Astra: AI Safety, Trust, & Risk Assessment

One-size-fits-all AI safety frameworks miss critical local contexts: The paper demonstrates that Western-developed AI safety guidelines overlook India-specific challenges like caste-based discrimination in algorithms, exclusion of vernacular language speakers, and infrastructure gaps in rural areas—issues affecting 1.5 billion people that require tailored solutions rather than generic global standards.

Technical fixes alone aren't enough; we need to categorize who is responsible: ASTRA's framework distinguishes between risks caused by design flaws (fixable through better training data or guardrails) versus risks stemming from user misuse or systemic inequities, helping policymakers target interventions at the right actors rather than assuming all problems have technical solutions.

AI risks should be mapped by when they occur, not just what they are: By organizing risks across development, deployment, and usage stages, the framework enables earlier intervention—catching discrimination in training data before systems launch, rather than only addressing harms after they've already affected people.

Starting with high-impact sectors creates a foundation for broader protection: Focusing initial risk assessment on Education and Financial Lending—sectors directly affecting economic opportunity and social mobility—provides a practical, scalable model that can expand to other domains while remaining grounded in real-world consequences.

AI safety frameworks should be "living" documents that evolve with technology: Rather than static regulations, ASTRA proposes a dynamic regulatory approach that can adapt as India's AI ecosystem grows, ensuring safety measures remain relevant instead of becoming outdated.