AI for Social Good: Using Technology to Create Real-World Impact

Contents

Executive Summary

This India AI Impact Summit panel discussion explores how AI, paired with open digital public infrastructure (DPI) and decentralized networks, can achieve population-scale transformation in healthcare, agriculture, and education across the global south. The speakers argue that AI's greatest potential lies not in frontier model development but in embedding intelligence into existing government digital infrastructure to serve billions of economically vulnerable populations at minimal cost.

Key Takeaways

AI's social impact depends on infrastructure, not just models: Brilliant inference models fail without DPI connecting them to public sector workers and citizens. India's decade of DPI investment is now the enabling layer for AI-driven health, agriculture, and education outcomes.
Open networks + low-cost inference + multilingual agents = population-scale inclusion: This combination is the formula for bringing billions into AI-powered services. Closed platforms and high inference costs guarantee limited reach.
Data sharing (and trust) is India's strategic advantage: The ability to build consent-based, secure health and agricultural data stacks uniquely positions India to train locally relevant models and export solutions to the global south.
Replicability requires standardization, not one-size-fits-all copying: The World Bank and others are working to identify minimal, modular components (e.g., the agricultural network blueprint) that can be adapted to local contexts without requiring full infrastructure rewrites.
The next frontier is biological-AI convergence for medicine: AI should not replace but rather augment biological understanding. Cell reprogramming, regenerative medicine, and precision health will emerge from combining distributed biological intelligence principles with AI's pattern-recognition power.

Summit Talk Summary

Key Topics Covered

Digital Public Infrastructure (DPI) as enabler: Role of coordinated, open-network systems in distributing AI solutions at scale
Open networks vs. closed platforms: Why decentralized, interoperable systems outperform proprietary approaches
AI inference cost: The critical importance of reducing inference costs to enable AI adoption in developing nations
Language and localization: Multilingual AI agents that serve non-English-speaking populations
Health stack development: Building integrated health data systems (phenotypic, genomic, demographic, radiological) for universal healthcare
Agricultural innovation: AI-powered agents for credit, crop prediction, and pest surveillance serving smallholder farmers
Education at scale: AI-driven literacy diagnosis and remediation reaching millions of students
Biological intelligence and AI convergence: Learning from distributed biological systems to improve AI efficiency
Tuberculosis detection via cough analysis: Real-world case study of AI applied to infectious disease diagnosis
Global replication of India models: Scaling Indian DPI-based solutions to Africa, Brazil, and Southeast Asia

Key Points & Insights

Population-scale impact requires coordination infrastructure: AI models alone don't solve problems; they must integrate into pre-existing DPI that connects policy, program implementation, public workers, and citizens. Without this infrastructure, deployment costs are prohibitively high.
Open networks unlock exponential innovation: The UPI payment system model—open, decentralized, interoperable—should inform AI infrastructure. Open networks allow multiple innovators to build applications at the edge using AI agents, enabling massive diffusion of technology without vendor lock-in.
Multilingual AI agents are the distribution mechanism: Language barriers dissolve when users interact with AI agents in their native language (Hindi, Tamil, local dialects), which then execute complex transactions behind the scenes. This is "inclusion at massive scale."
Cost of inference is the critical bottleneck: Inference must cost near-zero in the global south (under 5 rupees per query). Current inference pricing makes AI economically inaccessible to populations the technology could most help. This must be addressed as a priority equal to model training.
Data access without data sharing remains problematic globally: India is unique in having an open, consent-based, secure data-sharing model (as in UPI). Most developing countries lack both the infrastructure and institutional willingness to share health, agricultural, and educational data needed to train locally relevant AI models. This is a competitive disadvantage.
Real-world impact is measurable and already occurring:
- TB diagnosis via cough sound: 25% increase in case detection within one year
- Early-grade reading diagnosis: 3–8 million students enrolled; mandatory rollout across multiple states
- Agricultural networks: Millions of smallholder farmers accessing credit and crop prediction in Uttar Pradesh
- Protein structure prediction: India is 4th largest user of AlphaFold; 3 million+ researchers in 190 countries using it
Biology and AI should inform each other: Biology operates on "sips of energy" (distributed, efficient) vs. AI's gigawatt-scale data centers. Future AI should learn from biological principles: distributed processing, generational learning (DNA as memory), and low-power inference.
Inference model interoperability drives multiplier effects: Open networks allow cutting-edge models (e.g., Google's improved weather models) to be plugged into infrastructure serving millions (e.g., 38 million Indian farmers receiving monsoon predictions), without reimplementing systems.
Cost of delivery often goes undiscussed but is central to scaling: Early-grade reading assessment at 5 rupees per student is economically feasible; at 500 rupees, it's not. Cost modeling must be integral to solution design from the start.
India's DPI success is a global template: UPI, Bhashini, Nikshay (TB database), Rakshuk (education DPI), and health stacks demonstrate that governments can build the coordination layer. Other developing nations can adapt, localize, and replicate these models rather than reinventing them.

Notable Quotes or Statements

James Manika (Google): "We need to ensure that the digital divide does not become an AI divide."
Nandan Nilekani (Infosys, Networks for Humanity): "AI agents on an open network [are] the fundamental construct for massive diffusion of technology... the real power of agents is removing complexity for the user."
Nandan Nilekani: "If you're serving a customer with one query and that costs 500 rupees, it's not going to work... inference has to drop dramatically."
Kiran Mazumdar Shaw (Biocon): "If you combine biological intelligence with artificial intelligence and converge that, I think we're in for huge transformation... How do you convert a cancer cell into a non-malignant cell? That's the holy grail."
Sunil Wadhwani (Wadhwani AI): "The government of India is asking: can you help? We identified three or four key pain points in the patient's journey [for TB]... all of that was enabled by this DPI."
Kiran Mazumdar Shaw: "Biology works through distributed data centers. And when it wants to build intelligence, retrieve memory and inference data, it does so with sips of energy, not with gigawatts of power."
Sangu Kim (World Bank): "We are a sommelier... [trying to] understand the wines and then find [the] better wines for our customers' taste" (analogy for curating AI solutions for developing-world governments).
Sunil Wadhwani: "We have 25 AI platforms in India in education, healthcare, agriculture which are scaling up... Governments throughout the global south are hungry for these solutions and looking to India to provide these."

Speakers & Organizations Mentioned

Speaker	Role/Title	Organization
James Manika	Senior Vice President, Research & Society	Google; Co-chair, UN High-Level Advisory Board on AI
Nandan Nilekani	Co-founder & Chairman	Infosys; Networks for Humanity Foundation
Sangu Kim	Vice President, Digital Development	World Bank
Kiran Mazumdar Shaw	Chairperson	Biocon Group
Sunil Wadhwani	Co-founder	Wadhwani Institute for Artificial Intelligence

Government & Institutional Partners Referenced:

Government of India (Ministry of Agriculture, Ministry of Health)
Government of Uttar Pradesh
Government of Rajasthan
Indian Institute of Science (ISC)
World Bank
UN High-Level Advisory Board on AI

Technical Concepts & Resources

AI Models & Systems

AlphaFold / AlphaFold Protein Database: Nobel Prize–winning protein structure prediction; 3M+ researchers in 190 countries; India 4th largest adopter
AlphaGenome: Referenced as advancing biological understanding via AI
Gemini: Google's AI powering multilingual agents in agricultural networks (Uttar Pradesh pilot)
Neural GCM: Google weather model; delivered monsoon predictions to 38M Indian farmers
Cough-to-TB detection model: AI model diagnosing tuberculosis from smartphone audio; achieved 25% increase in case detection in rollout
Reading diagnosis model: 20-second assessment of early-grade literacy; identifies specific word/phrase/sentence difficulties

Digital Public Infrastructure (DPI)

UPI (Unified Payments Interface): Model for open, decentralized payment networks; now adapted for AI integration
Bhashini: Government of India's multilingual language infrastructure; supports 100+ Indic languages, 20 newly digitized
Nikshay: TB patient management database; enables AI-driven diagnosis and treatment optimization
Rakshuk: Rajasthan state-level education DPI; reaches 400,000 schools, 8M students
Health Stack: Phenotypic, genomic, demographic, radiological health data infrastructure (in development)
AgriConnect: World Bank–Google initiative; open network for smallholder farmers; being replicated to Kenya, Nigeria, Ethiopia, Brazil, Philippines

Datasets & Initiatives

Project Vani: Google–ISC collaboration for Indic language speech data; Phase 2 completed across all Indian states
AI for Bharat: Government initiative for language accessibility
VoiceAI: Multilingual voice infrastructure

Key Metrics & Scale Targets

TB Impact: 2M deaths/year globally; 500K deaths/year in India; 4–5M TB cases annually in India; 40% of TB cases globally undiagnosed
Education Scale: 75M children targeted by end of 2027 (currently 10M students, 2M educators reached via AIEL)
Frontline Worker Empowerment: 1.4M health workers targeted for multilingual AI assistance (child malnutrition focus)
Reading Intervention Cost: 5 rupees per student (cost metric for scalability)
Inference Cost Target: Stated as critical; must approach near-zero in global south

Governance & Frameworks

Consent-based data sharing model: India's framework for secure health and agricultural data access
Open-network principles: Decentralization, interoperability, plug-in architecture for models
Google.org Impact Challenges: Two ongoing funding mechanisms—AI for Science and Government Innovation

Geographic Expansion

Current deployment: India, Kenya, Nigeria, Ethiopia, Brazil, Philippines, Singapore, Switzerland (innovation labs)
Replication approach: Standardization of modular components; localization by national context; World Bank facilitating multi-country scaling

Context & Significance

This summit represents a strategic moment in global AI governance: India's decade-long investment in DPI is yielding measurable public health, education, and agricultural outcomes. The speakers collectively argue that this model is both replicable and necessary—that the global south cannot wait for frontier AI labs to solve localized problems, and that decentralized, open infrastructure is more cost-effective and sustainable than proprietary solutions. The emphasis on inference cost, multilingual capability, and integration with existing government systems reframes the AI conversation from "bigger models" to "cheaper, locally relevant deployment."