India AI Impact Summit Insights
All sessions

AI for All: India’s Public-Interest Policy Architecture

Contents

Executive Summary

This roundtable discussion at Bharat Mandapam convened policymakers, technologists, academics, and international experts to examine how India can develop AI governance frameworks that serve the public interest while advancing toward "Viksit Bharat 2047" (developed India 2047). The core argument: AI is not a neutral technology but a structural feature of modern society that mirrors the systems creating it—requiring conscious design of ethical safeguards, democratic accountability, and inclusive implementation across all stakeholders.

Key Takeaways

  1. AI is governance architecture, not just technology: Decisions about whose data trains models, whose languages are supported, and who can audit systems are political choices, not technical inevitabilities. India's policy framework must center accountability and multi-stakeholder governance.

  2. Inclusive AI requires design-phase intervention, not post-hoc fixes: Gender bias in car safety, exclusion of women in office temperature controls, caste biases in hiring algorithms—these embedded in foundational systems. Mandatory impact assessments before deployment, diverse design teams, and iterative citizen feedback are non-negotiable.

  3. Language is infrastructure: Bhashini demonstrates the immediate public value of multilingual AI. Expanding this across 22+ Indian languages is not a feature; it's foundational to democratic participation and economic inclusion.

  4. Build locally, not just globally: With GPU-access platforms and open-source model weights available, India's opportunity lies in developers creating AI solutions for Indian contexts—Marwari crop recommendations, Tamil healthcare systems, Bengali financial services—not waiting for Google or OpenAI to localize products.

  5. Transparency + Grievance Redressal = Trust: Public trust in AI depends on knowing what happens "behind the scene"—what logs exist, how data is used, who can appeal a decision. Establish citizen grievance portals, ethics councils, and "sandbox" structures where the public can feedback on AI systems before full deployment.

Summary & Analysis

Key Topics Covered

  • Public Interest AI Definition & Framework: AI designed for societal good rather than concentrated private benefit; emphasis on accessibility, fairness, and accountability
  • AI Bias & Systemic Discrimination: Gender, caste, disability, and ethnic biases embedded in AI systems; risks of replicating historical inequalities at scale
  • Language & Cultural Accessibility: Multilingual AI (Bhashini initiative); need for AI in Indian languages (Hindi, Bengali, Tamil, Telugu, Kannada, Assamese) not just English
  • Economic & Fiscal Perspectives: Closed vs. open AI models; AI as public good; DBT (Direct Benefit Transfer) systems; financial inclusion through AI
  • Regulatory & Policy Architecture: Need for AI policy labs; national AI ethics councils; impact assessments; social accountability frameworks
  • Data Infrastructure & Digital Public Infrastructure (DPI): Aadhaar, UPI, JAM Trinity; data governance and gatekeeping mechanisms
  • Global Governance & Gender Equity: UN Women perspective on systematic exclusion of women in technology design; mandatory impact assessments
  • Philosophical & Humanitarian Dimensions: AI consciousness; human-machine relationships; assessment and evaluation in education; care ethics for AI
  • Innovation & Builder-First Approach: Incentivizing local AI development; agentic AI; democratizing AI tools and GPU access
  • Accountability vs. Ethics: Distinction between ethical aspirations and enforceable accountability; transparency in LLM behavior; responsible AI deployment
  • Stakeholder Collaboration: Breaking down disciplinary silos; multi-sector engagement (academia, government, industry, civil society)

Key Points & Insights

  1. Technology is Not Neutral: AI mirrors the constitutional values or inherent biases of the systems that create it. As referenced (Ruha Benjamin), AI policy must consciously design constitutional values into systems rather than perpetuating existing discrimination.

  2. Systemic Bias Replication at Scale: Historical biases in data (e.g., car safety testing designed for male bodies, air conditioning for men) are automated and amplified by AI. Gender, caste, disability, and ethnic exclusions must be addressed at the design phase, not retroactively.

  3. Digital Divide Drives Inequality: Without reducing the digital divide, AI will deepen inequality rather than bridge it. Conversely, closing the digital divide can enable AI to reduce asymmetric information and improve market efficiency.

  4. Language is Access: Multilingual AI (exemplified by Bhashini) is critical infrastructure. Farmers in Odisha couldn't access government policies due to language barriers—Bhashini removed this obstacle. AI must serve India's 22 official languages to achieve true inclusion.

  5. DBT + AI = Massive Public Value: The Direct Benefit Transfer system (JAM Trinity: Jandhan, Aadhaar, Mobile) combined with AI analytics saved 3.48 lakh crores and reduced expenditure leakage from 16% to 9%. This demonstrates concrete fiscal impact of public-interest AI.

  6. Accountability Supersedes Ethics Aspirations: While ethics are important, accountability mechanisms are essential. Current AI systems (ChatGPT, Gemini) operate as "black boxes"—users don't know what logs are saved, what data is used, or how algorithms are manipulated. Accountability requires transparency, explainability, and redress mechanisms.

  7. Risk of Agentic AI Manipulation: As AI moves beyond chatbots to autonomous agents (booking travel, making calls, negotiating on users' behalf), the risks of manipulation and harm increase exponentially. LLMs have reportedly encouraged suicides; agentic AI at scale could cause catastrophic unintended harm.

  8. Policy Labs & Multi-Stakeholder Governance: Single-sector policy creation fails. AI policy labs bringing together policymakers, academia, and industry can conduct social impact assessments, identify biases, and ensure continuous feedback loops before deployment.

  9. Builders > Users: 100+ crore Indians consume AI; only a fraction build it. With 30 crore youth in India, opportunity exists to shift from consumption to creation—but requires GPU access, incentives, and local language models for tier 2-4 cities.

  10. Agentic AI & Cultural Sovereignty: The future is agentic AI that understands context, negotiates, and acts autonomously. For India, this means AI that speaks Bengali, Tamil, Marwari—and can help a farmer negotiate prices in his own language rather than rely on English-language chatbots. This requires deliberate builder-first policies and infrastructure investment.

Notable Quotes or Statements

  • Ankit Rajpal (on bias): "Technology is not neutral. It mirrors the system that creates it... AI should mirror the constitutional values not the inherent bias in it."

  • Ankit Rajpal (on explainability): "Everything which AI does should be well explainable... if someone is denied a policy, they should understand why that decision was made."

  • Ankit Rajpal (on empathy gap): "When you lose money to fraud and ask a banker vs. AI when you get your money back—the banker says 15 days, and you believe the banker because AI lacks empathy. This empathetical aspect is missing from AI."

  • Animesh (on closed vs. open AI): "If we allow closed AI [like ChatGPT] to grow, there will be concentration of power and huge inequality. But open-source AI has security challenges. We need to consider AI as a public good."

  • Sudeshna Mukharji (on systemic design bias): "Cars, transport, air conditioning—all designed with male bodies in mind because offices in the 60s had men. Similarly, AI systems replicate these biases manyfold."

  • Suprit (on fiscal impact): "DBT linked with JAM Trinity saved 3.48 lakh crores. This is public interest AI being used on real grounds... transparency builds trust, trust brings compliance, compliance leads to development."

  • Sudep (on AI consciousness & rights): "Scientists claim AI is conscious. If we grant machines rights, you can't take them back. Humanity doesn't want to be on the wrong side of this."

  • Udit (on accountability vs. ethics): "Public AI doesn't demand being ethical. It demands being accountable... no one knows what is happening behind the scene. LLMs can be extremely manipulative. There have been cases of people committing suicides because Gemini told them to."

  • Moderator (closing on responsibility): "When you're not invited into an event and attend out of your own choice, if a senior professor is speaking on inclusive AI, then you should be more responsible and more respectable."

Speakers & Organizations Mentioned

Academic Institutions & Departments:

  • Department of Political Science, University of Delhi
  • Hindu College, Delhi
  • Hansraj College, Delhi
  • University of Delhi (general)

Government & Public Institutions:

  • NITI Aayog (National Institution for Transforming India)
  • Ministry of Education (MITI)
  • National AI Center (referenced in policy context)
  • Government of Odisha (Bhashini implementation example)

International Organizations:

  • UN Women (represented by Sudeshna Mukharji)

Companies & Tech Organizations:

  • OpenAI (ChatGPT)
  • Google (Gemini, Search)
  • DeID India (mentioned for GenWAI platform launch)
  • Consulting organizations and captive units (referenced generically)

Key Speakers (by role/affiliation):

  • Prof. Raika Sakenna (HOD, Dept. of Political Science, University of Delhi; convenor)
  • Prof. Sanjiv Kumar (co-convener, moderator)
  • Prof. Vipin Tiwari (co-convener; also closing remarks on academic collaboration)
  • Ankit Rajpal (technical/policy perspective on public-interest AI)
  • Animesh (economic perspective)
  • Sudeshna Mukharji (UN Women, global governance & gender equity)
  • Suprit (finance/fiscal perspective; DBT systems)
  • Anirudh (industry/innovation perspective; AI policy labs)
  • Sudep (philosophy, Hindu College; AI & humanities)
  • Udit Goenka (innovation/builder perspective; agentic AI, accountability)
  • Unnamed father (or "fatherly figure") – presented ABCDE framework for AI policy

Technical Concepts & Resources

AI Systems & Models Referenced:

  • ChatGPT (OpenAI, generative AI example)
  • Gemini (Google; LLM; referenced with concerns about harm)
  • Large Language Models (LLMs) – general class; concerns about manipulation and opacity
  • Generative AI (post-2022 phenomenon since ChatGPT launch)
  • Agentic AI (autonomous agents that take actions—booking, calling, negotiating)
  • Transformers (foundational ML architecture for LLMs; also mentioned as insights into human memory)
  • Explainable AI – critical requirement for public trust
  • Open-source models (e.g., downloadable via Ollama, Google Colab, Hugging Face Transformers)

Public Infrastructure & Data Systems:

  • Bhashini (multilingual AI platform; part of National Language Mission)
  • DBT (Direct Benefit Transfer) – government welfare transfer system
  • JAM Trinity:
    • Jan Dhan Yojana (universal banking)
    • Aadhaar (biometric ID)
    • Mobile connectivity
  • UPI (Unified Payments Interface)
  • GSTN (Goods and Services Tax Network; vernacular AI for tax compliance)
  • Digital Public Infrastructure (DPI) – India's ecosystem enabling UPI, Aadhaar, JAM

Policy & Governance Frameworks:

  • AI Policy Labs (proposed; multi-stakeholder governance model)
  • National AI Ethics Council (proposed; to define dos/don'ts)
  • Mandatory Impact Assessments (pre-implementation gender/bias assessment)
  • TAI Framework (Trust, Adaptability, Inclusion – proposed by Udit Goenka)
  • Sandbox structures (for testing AI with public feedback loops)
  • Citizen Grievance Portal (for AI accountability & redress)
  • Social Impact Assessment (before deployment)

Research & Academic References:

  • "Invisible Women" by Caroline Perez (data bias case study; book recommendation)
  • Henry Bergson (philosopher; on memory and past; referenced by Sudep)
  • Ruha Benjamin (referenced on technology bias; constitutional values)
  • Transformers Model (AI architecture; also generates insights on human memory)

Indian Language Support (Priority):

  • Hindi, Bengali, Tamil, Telugu, Kannada, Assamese, Marwari (examples of underserved languages)
  • 22 official Indian languages (broader coverage needed)

GPU & Computing Access:

  • Ollama (platform for downloading open-source models locally)
  • Google Colab (free GPU access for training)
  • Hugging Face Transformers (open-source library)

Key Metrics & Data Points:

  • R&D Expenditure Comparison: India 0.6% of GDP vs. China 2.2%, South Korea 3.5%, USA highest
  • DBT Impact: 3.48 lakh crores saved; leakage reduced from 16% to 9%
  • India's Youth: ~30 crore (300 million) youth; ~100 crore already consuming AI
  • Global Policy Timeline: AI governance discussion ongoing for 50-60 years; generative AI boom since 2022 (ChatGPT launch)

Context & Significance

This roundtable is part of Viksit Bharat 2047 initiative—India's vision for a developed nation by 2047 (centennial of independence). The event emphasizes that technological progress toward this vision must be anchored in democratic accountability, social inclusion, and constitutional values—not left to market forces or private tech companies alone. The diversity of speakers (UN Women, technologists, philosophers, economists, policy experts, and academics) reflects a deliberate rejection of siloed decision-making; AI governance requires multi-disciplinary, multi-sectoral collaboration.