AI for Good: Technology That Empowers People

Contents

Executive Summary

This session focused on edge AI deployment in underconnected regions, emphasizing how moving computation closer to data sources enables faster, more private, and more cost-effective AI solutions. Multiple speakers demonstrated that edge AI is not just a technical optimization but a critical tool for achieving inclusive, equitable AI development in the Global South, where cloud infrastructure and reliable connectivity are limited.

Key Takeaways

Edge AI is not optional for the Global South—it is foundational. Limited cloud infrastructure, connectivity gaps, and latency-sensitive applications make edge computation the pathway to inclusive AI development.
Start with the problem, not the model. Successful deployments begin by defining the specific task (crop disease detection, traffic prediction, patient monitoring), then design or distill appropriately-sized models—not the reverse.
Privacy and efficiency are co-benefits of edge architecture. Federated learning and on-device inference simultaneously reduce privacy risks, lower bandwidth costs, decrease latency, and enable real-time personalization.
Standardization across regions drives adoption at scale. Without common standards for MEC, RAN intelligence, quality of experience, and data formats, edge AI solutions remain siloed; ITU, 3GPP, and regional SDOs are actively closing this gap.
Open-source, sandbox-based validation before deployment reduces risk. Testing AI models in the AI for Good sandbox with real datasets before full standardization and deployment builds confidence and enables rapid iteration across geographies.

Key Topics Covered

AI for Good framework and its three pillars: Solutions, Skills, and Standards
Edge AI architecture and its role in 5G/6G networks
Haptics and tactile feedback systems requiring sub-10ms latency
Federated learning as a privacy-preserving alternative to centralized data collection
Real-world XR and medical AI applications enabled by 5G/private networks
Model optimization techniques (quantization, pruning, distillation) for resource-constrained devices
Global AI governance dialogue and standards development
Use cases across multiple geographies: agriculture (Portugal, Sri Lanka), healthcare, emergency response, traffic management
Hardware accessibility and deployment challenges in underconnected regions

Key Points & Insights

Edge AI is essential for latency-critical applications: Tasks involving haptics, autonomous vehicles, and real-time medical response require sub-10ms latency, which can only be achieved by processing at the edge rather than in distant cloud data centers.
Privacy-by-architecture through federated learning: Rather than centralizing sensitive data (traffic patterns, health records, location history), federated learning trains models locally and shares only aggregated insights, preserving individual privacy while enabling global intelligence.
Task-driven model design, not model-first deployment: Successful edge AI requires starting with the specific problem to solve, then backward-engineering an appropriately sized model—not forcing large models onto constrained devices. Example: An agriculture AI should optimize for crop-related queries, not general knowledge.
Connectivity gaps are not disappearing: Even in developed regions (e.g., rural USA), reliable internet connectivity is sparse. Edge computing solves this by allowing offline inference on devices or through periodic mobile processing centers (e.g., "compute-on-wheels" tuktuk concept).
Standardization is critical for scalability and interoperability: ITU and regional SDOs (3GPP, TSDSI, O-RAN Alliance) are actively developing over 400 AI-related standards covering MEC, RAN intelligence, tactile applications, and 6G architectures—critical for reproducible, globally deployable solutions.
Open-source and sandbox approaches accelerate adoption: The AI for Good sandbox allows researchers and developers to test solutions with real data before standardization, leading to replicable deployments at scale.
Split control and intent-based signaling enable interoperability: Rather than tightly coupling device-specific haptic signals, converting to intent-level abstractions allows different manufacturers' systems to interoperate seamlessly.
Infrastructure-agnostic solutions work globally: The same AI agricultural advisory system deployed in Portugal with internet connectivity was successfully adapted for offline Sri Lankan villages using Raspberry Pi edge devices—demonstrating universal applicability.
AI availability on consumer devices is accelerating: Modern smartphones already run 10B+ parameter models on-device; AI is expanding to cars, IoT devices, and smart glasses, making edge inference a commodity feature.
Inclusive governance requires multistakeholder participation: The UN Global Dialogue on AI Governance emphasizes practical outcomes, human rights alignment, capacity building, and equal participation—not theoretical frameworks alone.

Notable Quotes or Statements

"What if the last thing that humans ever invent is invention itself?" — Fred Vana (ITU), posing the philosophical question of whether AI will become humanity's final invention, framing the urgency of ensuring AI serves humanity.

"Edge is needed especially in the global south...and sometimes when we think about edge we think it's needed only in places that are not really connected, but we have built solutions for America. When you go out of the city, you realize some parts are very disconnected." — Alagan Mahaling (Root Code), contextualizing edge AI as a global necessity, not just a developing-world patch.

"You don't want to use an LLM for everything...don't want the AI to tell why two famous CEOs didn't want to hold hands. You want it to answer about plants and agriculture." — Alagan Mahaling, emphasizing task-specific model design over generalist approaches.

"Edge AI means simply using AI closer to where things happen. That means closer to people, closer to services, communities rather than deepening only far away systems." — H.E. Claudia López (El Salvador, UN Permanent Representative), reframing edge AI as an equity issue.

"Capacity building was absolutely crucial element...the dialogue needs to be inclusive." — Ambassador Rein Tamsar (UN Global Dialogue on AI Governance), highlighting that AI governance must include underrepresented stakeholders.

Speakers & Organizations Mentioned

Primary Speakers

Fred Vana — Chief, Strategic Engagement, ITU (UN specialized agency for digital technologies)
Prof. Vijay Bhaskar (referenced as "Lal Brides") — Bharti School chairman, IIT Delhi; active in AI research and 6G initiatives
Ranjida (full name not fully transcribed) — Federated Learning researcher, IntelliomLab, IIT Delhi; PI on security and federated learning projects
Mala — Technologist, Center of Excellence, Wired and Wireless Technologies, ART Park (AI and Robotics Technological Park), India; researcher in 6G, millimeter-wave communications
Alagan Mahaling — Founder & CEO, Root Code; ICT Entrepreneur of the Year (2021), Young Entrepreneur of the Year (2024); Estonia e-residency envoy
Sakshi Gupta — Global Government Affairs, Qualcomm; AI and emerging technology policy professional
H.E. Claudia López — Permanent Representative, Republic of El Salvador to UN; Co-chair, UN Global Dialogue on AI Governance
Ambassador Rein Tamsar — Co-chair, UN Global Dialogue on AI Governance

Organizations & Institutions

ITU (International Telecommunication Union) — Organizer of AI for Good; 50+ UN sister agencies as partners
IIT Delhi — Research on haptics, MEC, federated learning
TSDSI (Telecommunications Standards Development Society, India) — SDO developing edge-centric technical reports and standards
3GPP — Standards for mobile networks
O-RAN Alliance — Open RAN with RAN Intelligence Controller (RAIC)
ART Park (AI and Robotics Technological Park) — India DST innovation hub
Root Code — AI solutions for agriculture, healthcare; operates in 27 countries, 92M+ users
Qualcomm — Hardware manufacturer; Tech for Good program; on-device AI at scale
UN Global Dialogue on AI Governance — Multi-stakeholder initiative on AI policy and cooperation
Draxa Health — Qualcomm Tech for Good partner; on-device AI healthcare assistant (India)

Technical Concepts & Resources

Architectures & Frameworks

Edge Computing / Multi-access Edge Computing (MEC) — Computing at the network edge, closer to users and data sources
Split Control — Distributing AI processing between cloud and edge; enables haptic feedback, real-time control
Intent-Based Signaling — Abstraction of device-specific sensor signals into intent-level messages for interoperability
Federated Learning — Distributed training where models remain on local devices; only aggregated updates sent to central server
AI-Native Networks — 5G/6G networks with AI embedded in RAN (Radio Access Network) layer, not as a peripheral add-on
ORAN (Open RAN) with RAIC (RAN Intelligence Controller) — Open architecture for programmable radio access networks
Hierarchical Architecture — Core network at top; UEs (user equipment) and MEC at edge; enables personalized model training locally

Standards & Organizations

ITU AI for Good — Three pillars: Solutions, Skills, Standards
400+ AI standards (published/in development) covering:
- Future networks (5G, 6G)
- AI-native networks
- MEC and quality of experience
- Tactile/haptic applications
- V2X (Vehicle-to-Everything) communications
- Digital twins
- Security in distributed AI
IMT-2030 Framework — ITU-R framework for ubiquitous intelligence and AI in next-generation networks
1M2M Standards — Machine-to-machine communications
3GPP Standards — Mobile network specifications including AI-in-RAN

AI Models & Techniques

Gemma — LLM referenced as testable on edge devices (Raspberry Pi)
Quantization — Reducing model precision to shrink size and memory footprint
Pruning — Removing non-essential weights/connections from neural networks
Distillation — Training smaller models to mimic larger models
Convolutional Neural Networks (CNNs) — For image-based tasks (e.g., plant disease detection)
10B+ Parameter On-Device Models — Modern smartphones running ~10 billion parameter models natively

Applications & Use Cases

XR (Extended Reality) for Medical Emergency Care — AR-guided CPR with real-time vital signs; IoT sensors integrated; public 5G
XR Assisted Facility Tours — Multilingual immersive tours in regional languages; private 5G; inclusive experience design
Agricultural Advisory AI — Soil nutrient sensing + image-based plant health diagnosis; offline-capable on Raspberry Pi; deployed Portugal, Sri Lanka
Traffic Prediction (V2X) — Predicting spikes for stadium events; dynamic resource allocation; federated learning across base stations
Vehicle-to-Everything (V2X) — Sharing road conditions, accident data; federated learning with cloud-trained global models
Remote Patient Monitoring (Rural) — High-risk patient monitoring with edge processing; offline-capable; deployed in rural USA
Security Incident Detection (Federated) — Distributed network detecting novel security threats; knowledge shared globally without exposing raw logs
Haptic Feedback for Telesurgery & Remote Manipulation — Sub-10ms latency requirement; requires edge-local processing and intent abstraction

Hardware Targets

Smartphones — 10B+ parameter models running natively
Raspberry Pi — Low-cost edge device for agriculture and IoT
IoT Devices — Sensors, wearables, embedded systems
Smart Glasses / XR Headsets — AR/VR inference on-device
Vehicles (Autonomous & Connected) — On-car AI processing
Mobile Processing Centers ("Compute-on-wheels")** — Tuktuk/mobile van with local data center for periodic village processing

Quality of Experience (QoE) Metrics

Latency — Sub-10ms requirement for mission-critical applications (haptics, telemanipulation)
Bandwidth — Reduced by local processing; federated learning avoids centralizing raw data
Privacy — Preserved through on-device inference and federated learning (no data exfiltration)
Personalization — Real-time model adaptation at edge; localized fine-tuning without cloud round-trips
Power Efficiency — Edge inference reduces transmission energy; critical in battery-constrained devices

Additional Context

Temporal Framing:

AI for Good launched in 2017; evolved from hype-driven PowerPoints to substantive deployments
2023: Generative AI era (GPT-style models)
2024 (recent): AI agents, embodied AI, robotics, brain-computer interfaces, space AI computing
UN Global Dialogue on AI Governance: First of its kind; scheduled July 2024, Geneva (back-to-back with ITU AI for Good Summit)

Policy & Governance Themes:

Human-centered AI; protection and empowerment of people
Closing AI capability gaps (Global South access)
Cooperation over fragmentation across national/regional approaches
Actionable, practical outcomes preferred over theoretical frameworks
Inclusive participation of governments, industry, civil society, academia