A new handbook from India's IT Ministry aims to solve one of AI's biggest puzzles: how to access vast datasets for training without compromising privacy or data ownership. This move signals a major shift towards structured, rights-focused AI development.
A New Rulebook for an AI-Powered India
On October 31, the Ministry of Electronics and Information Technology (MeitY) released a pivotal document that could shape the future of artificial intelligence in India. Titled the 'Handbook on Data Empowerment and Protection Architecture (DEPA) for AI Training', it was launched by MeitY Secretary S. Krishnan in Bengaluru. Developed with support from the iSPIRT Foundation, this handbook isn't just another policy paper; it's a practical techno-legal framework designed to untangle a complex knot. The central challenge it addresses is enabling startups, researchers, and large companies to train their AI models on high-quality, large-scale datasets while ensuring the organisations providing that data—like hospitals, banks, and government agencies—don't lose control over it. In essence, it aims to build the rails for a data-sharing superhighway with robust guardrails for privacy, confidentiality, and intellectual property.
The Privacy Imperative: 'Model-to-Data'
The handbook introduces a core concept that fundamentally changes the data-sharing dynamic, summarized by Principal Scientific Adviser Ajay Kumar Sood as letting 'the model travel to the data, rather than the data travelling to the model'. This approach is a direct response to widespread privacy concerns. Instead of data providers sending huge, sensitive datasets to AI developers, the developers' AI models are brought into a secure, controlled environment where the data resides. The model trains inside this 'secure data environment' without the developers ever directly accessing or exposing the raw data. This method is supplemented with other advanced techniques, such as 'differential privacy', which adds statistical noise to data to make it impossible to re-identify individuals from the results. The framework acknowledges a modern reality: simply removing names and addresses is not enough to guarantee anonymity when datasets can be combined. This new approach prioritises privacy by design, making it a foundational element rather than an afterthought.
Empowering Institutions with Data Rights
For years, institutions have accumulated vast and valuable information—from patient records in hospitals to financial data in banks and research data in universities. The rise of AI has turned these datasets into digital gold, but also created a power imbalance where institutions feared losing control over their most valuable asset. The DEPA handbook directly tackles this by formally recognising the roles of 'Training Data Providers' (the institutions) and 'Training Data Consumers' (the AI developers). It provides a contractual and technical basis for providers to grant access for specific purposes without surrendering ownership. Furthermore, the framework proposes the creation of 'Data Collabs', which will act as sector-specific intermediaries, pooling data from multiple providers to create richer, more diverse datasets for training AI in fields like healthcare or finance. This structure empowers institutions to participate in the AI economy on their own terms, ensuring they remain sovereign over their data.
Why Now and What's Next?
The release of this handbook is a timely intervention. Globally, regulators are tightening the screws on how AI models are trained. High-profile cases of data scraping and the insatiable appetite of generative AI for information have pushed the issues of consent and data rights to the forefront. India’s Digital Personal Data Protection (DPDP) Act of 2023 already laid the groundwork, and the DEPA framework builds on it, providing specific guidelines for the complex use case of AI training. In his foreword, IT Minister Ashwini Vaishnaw highlighted that DEPA provides a 'consent-based framework' for user-controlled data sharing. The handbook has been opened for public recommendations, indicating that this is the beginning of an iterative process. The success of this initiative will depend on its adoption by the industry and the effective functioning of the proposed 'Data Collabs'. However, its launch marks a clear declaration from the government: India intends to be a major player in the global AI race, but it plans to do so responsibly, by building a foundation of trust, transparency, and accountability.
