Astra: AI Safety, Trust, & Risk Assessment

1. Problem Statement

The paper addresses the critical gap in applying global AI safety frameworks to India's unique socio-technical landscape. While India has successfully deployed large-scale Digital Public Goods (DPGs) like Aadhaar and UPI, the rapid integration of AI and Machine Learning (ML) into these systems introduces risks that existing Western-centric taxonomies fail to capture.

Key Challenges Identified:

• Contextual Blindness: Global frameworks (e.g., EU AI Act, MIT Risk Repository) often ignore India's specific realities, such as caste-based discrimination, linguistic diversity, low digital literacy, and a massive informal economy.
• Definition of Risk: Standard definitions of "risk" often fail to account for "opportunity costs" (failure to achieve public interest goals) and "techno-colonialism" (dependency on foreign APIs/compute).
• Nature of AI Failures: Unlike classical engineering, ML systems lack formal error bounds. They suffer from hallucinations, distributional shifts, and emergent behaviors that are difficult to predict using historical data (unlike actuarial risks).
• Lack of Empirical Data: There is no existing, empirically grounded database of AI safety risks specifically for Indian DPGs, leading to superficial engagement with terms like "hallucination" or "bias" without rigorous analytical categories.

2. Methodology

The authors employ a Grounded Theory approach to construct a bottom-up, empirically driven taxonomy rather than imposing a top-down framework. The process followed a structured seven-step protocol:

1. Resource Identification: Review of literature and reporting on AI safety within the Indian context, focusing on DPGs and Digital Public Infrastructure (DPI).
2. Domain Selection: Identification of high-impact sectors: Financial Lending, Education, Law, Agriculture, and Healthcare.
3. Use-Case Compilation: Aggregation of documented risk use-cases within these domains.
4. Classification: Differentiation between AI Safety Risks (ASRs) (hazards stemming from design flaws) and general AI risks (operational failures or intentional misuse).
   • Heuristic: A risk is an ASR if it can be mitigated by deploying a new version of the AI (e.g., better training data or architecture).
5. Category Taxonomy Formulation: Thematic analysis using open coding to group risks into granular sub-categories and meta-categories.
6. Causal Taxonomy Formulation: Mapping risks based on Timing (Development, Deployment, Usage), Stakeholder (System vs. User), and Intent (Unintentional vs. Intentional).
7. Ontology Consolidation: Synthesizing domain-specific findings into a unified, domain-agnostic AI Safety Risk Ontology.

3. Key Contributions

A. The Indian AI Safety Risk Database (AIRD)

The paper proposes the AIRD, a structured repository cataloging potential safety risks. Unlike generic databases, it is localized to the Indian context, mapping risks against sociotechnical, regulatory, and operational dependencies.

B. The ASTRA Ontology (Taxonomy)

The authors developed a comprehensive ontology classifying risks into two primary meta-categories and several sub-categories:

1. Social Risks (Technical/Systemic):

   • Bias and Exclusion: Specific to India, including Caste Bias (e.g., credit models penalizing lower castes), Religious Bias, Gender Bias, and Linguistic Bias (failure to recognize regional dialects).
   • Toxicity: Hate speech amplification, sycophantic reinforcement of false beliefs, and psychological manipulation (e.g., chatbots exploiting mental health vulnerabilities).
   • Hallucination and Error: Factual fabrication (e.g., fake loan terms), linguistic inaccuracies, and mathematical errors in educational tools.
   • Situational Awareness: Failure to account for low-bandwidth environments, hardware mismatches (e.g., old Android versions), and infrastructure exclusion.
   • Security and Privacy: Data poisoning, adversarial evasion (fraud induction), and inference-based privacy compromises (e.g., inferring pregnancy from spending patterns).

2. Frontier and Socio-Structural Risks (Governance/Systemic):

   • Human-Computer Interaction: Automation bias (over-reliance by front-line workers) and technological solutionism.
   • Socioeconomic & Environmental: Power centralization (foreign compute dominance), digital divide, and environmental hazards (energy/water usage).
   • Dangerous Capabilities: Emergent behaviors like generating instructions for harmful biological compounds.
   • Uninterpretability: "Black box" decision-making preventing accountability (e.g., opaque credit scoring).

C. Causal Framework

The paper introduces a causal taxonomy distinguishing risks by:

• Timing: Development (training data flaws), Deployment (contextual misalignment), Usage (user error or malicious use).
• Stakeholder: Is the failure intrinsic to the AI system or caused by the end-user?
• Intent: Distinguishing between unintentional design flaws and intentional misuse (e.g., state-sponsored disinformation).

4. Results and Findings

• 37 Leaf-Level Risk Classes: The study identified and categorized 37 specific risk classes across the selected domains, moving from abstract categories to concrete, measurable harms.
• Context-Specific Manifestations:
  • Credit Lending: Models trained on upper-caste, urban data systematically exclude rural and lower-caste borrowers.
  • Education: AI tutors in regional languages often fail to understand tribal dialects or provide incorrect historical facts, reinforcing myths.
  • Healthcare: Personalized AI layers in public health apps may create "filter bubbles" that confirm local myths rather than providing medical facts.
• Differentiation of ASR: The study successfully demonstrated that many "operational" failures in India (e.g., connectivity issues) are actually design failures (lack of offline-first architecture), thus classifying them as ASRs.

5. Significance and Impact

• Decolonizing AI Safety: The paper challenges the "algorithmic colonization" of Indian systems by Western benchmarks. It argues that safety taxonomies must be rooted in local socio-political realities (e.g., caste, vernacular languages) rather than imported frameworks.
• Policy and Governance: Provides a "common diagnostic language" for Indian regulators, auditors, and policymakers to evaluate AI safety. It shifts the focus from reactive mitigation to proactive safety design.
• Developmental Lens: Reframes risk to include "opportunity cost." If an AI tool fails to serve the poor due to connectivity issues, the risk is the failure of the developmental mandate, not just a technical glitch.
• Living Framework: The authors propose the AIRD as a "living repository" that will evolve through community participation (hackathons, panchayat-level studies) and quantitative calibration (Bayesian updating of incident feeds).

Scope Limitations:
The paper explicitly excludes speculative "existential" risks (e.g., superintelligence, runaway optimization) that are not yet empirically observable in India's current DPG landscape, focusing instead on immediate, tangible harms.

Conclusion:
ASTRA represents a foundational step toward a sovereign, context-aware AI safety framework for India. By grounding risk assessment in empirical data and local socio-technical realities, it aims to ensure that AI serves as a catalyst for domestic empowerment rather than a vector for new forms of exclusion and exploitation.