in AI for Good challenges, hackathons, and machine learning competitions

Development of AI Skills Coalition with machine learning environment sandboxes for government training and capacity building

Unresolved issues

Specific metrics and evaluation frameworks for measuring Edge AI deployment success across different regions and use cases

Standardization challenges for interoperability between different Edge AI devices and manufacturers, particularly in haptics applications

Infrastructure requirements and investment strategies for deploying Edge AI in underconnected areas of the Global South

Balance between centralized training and edge inference in federated learning architectures for optimal performance

Regulatory and governance frameworks needed to support Edge AI deployment while ensuring privacy and security

Scalability challenges for moving from prototype Edge AI solutions to large-scale commercial deployment

Technical specifications for innovative deployment methods like ‘compute-on-wheels’ for rural areas

Suggested compromises

Hybrid approach combining cloud-based training with edge-based inference to balance computational requirements with privacy and latency needs

Gradual transition from cloud-centric to edge-native AI architectures rather than complete replacement

Use of intent-based signaling in haptics to enable interoperability between different device manufacturers while maintaining performance

Combination of public and private 5G networks depending on specific use case requirements and security needs

This comment cuts through technical discussions to address the fundamental governance and equity issues surrounding AI deployment. It challenges the audience to move beyond technical solutions to consider systemic approaches to ensuring AI benefits are distributed equitably, and warns against fragmented approaches that could exacerbate global inequalities.

Impact

This comment brought the discussion full circle by connecting the technical innovations discussed throughout the session to broader questions of global governance and equity. It elevated the conversation from implementation details to policy implications and reinforced the session’s theme about AI serving humanity rather than the reverse, providing a framework for thinking about how the technical solutions discussed could be scaled and governed responsibly.

Overall assessment

Explanation

Standardized evaluation metrics are needed to assess and compare edge AI implementations across different contexts and regions

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.