India AI Impact Summit 2026

responsible-ai-in-practice-a-data-perspective

8 min read

TRACK:Safe and Trusted AI
SPEAKERS
  • Dr. Richa Singh, IIT Jodhpur
  • Dr. Shruti Nagpal, Sony AI
  • Ms. Aprajita Rana, AZB & Partners
  • Shri Nitendra Rajput, AI Garage, MasterCard
SESSION OVERVIEW

The session will begin with a 15-minute introduction outlining the research we have been leading at Sony AI, after which we will open into a moderated panel discussion. The aim is to examine how responsible AI can be meaningfully advanced by addressing data practices, evaluation standards, and governance considerations.

VIDEO RECORDING

Responsible AI in Practice: A Data Perspective

Abstract

The session opened with Dr. Shruti Nagpal presenting Sony AI’s new “PHEBE” dataset – a globally diverse, consent‑driven, GDPR‑compliant benchmark for human‑centred computer‑vision tasks. She highlighted the core pillars of responsible AI (consent, revocation, diversity, copyright protection and fair compensation) and demonstrated how the dataset operationalises those principles. A moderated panel then explored what “responsible AI” means across domains, covering legal, academic and industry viewpoints. Topics ranged from data‑governance, model‑card standards and unlearning, to organisational processes for fairness, security‑by‑design, and the tension between performance and ethical safeguards. The discussion concluded with a call for structural changes – from school‑level curricula to dedicated “Chief Responsible‑AI Officer” roles – to embed responsibility throughout the AI lifecycle.

Detailed Summary

1.1 Framing the Problem

  • Shruti opened by noting the proliferation of the term responsible AI at the summit and argued that the conversation must move from buzzwords to concrete data practices.
  • She positioned fairness not only as a model‑level concern but as something that can (and should) start with the underlying dataset.

1.2 Core Pillars of Responsible AI (Sony AI Perspective)

PillarRationale & Implementation
Consent (and revocation)Data contributors must be able to withdraw consent at any time. Current tools exist but are not yet practical; Sony AI stresses making revocation enforceable.
Globally diverse representationDatasets must capture the full spectrum of human variation to avoid demographic bias.
Copyright protectionPrevent unauthorized reuse of contributed data.
Fair compensationContributors, annotators and pipeline workers should receive just remuneration; this is often omitted in AI pipelines.

1.3 Introducing the PHEBE Dataset

  • Full name: Fair Human‑Centric Image Benchmark (PHEBE).
  • Scope: 1 dataset enabling nine distinct computer‑vision tasks (e.g., face‑recognition, person‑segmentation, pose estimation).
  • Scale & Richness: Each image carries 40+ annotations covering:
    • Subject‑level attributes (skin tone, facial hair, hair colour, multiple simultaneous features).
    • Environmental context (indoor/outdoor, lighting source).
    • Device metadata (camera model, focal length).
  • Ethical provenance:
    • All images collected with explicit, documented consent.
    • Annotators were fairly compensated and “quality‑checked” by a dedicated QA team.
    • GDPR‑compliant; publicly released with a data‑card outlining provenance, consent status and usage limits.

1.4 Vision & Call to Action

  • The dataset is positioned as a blueprint for ethical data collection, a test‑bed for bias diagnosis, and a practical resource for researchers.
  • Shruti shared a QR code and a download link, encouraging attendees to experiment with PHEBE and adopt its principles in their own pipelines.

2. Panel Introduction

  • Moderator (Shruti) introduced the panelists:
    • Ms. Aprajita Rana (partner, AZB & Partners) – specialist in AI technology law and data governance.
    • Shri Nitendra Rajput (VP, AI Garage, MasterCard) – referred to as “Ankur” during the discussion.
    • Dr. Mayank [Last Name not supplied] – academic researcher from IIT Jodhpur (the transcript labels him “Dr. Mayank”).
  • A brief note that a fourth listed speaker (Dr. Richa Singh) was not audible in the recording.

3. Defining “Responsible AI” – Multi‑Domain Perspectives

3.1 Industry View (MasterCard) – Shri Nitendra Rajput

  • Core mantra: “Security‑by‑design” (fairness, transparency, accountability, compliance).
  • Fairness – Models must not discriminate on protected attributes (race, gender, income).
  • Transparency – Explainability is mandatory; users have a right to know why a loan or transaction was declined.
  • Security & Data Integrity – Data must be sourced from verified, authorized origins; custodians must prevent leaks.
  • Regulatory compliance – Must align with regional frameworks (e.g., EU AI Act).
  • Process – MasterCard’s Studio Process with multiple “gates” (problem definition, data source vetting, bias checks) ensures responsible AI is embedded early and continuously rather than as a post‑hoc review.

3.2 Legal Perspective – Ms. Aprajita Rana

  • Data‑governance gap – India’s past lack of a robust data‑protection law led to widespread use of public data for AI training; recent Data Protection & Privacy (DPTP) Act now obliges explicit consent, purpose limitation, and right‑to‑erasure.
  • Compliance challenges – Companies often lack mechanisms for “data unlearning” (removing a contributor’s data from already‑trained models).
  • Corporate accountability – Boards and CEOs need clear guidance on AI risk, especially in M&A deals involving AI assets.
  • Regulatory approach – A mix of voluntary standards (industry best practices) and binding legislation for high‑risk sectors (finance, health, insurance) is recommended.

3.3 Academic View – Dr. Mayank

  • Design‑first mindset – Responsible AI should be posed as a question at the problem‑statement level: “Can this problem be solved responsibly?”
  • Curriculum evolution – IIT Jodhpur has introduced a dedicated Dependable AI course; a new ML‑DL Ops class teaches pre‑ and post‑hoc responsible‑AI techniques.
  • Future‑ready education – Plans to embed AI ethics from class 11‑12 textbooks (targeting ~20 million students annually).

4. Data Governance & Legal Compliance

  • Consent & Revocation – Both Shruti (Sony) and Aprajita (AZB) stressed that consent must be revocable in practice, not only on paper.
  • Right to Erasure (DPTP Act) – Mayank highlighted the need for model unlearning mechanisms so that a data‑subject’s request can be honoured even after model training.
  • Model & Data Cards – The panel agreed that publishing standardized model cards (architecture, training data provenance) and data cards (collection process, consent status) is essential for transparency and auditability.
  • Legal Cases – Reference to the ongoing ANI vs. OpenAI litigation in India underscores the rising enforcement of data‑rights.

5. Model Governance & Organizational Processes

  • Security‑by‑Design – MasterCard’s gated studio process forces teams to address fairness, bias, and compliance before model development proceeds.
  • Cross‑functional Collaboration – Data‑strategy, AI‑governance, product, and business teams work together from day one; silos are discouraged.
  • No Trade‑off Policy – MasterCard insists that performance deadlines must factor in governance approvals; there is no compromise on responsible AI.
  • Emerging Roles – The discussion noted a trend toward titles like Chief Responsible‑AI Officer and Responsible AI Scientist, reflecting the need for dedicated stewardship.

6. Industry Challenges – Fairness vs. Performance

  • Tension points – While fairness is non‑negotiable, pressures to deliver fast can surface. MasterCard’s studio gates aim to pre‑empt such trade‑offs.
  • Stakeholder alignment – Shareholder interests (profit, speed) must be balanced against ethical safeguards; the panel stressed that ethical metrics should be part of KPIs.
  • Accountability in the supply chain – Aprajita warned that downstream vendors often lack clarity on the responsible‑AI standards they must follow, creating legal exposure for the principal company.

7. Voluntary Standards vs. Regulation

  • Voluntary Frameworks – Industry groups are already drafting governance models, explainability norms, and auditing practices (especially for generative AI).
  • Regulatory Necessity – High‑risk domains (healthcare, finance, insurance) require statutory accountability; otherwise, consumers lack recourse.
  • Hybrid Approach – The consensus: combination of robust regulations for critical sectors and flexible, industry‑driven standards for rapidly evolving technologies like Gen‑AI.

8. Foundation Models, Opacity, and Unlearning

  • Opaque Training Data – Mayank argued that responsible AI is still aspirational when training data remains undisclosed.
  • Model‑Card Adoption – Few presenters at the summit displayed model cards; the panel urged community-wide adoption.
  • Unlearning Challenges – Implementing true data‑subject erasure requires technical capabilities beyond current practice; research on machine unlearning is still nascent.

9. Structural Changes Needed to Move from Principle to Practice

Proposed ChangeRationale
Integrate AI ethics into school curricula (class 11‑12)Early exposure builds a generation that questions AI use and understands ethical constraints.
Create dedicated “Chief Responsible‑AI Officer” rolesEmbeds accountability at the executive level; clarifies governance responsibilities.
Mandate continuous responsible‑AI checkpoints throughout the product lifecycle, not just at launch.Guarantees that new data, model updates, and deployments stay aligned with ethical standards.
Standardize and require public model‑card/data‑card releases for any publicly released model.Increases transparency, facilitates external audits, and supports regulatory compliance.
Develop clear legal frameworks for AI accountability (e.g., who is liable in M&A, vendor contracts).Reduces ambiguity for corporations and protects end‑users.

10. Closing Remarks & Announcements

  • Token of Appreciation – Sony AI’s Director presented a gratitude token to all participants.
  • Group Photo – Attendees were asked to stay for a collective photograph.
  • Future Outlook – The moderator reiterated that responsible AI cannot be retro‑fitted; it must be embedded at design time and continuously reinforced.

Key Takeaways

  • Responsible AI must start at the data level; fairness, consent, diversity and fair compensation are inseparable from dataset design.
  • Sony AI’s PHEBE dataset offers a concrete, GDPR‑compliant benchmark with rich multi‑task annotations, exemplifying ethical data collection.
  • Security‑by‑design and early‑stage governance gates (MasterCard’s “studio process”) are effective ways to prevent ethical shortcuts later in the pipeline.
  • Legal compliance in India is tightening (DPTP Act, right‑to‑erasure); companies need operational unlearning capabilities to honor data‑subject requests.
  • Cross‑disciplinary collaboration (lawyers, technologists, academics) is essential; each brings unique constraints and solutions to the responsible‑AI puzzle.
  • Voluntary standards are valuable for fast‑moving areas like generative AI, but regulation remains crucial for high‑risk sectors.
  • Transparency tools – model cards, data cards, and documentation of consent/compensation – must become mandatory artifacts for all released AI systems.
  • Education is a structural lever: embedding AI ethics in secondary‑school curricula will nurture future practitioners who internalise responsibility from the outset.
  • Organisational roles dedicated to AI responsibility (e.g., Chief Responsible‑AI Officer) help institutionalise accountability across the product lifecycle.
  • Continuous responsibility is a lifecycle commitment, not a one‑off checklist; ongoing monitoring, updates, and stakeholder engagement are required to keep AI systems trustworthy.

See Also:

Graph View

Backlinks