As artificial intelligence continues to rapidly evolve, the conversation around its responsible development and deployment has become increasingly urgent. In January 2025, India took a significant step toward leading this global dialogue by announcing the establishment of the IndiaAI Safety Institute (AISI). Spearheaded by the Ministry of Electronics and Information Technology (MeitY) under the broader IndiaAI Mission, this initiative marks a landmark effort to ensure that AI technologies developed and deployed within the country are safe, ethical, and aligned with national priorities.

The institute will serve as a cornerstone of the Safe and Trusted AI pillar within the IndiaAI Mission and seeks to build an ecosystem where innovation coexists with accountability, transparency, and public trust.

The Vision and Framework: Building a Contextual AI Safety Ecosystem

The IndiaAI Safety Institute has been conceptualized to act as a central authority on AI safety, risk mitigation, and responsible AI deployment. It is structured around a hub-and-spoke model, where a central institution collaborates with a network of academic institutions, research organizations, startups, and industry players.

This model enables both centralized coordination and decentralized execution. For example, while the central institute may define the safety standards and research priorities, specialized research labs at IITs or private AI firms may execute projects in their areas of expertise. This allows for a highly adaptive system that evolves with the technology landscape.

A critical strength of this initiative is its focus on indigenous R&D using Indian datasets. AI systems built and trained exclusively on Western datasets often fail to perform accurately in Indian contexts due to vast linguistic, cultural, and socio-economic differences. For instance, a facial recognition algorithm trained on Western datasets may exhibit significantly lower accuracy when identifying faces from diverse Indian ethnic groups. By emphasizing research tailored to Indian realities, the institute aims to close these gaps and build inclusive AI systems.

First Round of Responsible AI Projects: From Theory to Application

To kickstart the Safe and Trusted pillar, IndiaAI launched its first Expression of Interest (EoI) in late 2024, which saw an overwhelming response from the academic and industrial research community. From over 2,000 applications, eight projects were selected based on their relevance, feasibility, and alignment with national AI safety priorities. These projects reflect a diverse set of problem statements, ranging from technical challenges to ethical frameworks.

1. Machine Unlearning – IIT Jodhpur

As data privacy norms tighten globally, the concept of machine unlearning has gained importance. It refers to the ability of a model to forget specific data points, ensuring that users can exercise their “right to be forgotten.” For example, if an individual withdraws consent for their medical data to be used in training an AI diagnostic tool, machine unlearning would ensure the model no longer retains or uses that data.

2. Synthetic Data Generation – IIT Roorkee

Given the scarcity of high-quality, annotated datasets in regional languages and domains like agriculture or informal labor, generating synthetic datasets has become essential. IIT Roorkee is working on methods to produce synthetic yet realistic datasets that can help train AI systems in low-data environments without compromising on privacy.

3. Bias Mitigation in AI Models – NIT Raipur

AI models often inadvertently perpetuate societal biases. A recruitment tool trained on historical hiring data might systematically disadvantage women or marginalized communities. NIT Raipur’s project explores algorithmic techniques to detect and mitigate such biases before they become embedded in AI systems.

4. Privacy-Enhancing Technologies – DIAT Pune and Mindgraph

This collaboration focuses on cryptographic methods like homomorphic encryption and federated learning, which allow AI models to be trained without direct access to sensitive data. This is particularly relevant in sectors such as healthcare, where patient confidentiality is paramount.

5. Explainable AI for Public Services – IIT Delhi, IIT Dharwad, IIIT Delhi

As government services increasingly rely on AI (e.g., predictive policing or welfare benefit distribution), ensuring that decisions made by these systems are understandable and auditable is essential. These institutions are developing explainability frameworks to make AI decisions transparent to both administrators and citizens.

6. Ethical AI Frameworks for Governance – IIIT Delhi

This project aims to translate abstract ethical principles into actionable policies for public sector AI use. For example, guidelines on consent management, data provenance, and algorithmic fairness for government departments using AI in education or transportation.

7. Algorithm Auditing Tools – Civic Data Labs

Independent auditing is necessary to verify whether deployed AI systems comply with ethical norms and legal frameworks. Civic Data Labs is developing open-source tools that allow such audits to be conducted efficiently, making AI systems more accountable to the public.

8. AI Risk Testing Frameworks – Amrita Vishwa Vidyapeetham

Similar to stress testing in the financial sector, this project aims to develop a comprehensive framework for testing AI systems under high-risk scenarios, such as adversarial attacks or deployment in critical infrastructure.

Second Expression of Interest: Scaling the Mission

Following the momentum of the first EoI, IndiaAI launched the second round of EoI in early 2025, broadening the scope and inviting participation from a wider pool of Indian academic institutions, startups, and R&D organizations.

The themes in this round are deeply strategic and forward-looking:

• Deepfake Detection: In light of the increasing misuse of AI-generated videos in misinformation campaigns, this initiative seeks technical tools for early detection and provenance verification.
• Stress Testing Tools: Ensuring the robustness of AI systems, especially those used in mission-critical applications such as disaster response or national security.
• AI Risk Assessment & Management: Developing comprehensive risk taxonomies and mitigation strategies applicable across industries.
• Watermarking and Labelling: Techniques to label AI-generated content to prevent confusion or misuse, particularly on social media.
• Responsible Use Protocols: Guidelines and compliance frameworks that organizations must adhere to when deploying AI technologies.

The deadline for submission was extended to February 28, 2025, signaling the government’s intent to make the process more inclusive and participatory.

A Globally Aligned, Locally Rooted Approach

What sets the IndiaAI Safety Institute (AISI) apart from similar global initiatives is not only its ambition but its commitment to designing AI systems rooted in India’s cultural, linguistic, and socio-economic diversity.

While countries across the world—from the EU’s AI Act to the U.S. Executive Order on AI—have made strides in establishing frameworks for responsible AI, India’s approach is uniquely positioned to reflect the realities of a complex, pluralistic society. The AISI aims to bridge the gap between global AI governance ideals and India’s grassroots-level deployment challenges, ensuring that safety, fairness, and trust are embedded in technologies meant for all layers of society.

1. Linguistic Diversity as an Inclusion Imperative

India is home to 22 constitutionally recognized languages, with thousands of dialects spoken across rural and urban landscapes. In contrast to markets where English or a single national language dominates, AI systems in India must function across multilingual environments—be it chatbots for agriculture advisory in Bundeli or healthcare applications offering instructions in Tamil.

This presents a multi-layered challenge for Natural Language Processing (NLP):

• Language equity: Prioritizing major Indian languages without marginalizing smaller dialects.
• Code-mixed language handling: Building models that understand text mixing Hindi and English (commonly used in messaging apps).
• Data collection and annotation: Creating high-quality, labeled datasets in underrepresented languages is often resource-intensive and commercially unappealing for global firms.

The IndiaAI Safety Institute can play a critical role in developing foundational language models fine-tuned on Indic datasets, while also guiding ethical practices in regional language data sourcing.

2. Digital Literacy and Interface Design

The World Bank reports that nearly 40% of India’s population remains offline, and a significant portion of digital users have only basic literacy. This makes explainability and accessibility not just ideal traits—but absolute requirements.

For example:

• Voice-based navigation: AI platforms serving farmers, frontline health workers, or elderly users must go beyond text-heavy interfaces.
• Visual explanations: AI predictions—say, in weather forecasting or credit scoring—need to be backed with intuitive, image-based cues that are easy to grasp.
• Low-bandwidth optimization: AI apps must be lightweight, capable of running on entry-level smartphones without constant connectivity.

In this context, the “explainable AI” principle becomes deeply local. It isn’t just about model interpretability for a data scientist—it’s about ensuring that a rural homemaker using an AI-enabled loan app understands why she was approved or denied credit.

The AISI is designed to support such socially rooted innovations, setting India-specific benchmarks for explainability, fairness, and usability.

3. Socio-economic Context as a Driver of Responsible AI

India's diversity is not only linguistic—it encompasses caste, gender, income, geography, and education. Algorithms that work seamlessly in urban Bengaluru may fail in tribal Odisha or among underbanked communities in Bihar.

Globally, there’s growing concern about algorithmic bias. But in India, these concerns are heightened by structural inequalities that can be unintentionally reinforced if AI models are trained on non-representative or biased data.

For instance:

• A facial recognition system trained primarily on lighter-skinned individuals might underperform for darker-skinned populations, reinforcing exclusion in welfare schemes.
• An ed-tech recommendation engine that only suggests English-language courses could disadvantage students from vernacular backgrounds, despite equal potential.

The AISI's mandate to work with contextualized Indian datasets ensures that responsible AI in India isn’t imported—it is co-designed with the communities it serves.

Conclusion: India as a Global Leader in Responsible AI

The establishment of the IndiaAI Safety Institute is a timely and strategic intervention. As countries worldwide debate regulations and ethical frameworks for AI, India has chosen to act by funding indigenous research, building institutional capacity, and nurturing a national ecosystem for AI safety and trust.

This is not just a policy announcement; it is the beginning of a paradigm where responsible innovation becomes the foundation of India’s AI ambition. If successfully implemented, India may not only emerge as a leader in AI technology but also as a global benchmark in AI governance and ethics.

References:

• https://indiaai.gov.in/article/india-takes-the-lead-establishing-the-indiaai-safety-institute-for-responsible-ai-innovation