Institutions across India are sitting on a goldmine of data essential for training the next generation of AI. A new government-backed handbook aims to provide a crucial roadmap for unlocking this potential while protecting privacy and data rights.
The High-Stakes Balancing Act
Artificial
intelligence is advancing at a breathtaking pace, fuelled by a voracious appetite for data. For AI models to become more powerful and accurate, they need to learn from vast and diverse datasets. Institutions like hospitals, banks, universities, and government agencies hold some of the most valuable data pools in existence—rich with real-world information crucial for developing sophisticated AI in sectors like healthcare, finance, and public services. However, this data is also incredibly sensitive. It contains personal records, confidential research, and proprietary information. Sharing it without ironclad safeguards creates immense risks, including devastating data breaches, loss of institutional control over valuable assets, and the erosion of public trust. Once information is entered into an external AI tool, the institution loses authority over how it is stored, processed, or shared, opening the door to legal and regulatory risks.
India's New Rulebook for AI Data
Recognising this challenge, India's Ministry of Electronics and Information Technology (MeitY) has released a new framework titled 'Data Empowerment and Protection Architecture (DEPA) for AI Training'. Unveiled in late October, this handbook serves as a guide for institutions, which it calls 'Training Data Providers', on how to safely share datasets with startups, researchers, and companies, referred to as 'Training Data Consumers'. The framework aims to enable access to high-quality data for AI innovation while strictly safeguarding privacy, commercial confidentiality, and institutional control. It even envisions a role for intermediaries called 'Data Collabs' to pool information from multiple sources within specific sectors to create richer, more useful datasets.
Privacy and Rights at the Core
The handbook acknowledges a critical fact of the digital age: simply removing names and other direct identifiers from a dataset is not enough to guarantee anonymity. Datasets can be combined to re-identify individuals, a significant privacy risk. To counter this, the framework proposes a two-pronged approach. The first is the creation of 'secure data environments', which allow AI models to be trained on data without developers ever being exposed to the raw information. The second is the use of 'differential privacy', a sophisticated technique that adds statistical 'noise' or randomness to the data. This makes it impossible to infer information about any single individual from the AI model's outputs. These privacy-enhancing technologies are central to building trust and ensuring that data can be used for the greater good without compromising individual rights.
How the Model Travels to the Data
A key principle articulated by Principal Scientific Adviser Ajay K. Sood is that under this new architecture, 'the model travels to the data'. This represents a paradigm shift from the traditional method of moving large, sensitive datasets to a developer's servers. Instead, AI developers send their models into the secure, controlled environment where the data resides. The model is trained inside this digital vault, and only the refined, privacy-safe model is allowed to leave. The handbook outlines two potential models for this exchange. In one, the data provider anonymises the data, and the consumer can either download it or access it in the secure environment. A more secure second option involves a platform that anonymises and encrypts the data, forcing the consumer to work on it exclusively within that protected space.
The Road Ahead
While this framework provides a much needed guide, implementation will be the true test. Institutions will need to develop robust internal data governance policies and technical capabilities to manage these new workflows. Building a cross functional AI committee with representatives from IT, legal, and academic leadership is a recommended first step for many organisations. Furthermore, clear agreements will be essential, specifying data ownership, usage rights, and liability if something goes wrong. For India to fully harness its AI potential, the principles laid out in the MeitY handbook must be translated into concrete, auditable practices that build confidence among institutions, innovators, and the public alike.
