India AI Impact Summit Insights
All sessions

AI for Social Good: Aligning Research with National Priorities | India AI Impact Summit 2026

Contents

Executive Summary

The India AI Impact Summit 2026 opening session established a framework for evaluating AI's real-world impact on development outcomes across health, governance, education, agriculture, and labor. Key speakers—Ekbal Aiden (JPAL), Nobel laureate Michael Kremer, and India's Ministry of Electronics and IT Secretary S. Krishnan—emphasized that rigorous impact evaluation, not hype, should drive AI deployment in public services, particularly in the Global South.

Key Takeaways

  1. Evidence, Not Enthusiasm, Should Drive Deployment: Rigorous randomized control trials (RCTs) and impact evaluations prevent scaling failed technologies. The "silver bullet" mindset—common in government—must be replaced with systematic testing and learning.

  2. AI's Real Advantage is Amplifying Humans, Not Replacing Them: The most documented successes augment frontline workers (farmers with forecasts, teachers with adaptive software, drivers with improved testing). AI works best as a productivity multiplier for constrained capacity (India cannot hire all needed doctors, nurses, teachers).

  3. Procurement & Incentive Design Matters as Much as Technology: Pay-for-performance, multi-vendor approaches, and usage-linked accountability drive quality and sustained adoption better than lowest-cost bidding or single-vendor lock-in.

  4. Multi-Phase Funding (Pilot → Test → Scale) Reduces Waste: Small exploratory grants, medium grants for rigorous evaluation, and larger grants for proven scales allow evidence to guide resource allocation and reduce failed experiments.

  5. Global South Model: Frugal Innovation + Open Infrastructure: India's approach—public compute, open-source model adaptation, distributed state experimentation, and data platforms—is replicable in African, Southeast Asian, and Latin American contexts and requires less capital than proprietary alternatives.

India AI Impact Summit 2026 — Opening Session Summary

Key Topics Covered

  • Impact Evaluation Framework: Moving beyond lab results to field-tested, evidence-based AI deployment
  • AI Applications in Development: Six priority areas where AI can reduce poverty and improve outcomes
  • Government Procurement & Experimentation: Structuring contracts to incentivize quality and enable policy learning
  • Addressing Technology Hype Cycles: Learning from failed initiatives (e.g., One Laptop per Child) to avoid AI overhype
  • State-Level Implementation: Role of states as testbeds for AI policy and evaluation
  • Data Accessibility & Privacy: Balancing open data access with privacy protection
  • India's AI Mission: Compute, models, and datasets as public goods for researchers and startups
  • Organizational Change: Building institutional capacity to adopt and sustain AI in government
  • Global South Perspectives: Frugal innovation models applicable across LMICs

Key Points & Insights

  1. Six Priority Areas for AI Impact:

    • Targeting & needs prediction (e.g., flood forecasting, disease risk mapping)
    • Personalized timely support (e.g., crop-specific disease recommendations, adaptive learning)
    • Frontline service amplification (augmenting teachers, nurses, doctors, extension workers)
    • Reducing human biases in hiring, promotion, and recruitment (vs. algorithmic bias concerns)
    • Organizational efficiency (court case management, one-call triaging, tax optimization)
    • Progressive taxation (using AI for fairer, more equitable tax systems)

  2. Critical Caution on Hype: The "One Laptop per Child" failure (launched 2013 in Uttar Pradesh, abandoned by 2014) demonstrates the danger of scaling technologies without rigorous field evaluation. Newspaper headlines reversed within 12 months, illustrating how even well-intentioned initiatives can fail without evidence.

  3. Documented Real-World Wins:

    • Monsoon forecasts: 38 million Indian farmers received AI-enabled forecasts in 2025; preliminary evidence from Odisha shows farmers receiving delayed-progression forecasts were more likely to plant hybrid seeds and adjust transplanting decisions.
    • Driver's license testing (HAMS): Microsoft Research India's AI system reduced unsafe driving reports by 20-30% across 56 RTOs and 345,000+ tests. Impact measured via on-road data from Ola (cab aggregator).
    • Adaptive learning (Andhra Pradesh): Personalized software doubled student learning pace with just 1 hour/week usage in grades 6–9; rolled out across 1,500+ schools.
    • Poshan Tracker & E-Sanjivani: Existing health systems show strong evidence of impact.

  4. Evaluation Gaps: Most AI in government services lacks rigorous evaluation because:

    • Commercial incentives favor lucrative private-sector applications over public goods
    • Government adoption is slower than private sector (no disruptive new-entrant dynamics)
    • Procurement systems often default to lowest-cost bidding, ignoring quality variation

  5. Multi-Stakeholder Coalition Model: Effective AI for social good requires integration of:

    • Researchers (academics testing real-world impact)
    • Implementing partners (governments, NGOs, social enterprises)
    • Donors (philanthropies and development agencies funding outcomes)

  6. Pay-for-Performance Contracts: NITI Aayog tested contracts that pay edtech providers based on learning gains, not just delivery. This aligns incentives toward continuous product improvement and addresses both software quality and utilization challenges.

  7. Rapid Feedback Loops: Edtech and digital services enable real-time usage data. Government partners can rapidly identify low-usage schools and intervene (as occurred in Andhra Pradesh), accelerating course correction without waiting for full evaluation cycles.

  8. State Capacity Building: India's 54 Centers of Excellence (one per state, funded by Ministry of Electronics and IT) serve as learning hubs to help states identify, evaluate, and scale AI solutions. This distributed model enables experimentation and cross-state learning.

  9. Data as Public Infrastructure: India's AI Kosh (platform) makes datasets openly available to researchers and startups, addressing traditional government data silos. This required 10+ years of cultural and policy work but unlocks research and innovation.

  10. Theory of Change Timelines: Evaluation duration is not bottlenecked by assessment methodology but by the underlying intervention. Whether AI, teacher training, or hardware, if a 6–9-month period is needed for learning outcomes to materialize, that cannot be accelerated. Modern digital data collection speeds up evaluation execution, not the inherent causal timeline.

Notable Quotes or Statements

"Unless we know a technology works in the lab but also works in the field, how it intersects with the field and what is going to be the final outcome, we should not get too carried away by these silver bullets." — Ekbal Aiden, JPAL

"I'm much more worried about the biases in us humans because the biases in us have taken thousands and thousands of years to evolve. Whereas with AI, if I do it well, it's a few lines of code that you can actually work on to try and get those biases out." — Ekbal Aiden

"Evaluation is going to be very important, and I think thinking through systems for organizational change in governments—including new systems for procurement—is going to be very important." — Michael Kremer, University of Chicago

"What takes time in an evaluation is not necessarily the evaluation but your theory of change as to how long does it take to make a difference in the outcome. And that has not changed, right—like whether it is AI, whether it is teacher training, whether it is one laptop per child, if it is going to take 6 to 9 months to increase the learning outcomes of children, that cycle cannot be accelerated." — Ekbal Aiden

"We are never going to have as many teachers as we want. We are never going to have as many doctors as we want. If we can actually enhance their productivity, not only is there a possibility that the quality of services will improve significantly, but there's a possibility that we would be able to provide the kind of services that would truly delight, something that we did not think was possible before, but we can do today." — S. Krishnan, India Ministry of Electronics and IT

"Impact evaluations, if they're done well, are not just telling you whether this particular AI application worked or not. It's telling you why something worked—how the technology intersected with our lives, how it intersected with human behavior, how it changed trust in technology." — Ekbal Aiden

Speakers & Organizations Mentioned

Core Speakers

  • Ekbal Aiden — Global Executive Director, Abdul Latif Jameel Poverty Action Lab (JPAL), MIT; Director, Project AI Evidence
  • Michael Kremer — Professor, University of Chicago; Director, Development Innovation Lab; 2019 Nobel Laureate in Economic Sciences
  • S. Krishnan — Secretary, Ministry of Electronics and Information Technology (MeitY), Government of India; Former Finance Secretary, Tamil Nadu
  • Shobini Mukharji — Head, JPAL India Office

Referenced Researchers & Practitioners

  • Abijit Banerjee, Esther Duflo — Co-founders of JPAL; Nobel laureates
  • Monik (Prof.) — Flood forecasting research, Bihar
  • Sam & Thiago — Education case studies; Thiago from Brazil
  • Abishek Singh — Government partner (mentioned as speaker)
  • Aperna Krishnan — (Implicit reference to data analytics work mentioned in same context as S. Krishnan's prior JPAL collaboration)

Organizations & Initiatives

  • JPAL (Abdul Latif Jameel Poverty Action Lab) — MIT-based RCT and impact evaluation center
  • Project AI Evidence — JPAL's AI research and funding arm
  • NITI Aayog — Government of India policy body (pay-for-performance contract testing)
  • Ministry of Electronics and Information Technology (MeitY) — India government; overseeing AI Mission, Centers of Excellence
  • Rocket Learning — Edtech organization (education case study)
  • Neura Health — Health technology organization
  • Adalat AI — Court case management efficiency (referenced for judicial backlog reduction)
  • Central Square Foundation — Partner on Andhra Pradesh adaptive learning evaluation
  • Microsoft Research India — Developer of HAMS (driver's license testing AI)
  • Ola — Indian cab aggregator (provided on-road driving data for HAMS evaluation)
  • Poshan Tracker — India government health nutrition tracking system
  • E-Sanjivani — India government telemedicine system
  • Gates Foundation, Community Jameel — Donor agencies
  • Anthropic, Google, Meta — Tech companies mentioned as summit participants
  • AI Kosh — India's AI infrastructure platform (compute, models, datasets)

Government Bodies

  • Government of India (overall)
  • Ministry of Electronics and Information Technology (MeitY)
  • Government of Andhra Pradesh (adaptive learning rollout)
  • Government of Telangana (monsoon forecast study)
  • Government of Odisha (2025 monsoon forecast evidence)
  • State RTOs (driver's license testing deployment)

Technical Concepts & Resources

AI/ML Models & Techniques

  • Large Language Models (LLMs) — Referenced as core technology (e.g., ChatGPT)
  • Multimodal AI — Ability to process text, video, audio (highlighted as key advantage)
  • Adaptive/Personalized Learning Software — Edtech system that adjusts to individual student pace
  • Open-source Model Adaptation — India's strategic choice vs. building proprietary LLMs
  • Foundational Models (Small, Multimodal, Visual, Quantitative, Mathematical) — India's diversified model development approach

Evaluation Methodologies

  • Randomized Controlled Trials (RCTs) — Gold-standard causal evaluation method used by JPAL
  • Administrative Data Analysis — Using government datasets for insights (vs. RCTs alone)
  • Real-world impact measurement — Evaluating field outcomes, not lab performance
  • A/B Testing — Rapid testing of product variations (mentioned as valuable)
  • Pay-for-Performance Metrics — Contracts tied to learning gains or outcome variables
  • Usage Analytics — Real-time digital data collection from edtech, traffic systems, etc.

Data & Infrastructure

  • AI Kosh — India's open platform for AI compute, models, and datasets
  • Data Sharing Memoranda of Understanding (MoUs) — Formal agreements enabling government data sharing
  • Anonymized/Privacy-Protected Datasets — Enabling open access while protecting individuals
  • Administrative Data — Government databases (education, health, tax, traffic, judicial)

Applications & Case Studies

  • Monsoon Onset Forecasting — 38 million farmers, evidence from Telangana & Odisha
  • Flood Forecasting & Warnings — Detailed geo-spatial predictions (Bihar case)
  • Crop Disease Diagnosis — Farmer photo uploads receiving personalized, soil-specific recommendations
  • Driver's License Testing (HAMS) — Microsoft Research India system; 345,000+ tests; 20–30% reduction in unsafe driving
  • Adaptive Learning Software — Andhra Pradesh grade 6–9; doubled learning pace at 1 hour/week
  • Court Case Management (Adalat AI) — Reducing judicial backlog
  • One-Call Triaging — Better emergency dispatch routing
  • Tax Assessment & Progressivity — AI for fairer tax systems (Senegal example referenced)
  • Traffic Violation Detection — AI cameras for enforcement transparency

Policy & Governance Frameworks

  • India AI Mission — 3-year national initiative (launched ~2023); compute, models, data, R&D
  • Centers of Excellence (54 nationwide) — State-level AI learning and experimentation hubs
  • Multi-Phase Funding Model:
    • Pilot grants (small)
    • Rigorous testing grants (medium)
    • Scale-up grants (larger)
  • Procurement Innovation: Multi-vendor approaches, usage-based accountability, pay-for-performance
  • Social Equity Capitalism — Non-profit investment model for social-impact AI (referenced as needed)

Comparative/Historical References

  • One Laptop per Child (OLPC) — Case study in technology hype cycle failure (Uttar Pradesh, 2013–2014)
  • Financial Engineering Pre-2008 Crisis — Parallel example of hype cycle and unvalidated solutions
  • Development Innovation Ventures (DIV, U.S.) — Multi-phase funding model referenced as successful precedent

Methodological & Conceptual Notes

  • Theory of Change: The causal pathway from intervention to outcome; timeline cannot be artificially shortened, but digital data collection speeds execution.
  • Hype Cycle vs. Evidence: A recurring theme; newspapers/media flip narratives when technologies fail without rigorous testing.
  • Distributed Experimentation Model: States as testbeds; central facilitation of evaluation and cross-learning (India's advantage due to federal structure).
  • Productivity as Core Challenge: Many government functions (judicial, health, education) are capacity-constrained; AI as amplifier more realistic than replacement narrative.

Context & Significance

This session frames AI policy for the Global South, emphasizing that impact—not innovation per se—should guide investment and deployment. The location (Delhi, India AI Impact Summit 2026, first held in Global South) and participants signal a deliberate effort to center evidence-driven AI governance in middle-income contexts, with lessons for Africa, Southeast Asia, and Latin America.