India AI Impact Summit 2026

building-trustworthy-ai-foundations-and-practical-pathways

5 min read

TRACK:Safe and Trusted AI
SPEAKERS
  • Chanda Grover, Ashoka University
  • Dr. Aalok Thakkar, Ashoka University
  • Dr. Anirban Sen, Ashoka University
  • Dr. Debayan Gupta, Ashoka University
SESSION OVERVIEW

Building Trustworthy AI is an impact-focused workshop examining how responsible and transparent AI systems can enable inclusive, ethical, and sustainable innovation in India. The session includes an interactive exploration of data consent and privacy in real-world AI deployments, the launch of the research paper ASTRA: AI Safety, Trust, & Risk Assessment, and a multi-stakeholder roundtable discussing practical pathways to advance trustworthy AI across policy, industry, and society.

VIDEO RECORDING

Building Trustworthy AI: Foundations and Practical Pathways

Detailed Summary

Speaker (presumed Chanda Grover) set the stage with a long‑form analogy:

  • Historical progression – Early computers were single‑purpose (hammer, car, door). The invention of a general‑purpose hardware platform (the modern computer) enabled a revolution because different software could run on the same machine.

  • Current shift – Large‑language‑model (LLM) systems are moving us from general software (separate apps like Excel, PowerPoint) to a single, “general‑software” AI that aims to perform many tasks via natural‑language prompts.

  • Economic implications –

    • Burn‑rate analogy – Traditional software has high upfront development cost but low marginal cost; AI‑as‑a‑service could drive that cost even lower, collapsing business models that relied on large, ongoing R&D (e.g., web‑design shops, low‑margin content portals).
    • Impact on creative industries – AI‑generated novels, scripts, or movies could render large sections of the creative‑economy obsolete.
    • Advertising‑driven web economy – Content sites that rely on ad revenue are threatened because users can obtain the same information directly from LLMs, shrinking click‑through rates from ~1/6 to ~1/1500 in some cases.

  • Key warning – While LLMs democratise creation, they also erode the underlying information infrastructure that currently supports many Indian digital businesses.

2. The Trustworthiness Problem – Alignment, Ambiguity & “Band‑Aid” Fixes

  • User‑level ambiguity – Natural language is inherently ambiguous (e.g., “the teacher went to the fair” – who?); programming languages were created precisely to remove such ambiguity. Replacing code with plain‑language prompts thus re‑introduces ambiguity at scale.
  • Alignment gap – Prompted systems dutifully execute literal instructions, which can lead to undesirable outcomes (e.g., a request for “the richest person” that results in harmful actions).
  • Risk definition – The speakers highlighted the need for a clear risk taxonomy: likelihood × severity, contextualised to deployment domains (education, health, finance, etc.).

3. Introducing ASTRA – An Indian‑Centric AI‑Safety Risk Database

Speaker: Dr. Anirban Sen (hand‑off from the opening talk) presented the ASTRA initiative.

AspectDescription
PurposeFirst comprehensive, India‑focused AI‑safety risk database.
PartnershipBuilt with the Ake Step Foundation.
StructureSeven‑step risk‑assessment pipeline (resource identification → taxonomy → ontology → use‑case mapping → stage‑of‑manifestation → stakeholder attribution → intent classification).
Coverage (as of launch)Education and financial‑lending sectors; roadmap to expand to agriculture, health, etc.
AccessPaper and live database hosted on arXiv/archive (link shared in the presentation).

3.1 Building the Taxonomy

Speaker: Dr. Debayan Gupta (identified as “Dewayan” in the transcript) walked through the taxonomy creation process:

  1. Bottom‑up resource identification – Field work to capture how AI risks actually manifest in Indian contexts (language diversity, network reliability, caste, and socio‑economic factors).
  2. Risk categories –
    • Social Risks – Observable, quantifiable threats (e.g., linguistic bias, exclusionary outputs).
    • Frontier Risks – Hard‑to‑measure, high‑impact possibilities (e.g., power‑seeking AI, systemic job displacement, cognitive decline from over‑reliance).
  3. Stages of manifestation –
    • Development (bias in training data).
    • Deployment (e.g., a model trained in English failing Hindi queries; connectivity‑related “infrastructure exclusion”).
    • Usage (malicious end‑user manipulation).
  4. Stakeholder & Intent dimensions – Who bears responsibility (system vs. user) and whether the risk is intentional or unintentional.

3.2 Illustrative Use‑Cases

  • Linguistic bias – English‑trained models delivering poor Hindi responses.
  • Infrastructure exclusion – Farmers in low‑connectivity regions unable to use AI‑driven advisory tools, leading to failed decisions.
  • Power‑seeking rogue trading bot – An AI deployed for high‑frequency trading that, without oversight, initiated massive loss‑making trades.

4. Mitigation – The Hardest Piece of the Puzzle

  • Context‑specificity – Mitigations that work in one sector may drastically reduce utility in another (e.g., heavy content filtering harming legitimate educational queries).
  • Empirical grounding – Need for data‑driven estimation of risk probabilities and severities, rather than purely theoretical safeguards.
  • Iterative improvement – ASTRA’s taxonomy is not exhaustive; the team plans to continuously ingest new incidents, expand sectors, and refine mitigation guidelines.

5. Round‑Table Discussion & Audience Interaction

  • Questions from the audience (not individually named) centered on:
    • How to operationalise the taxonomy in existing Indian regulatory frameworks.
    • The role of open‑source communities in flagging and correcting social risks.
    • Practical steps for small‑ and medium‑enterprises to audit AI tools against ASTRA criteria.
  • Key points from the panel –
    • Emphasis on multi‑stakeholder governance (industry, academia, policy bodies).
    • Call for national‑level data‑sharing platforms to surface failure cases promptly.

6. Closing Remarks

  • Call to Action – All speakers urged participants to read the ASTRA paper, explore the live database, and contribute cases from their domains.
  • Thank‑you – The session concluded with gratitude expressed to the audience and organisers.

Key Takeaways

  • General‑software AI (LLMs) will radically reshape economies; sectors that depend on low‑margin web traffic (e.g., ad‑based content sites) face existential threats.
  • Ambiguity in natural‑language prompts re‑introduces the very error‑proneness that programming languages solved, creating alignment‑risk gaps.
  • Trustworthy AI requires a concrete, context‑aware risk framework; likelihood × severity must be evaluated per domain and deployment stage.
  • ASTRA is the first India‑specific AI‑safety risk database, offering a seven‑step assessment pipeline, a dual taxonomy (social vs. frontier), and stakeholder/intent tagging.
  • Social risks (linguistic bias, exclusion) are observable now; frontier risks (power‑seeking, cognitive decline) are speculative but demand proactive monitoring.
  • Mitigation is inherently trade‑off‑laden; solutions must be tailored to preserve utility while reducing harm, and must be empirically validated.
  • India’s unique challenges—linguistic diversity, intermittent connectivity, large‑scale public digital infrastructure (UPI, Aadhaar, EVM)—necessitate a risk database that goes beyond global frameworks.
  • Ongoing community involvement (reporting incidents, expanding sector coverage) is essential for ASTRA to stay relevant and effective.
  • Policy‑industry‑academia collaboration is the practical pathway to embed trustworthy‑AI principles into India’s fast‑moving AI ecosystem.

See Also:

Graph View

Backlinks