India AI Impact Summit 2026

safe-ai-building-shared-trust-and-accountability-infrastructure

7 min read

TRACK:Safe and Trusted AI
SPEAKERS
  • Dr Akash Kapur, Princeton University
  • Dr Ponnurangam Kumaraguru, IIIT-Hyderabad
  • Mr Nitarshan Rajkumar, Anthropic
  • Ms Deepika Mogilishetty, EkStep Foundation
  • Ms Urvashi Aneja, Digital Futures Lab
SESSION OVERVIEW

As AI systems are deployed at population scale, ensuring trust, safety, and accountability becomes critical. This session focuses on how AI safety principles can be translated into operational governance frameworks, institutional mechanisms, and shared infrastructure that enable responsible AI adoption while protecting users and communities.

VIDEO RECORDING

Safe AI: Building Shared Trust and Accountability Infrastructure

Abstract

The panel examined why “good intentions” alone cannot guarantee safe AI at population scale. The discussion moved from high‑level reflections on how safety principles can be lost in translation, to concrete Indian challenges such as multilingual and code‑mixed communication, to institutional and regulatory mechanisms that can embed continuous accountability. Panelists highlighted the tension between rapidly evolving private incentives and the slower, democratic processes needed for public‑good safeguards, and they offered concrete ideas – from open‑protocol standards to user‑centric consent redesigns – that could form shared infrastructure for responsible AI deployment.

Detailed Summary

  • Moderator (Deepika) introduced the panel, noted two invited speakers were absent, and invited the audience to take a quick photo.
  • She set the agenda: moving AI safety from abstract principles to operational governance that can be applied at scale.
  • Emphasised that “good intentions don’t scale” and that the panel would explore what infrastructure and accountability mechanisms are required.

2. Translating Safety Principles to Real‑World Contexts

2.1 Akash Kapur on “lost in translation”

  • Stated that when technologies leave the lab, they confront demographic, cultural, and linguistic complexity.
  • Recalled his PhD fieldwork on ICT4D kiosks in Tamil Nadu, where caste‑based spatial placement determined who could actually use the service.
  • Argues that AI’s promise of universal knowledge is filtered through social structures, creating inclusion/exclusion risks that are integral to safety.
  • Introduced the temporal dimension of safety: systems are released into a configuration that evolves (e.g., the early open Internet vs. today’s walled gardens). Safety therefore requires ongoing audits and certification, not a one‑off checklist.

2.2 The “safe” terminology debate

  • Deepika asked whether “safety” has become a catch‑all buzzword.
  • Kapur responded that early AI safety was synonymous with existential risk; now the field also covers harms, bias, exclusion, and misuse.
  • He suggested that “responsible AI” may be a more precise umbrella term.

3. India‑Specific Challenges

3.1 Multilingual & Code‑Mixed Reality (PK)

  • PK surveyed the audience to gauge familiarity with multilingualism.
  • Highlighted India’s language diversity (Telugu, Tamil, Marathi, Hindi, etc.) and the prevalence of code‑mixing in everyday communication (e.g., switching between English and regional languages within a single sentence).
  • Noted that while language‑specific models exist for Hindi, Marathi, Telugu, code‑mixed data remain largely unsupported.

3.2 Concrete Safety Scenarios

  • Described EkStep’s chatbot deployment for farmers asking agronomic questions.
  • Illustrated a problematic interaction: a farmer asked the bot to draft a petition to the Maharashtra chief minister; the model complied, generating a politically charged document.
  • Raised the question: Should the bot have been allowed to generate such content? – a direct safety dilemma.

3.3 Bottom‑Of‑the‑Pyramid (BOP) deployment

  • Emphasised that BOP users often rely on voice interfaces and have limited digital literacy.
  • Lack of robust voice‑language support for multilingual contexts complicates safety monitoring.

4. Institutional Accountability & Incentive Alignment

4.1 Competing Stakeholder Interests

  • Moderator asked how to craft frameworks that manage divergent incentives across government, private firms, and civil society.
  • Kapur noted that private incentives can sometimes align with public good – e.g., geopolitical competition spurred the release of open‑source AI models.

4.2 Open‑Source Public Goods (Nitarshan)

  • Introduced the Machine‑Centric Protocol (MCP) – an open protocol for “agentic AI” akin to TCP/IP for the internet.
  • MCP was donated to the Linux Foundation, showing a concrete public‑good contribution from Anthropic.
  • Emphasised that standardised interfaces can reduce fragmentation and enable cross‑company interoperability, lowering systemic risk.

4.3 Need for Regulation

  • Kapur argued that unregulated private AI development risks repeating Internet‑era failures (e.g., lack of public‑interest prioritisation).
  • He appealed to democratic institutions to channel public values (equality vs. freedom, efficiency vs. justice) into AI governance.
  • Stressed that democracy already regulates highly complex domains (pharma, nuclear safety); AI should be treated similarly.

5. Existing & Proposed Regulatory Mechanisms

5.1 Self‑Regulation

  • PK described self‑regulatory codes (e.g., Data Security Council of India), where firms voluntarily adopt best‑practice pledges.
  • Noted that such initiatives have limited enforcement power and may lack impact at scale.
  • Deepika highlighted the terms‑and‑conditions fatigue problem: users habitually click “Agree” without reading.
  • Discussed parallels with pharma’s consent sheets – historically, regulators, not consumers, ensured safety.
  • Suggested re‑designing privacy policies and user agreements into concise, understandable “stickers” (e.g., using visual icons, plain‑language summaries).

5.3 Auditing & Certification Frameworks

  • Kapur advocated for periodic safety audits that are process‑oriented, not a one‑off checklist.
  • Suggested a DPI‑style (Data Protection Impact) approach for AI: systematic impact assessments before deployment and continuous monitoring thereafter.

6. Over‑Zealous Guardrails – A “Safety‑of‑Safety” Issue

  • In the closing minutes, Kapur warned that over‑restrictive guardrails can create new harms.
  • Example: Gemini (a large language model) was instructed to refuse any political‑information queries, leading it to block legitimate citizen queries about polling locations.
  • Such over‑blocking can disenfranchise users, especially in contexts where civic information is essential.

7. Synthesis – Towards Shared Building Blocks

7.1 Process‑Based Safety

  • Safety should be viewed as a continuous process: regular checks, impact assessments, and updates reflecting changing incentives and contexts.

7.2 Cross‑Cultural Templates

  • Develop modular safety templates focusing on the application layer (e.g., UI, data flow) rather than only the model layer.
  • Templates can be localised (language, cultural norms) while preserving a core methodology.

7.3 Evaluation Tools & Audits

  • Propose standardised evaluation suites (bias, misinformation, privacy leakage) that can be re‑run with every major update.

7.4 Multi‑Stakeholder Governance

  • Encourage joint working groups comprising regulators, industry, academia, and civil‑society representatives to co‑design accountability frameworks.

8. Audience Interaction & Illustrative Polls

  • The moderator ran several quick audience polls (e.g., number of apps downloaded, whether participants read T&C, languages spoken) to underline real‑world behaviour and awareness gaps.
  • These polls reinforced the panel’s point that user‑level practices (ignoring consent, low digital literacy) amplify safety risks.

9. Closing Remarks & Open Questions

  • Open Question 1: How can code‑mixed language models be built and evaluated reliably for safety?
  • Open Question 2: What light‑weight regulatory structures can be instituted in resource‑constrained settings like India?
  • Open Question 3: How to balance guardrails that prevent misuse without over‑blocking legitimate user needs?
  • Open Question 4: What concrete public‑good protocols (beyond MCP) can be nurtured to standardise AI interaction across ecosystems?

Final Takeaway: The panel concluded that collaboration across sectors—rather than finger‑pointing—will be essential to create shared, adaptable safety infrastructure that respects user agency while ensuring accountability.

Key Takeaways

  • Safety is a process, not a product. Continuous audits, temporal monitoring, and iterative certifications are required for AI systems deployed at scale.
  • Cultural and linguistic contexts matter. India’s multilingual, code‑mixed reality poses unique safety challenges that generic models cannot address.
  • Open protocols (e.g., MCP) demonstrate public‑good alignment and can reduce fragmentation, fostering safer ecosystems.
  • Self‑regulation alone is insufficient. Robust, democratically‑backed regulatory frameworks are needed to prioritize public interest over pure profit motives.
  • User consent mechanisms must be redesigned into plain‑language, visual “stickers” to avoid the “click‑agree” habit.
  • Over‑restrictive guardrails can backfire (e.g., political‑information blocks), creating new harms—guardrails need nuanced, context‑aware design.
  • Bottom‑of‑the‑pyramid deployments amplify safety risks due to limited literacy and reliance on voice interfaces; targeted interventions are crucial.
  • Multi‑stakeholder governance (government, industry, academia, civil society) is the most promising path to shared accountability.
  • Key research gaps remain in evaluating safety for code‑mixed language models, designing lightweight regulatory tools for low‑resource settings, and creating universally‑applicable safety templates.

End of summary.

See Also:

Graph View

Backlinks