AI for Social Good: Nonprofit Innovations at Scale

Contents

Executive Summary

This panel discussion examines how AI is being deployed by nonprofits and civil society organizations to address social justice, health, language preservation, and judicial reform across the Global South—with a particular focus on India. Rather than treating communities as passive beneficiaries of AI, the speakers advocate for community-centered AI development where affected populations become architects of the technology, own the data, and benefit economically from AI work.

Key Takeaways

Community Ownership, Not Charity: The most effective AI for social good centers affected communities as data owners, model builders, and evaluators—not as passive recipients. Karya's model of 140,000+ workers performing 65M+ AI tasks and owning datasets in 70+ Indian languages is the template.
"Painkillers Before Multivitamins": When scaling AI in government or entrenched institutions, solve acute, visible problems first (e.g., judges handwriting) before attempting comprehensive transformation. This builds trust and creates demand for deeper integration.
Multilingual AI Is Solvable, Not Theoretical: The Nandbar case proves that building production AI models for low-resource tribal languages is feasible within weeks—not years—if embedded teams, government buy-in, and proper incentives exist. The barrier is will, not capability.
Synthetic Media Threatens Epistemic Trust Itself: The existential risk from deepfakes isn't individual hoaxes but systemic loss of faith in evidence. Organizations must prepare for a world where people distrust all media, which empowers authoritarian narrative control.
Reframe Nonprofits as Revenue-Generating Institutions: Sustainable social impact comes from diverse revenue streams (government contracts, market services, foundational support) not primarily from grants. Nonprofits should operate like businesses in financial discipline while maintaining mission-driven incentives.

Key Topics Covered

AI in the Nonprofit Sector: How nonprofits use AI for operational efficiency, grant-making transformation, and program delivery
Misinformation & the Information Ecosystem: AI's role in creating synthetic media (deepfakes) and the erosion of trust in digital information
Language Preservation & Multilingual AI: Building AI models for India's 19,000+ dialects and tribal languages; the AI value chain from data collection to evaluation
Judicial Reform via AI: Automating clerical processes in courts to reduce case backlogs and accelerate justice delivery
Community-Centered AI Development: Ensuring communities own data, participate in model evaluation, and receive economic benefits
Nonprofit Business Models: Generating sustainable revenue through market partnerships, government contracts, and platform services rather than relying solely on grants
Philanthropic Strategy & Funding: How foundations like Pat McGovern Foundation support nonprofits in AI adoption
Scaling Justice at the District Level: Real-world case study of building multilingual AI models in Nandbar, Maharashtra in 45 days

Key Points & Insights

Communities Must Be Architects, Not Beneficiaries: Villas Dar (Pat McGovern Foundation) emphasizes that the fundamental question isn't "how do nonprofits use AI as customers" but "how do we ensure communities are architects of what AI looks like?" This requires moving AI capacity-building out of Silicon Valley and into communities.
Power Asymmetry vs. Information Asymmetry: Manu Chopra (Karya) articulates a critical distinction: India doesn't have an information asymmetry problem—it has a power asymmetry problem. Building better chatbots doesn't solve structural inequity; the technology itself must be decentralized and community-controlled.
The Nandbar Model: 45-Day Language AI Development: Karya completed data collection, model building, fine-tuning, and evaluation for six tribal languages in 45 days by embedding teams in the district, ensuring locals owned the datasets and built the models. This demonstrates that multilingual AI at scale is feasible if there is will.
Deepfakes Create an "Epistemic Crisis" Larger Than Individual Hoaxes: Vivian Schiller (Aspen Digital) warns that synthetic media's biggest danger isn't that people will believe individual fakes, but that mass exposure to deepfakes will create blanket cynicism—making people distrust all information, including genuine evidence of atrocities. This opens space for autocrats to manipulate discourse.
AI-Driven Stenography Solves the "Painkiller" Problem in Courts: Utkarsh Sakenna (Adalat AI) strategically chose stenography automation (not end-to-end solutions) because judges handwriting court records is an acute, visible pain point. This "painkiller vs. multivitamin" approach accelerates adoption and government mandate (Kerala and Andhra Pradesh have already mandated Adalat AI).
81% Revenue Sharing with Data Workers: Karya gave away 81% of all revenue from data contracts directly as wages to the 140,000+ workers engaged, with wages growing 6.8x year-over-year. This sustainable model demonstrates that communities can profit from the AI economy, not just participate in it.
Grant Guardian: AI-Driven Philanthropic Due Diligence: Pat McGovern Foundation built Grant Guardian—an AI tool that automates financial risk assessment for nonprofits, eliminating hundreds of pages of grant applications and labyrinthine documentation. This reduces friction in philanthropy itself.
Samudai Portal: Government-Nonprofit Collaboration at Scale: Launched yesterday (in the talk), Samudai allows any organization to directly hire Karya's trained workers without Karya as a middleman, provided they pay living wages and avoid harmful work. This enables government spending on language digitization ($50M+ per state) to directly reach workers.
AI + Analog Data = Multiplied Impact: The Kushy Baby maternal health case study shows that combining 20 years of analog field data with geospatial AI modeling identified six villages with abnormal low-birth-weight clusters in six months—leading to targeted interventions and measurable health improvements in a fraction of the time traditional research would take.
Nonprofits Must Maintain Revenue-Positive Operations: Both Villas and Vivian emphasize that "nonprofit" is a tax status and philosophy, not a magic exemption from financial sustainability. Nonprofits must bring in more revenue than they spend; the difference is where the surplus goes (back into impact, not shareholder returns).

Notable Quotes or Statements

"How do we make sure that people, communities, and the institutions that represent them are the architects of what AI looks like? How do we take the capacity to build frontier models and move it out of Silicon Valley and move it into the communities that we live in?" — Villas Dar, Pat McGovern Foundation

"India doesn't have an information asymmetry problem. We have a power asymmetry problem." — Manu Chopra, Karya

"The process itself has become the punishment." — Utkarsh Sakenna, Adalat AI (on India's 50M case backlog and 80% of prisons holding undertrials)

"I think people are now so cynical about what they see that the bigger problem is not that we will fall for some fake but that we will stop believing anything at all." — Vivian Schiller, Aspen Digital

"Every good nonprofit needs a path to its own irrelevance." — Manu Chopra, Karya (on building self-sustaining platforms so nonprofits aren't needed as intermediaries)

"Painkillers work better than multivitamins." — Utkarsh Sakenna, Adalat AI (on scaling strategy)

"Nonprofit is a philosophy, a tax code status, a state of mind. Nonprofits can generate revenue, can do so profitably, and can use it to sustain their work." — Villas Dar

Speakers & Organizations Mentioned

Primary Panelists:

Villas Dar — President, Pat McGovern Foundation (largest philanthropic funder of AI in nonprofits; $500M+ in grants over past 5 years)
Vivian Schiller — Director, Aspen Digital (Aspen Institute); former roles at New York Times, NPR
Manu Chopra — Founder/Lead, Karya (AI + employment, language preservation, data work)
Utkarsh Sakenna — Founder/Lead, Adalat AI (judicial reform, court automation)
Amar Bhidé — Moderator; Founder/Lead, Indiaspora (diaspora community across 200+ countries, chapters in US, UK, Singapore, Canada, UAE, Australia)

Organizations & Institutions:

Pat McGovern Foundation — Major funder of nonprofit AI initiatives
Aspen Institute / Aspen Digital — Think tank focused on AI's impact on society, media, policy
Karya — Nonprofit building AI for low-resource languages and worker economic empowerment
Adalat AI — Nonprofit deploying AI to automate judicial clerical processes
Indiaspora — Diaspora community organization; organizing global forum in Bangalore
Kushy Baby — Maternal health nonprofit in Rajasthan (case study)
Bhashini — Government of India language AI initiative
Samudai — Government portal for direct worker engagement (launched during talk)
JPAL (J-PAL) — Running randomized control trial on Adalat AI's impact
Anthropic — Major AI lab collaborating with Karya on evaluations
Microsoft Research — Partner in Karya's evaluation work
Government of India — Primary funder and partner for most initiatives

Technical Concepts & Resources

AI/ML Concepts:

Data Collection → Model Building → Fine-Tuning → Evaluation: The complete AI value chain, with emphasis on community participation at all stages
Synthetic Media / Deepfakes: AI-generated video, audio, photography used for misinformation (distinguished from benign uses like translation or transcription)
Geospatial AI Modeling: Combining analog field data with location data to identify health/development disparities (Kushy Baby case)
Multilingual & Multimodal Models: Training on multiple languages and input types; identified as priority for frontier AI labs
Fine-tuning for Low-Resource Languages: Adapting general models to tribal languages and regional dialects (6 languages in Nandbar study)
Model Evaluation at Community Scale: 500,000+ public and expert evaluations performed by Karya workers to assess trustworthiness, local relevance, and utility

Tools & Platforms:

Grant Guardian — Pat McGovern Foundation's AI tool for automated nonprofit financial due diligence
Samudai — Government of India portal connecting organizations with trained Karya workers for AI data tasks
Adalat AI Platform — Stenography automation, digital case filing, court management systems
Bhashini — Govt. of India language AI initiative (partnership with Karya)

Datasets & Initiatives:

Nandbar Tribal Language Datasets: Six tribal languages (Bili, Ahirani, others) in Maharashtra; datasets owned by district and community
65 Million AI Tasks: Completed by Karya's 140,000+ workers across 70+ Indian languages
Foundational Data in 70+ Indian Languages: Built through Karya's market-based model

Policy & Research:

Randomized Control Trial (RCT) with J-PAL: Evaluating Adalat AI's impact on case resolution times and judicial efficiency (advised by Esther Duflo, Abhijit Banerjee, Daron Acemoglu)
GILA (Global Alliance for Learning): Announced initiative on AI and education; endorsed by panelists

Relevant Statistics/Benchmarks:

India's Case Backlog: 50 million cases pending; at current disposal rate, 300 years to clear
Indian Prison Demographics: 80% are undertrials (not convicted), waiting for cases to resolve
Adalat AI Impact: Doubling/tripling judicial output; 30-50% reduction in case resolution times (internal studies)
Karya's Reach: 140,000+ workers engaged; 81% of revenue distributed as direct wages; 6.8x year-over-year wage growth
India's Linguistic Diversity: 19,000+ dialects requiring digital representation; government spending ($50M+/year per state on language digitization)
Maternal Health Case Study (Kushy Baby): Identified health disparities in six villages; vitamin supplementation interventions saw statistically significant improvements within six months

Note: This transcript contains some audio artifacts and repetition typical of live speech. Where claims are quantitatively specific (e.g., "$500M in grants," "81% revenue sharing," "65M tasks"), these are preserved as stated by panelists but should be verified against official sources for publication.