India Launches DEPA Handbook for Secure AI Training Data Sharing

India has taken a new step in its effort to give artificial intelligence developers access to better data without asking the institutions that hold that data to give up control of it. On Saturday, October 31, 2026, the Union Ministry of Electronics and Information Technology (MeitY) launched the Handbook on Data Empowerment and Protection Architecture (DEPA) for AI Training at a high-level roundtable in Bengaluru.

The handbook was unveiled by S. Krishnan, Secretary, MeitY, with support from the iSPIRT Foundation. It sets out a techno-legal framework for purpose-bound data collaboration, allowing institutions to share data for AI training while retaining control over their information. In practical terms, the idea is to make trusted data collaboration possible at scale, so that AI models can be trained on large, high-quality datasets without the data holders losing sight of how their data is used.

What DEPA is supposed to solve

Good AI models need good data, but the institutions that hold useful data have often had good reasons not to share it. The framework seeks to address challenges that often restrict data sharing, including concerns around privacy, confidentiality, intellectual property and institutional control. Those concerns are not abstract. A hospital, a bank, a university or a government department may hold data that could improve an AI system, yet still be unwilling to hand copies of that data to an outside developer, because once a copy leaves its control, it is hard to track, limit or withdraw.

DEPA is presented as an answer to that problem. It combines contractual permissions, controlled computing environments and privacy-enhancing safeguards. It also incorporates data provenance, allowing the origin, permissions and use of data to be tracked throughout the AI training process to ensure accountability and transparency. In other words, the framework is not only about opening access, but about keeping a record of where data came from, what it was permitted to be used for, and how it was used.

That emphasis on provenance matters because questions about training data are now central to the AI debate worldwide. As major AI companies promise publishers a single way to block some scraping, the argument over who controls training data is moving from courtrooms and company policies into technical standards. DEPA approaches the same question from the other side: instead of only blocking unwanted use, it tries to create a trusted route for wanted use.

The model travels to the data

The most striking idea in the launch is also the simplest to explain. Principal Scientific Adviser to the Government of India Ajay Kumar Sood highlighted that the framework allows AI models to access data without requiring institutions to transfer the underlying information.

Under DEPA, the model travels to the data, rather than the data travelling to the model.

Sood said that approach could facilitate trusted, multi-party and cross-border data collaboration and accelerate scientific discovery. He described the framework as India’s contribution to global science diplomacy and efforts to develop an inclusive and responsible global AI ecosystem. The distinction is important: if several institutions can each keep their data in place while a model learns from it under agreed rules, collaboration becomes possible without any one party having to surrender its dataset.

MeitY Secretary S. Krishnan framed the launch in similar terms. He said India’s AI ambitions must be matched by the ability to use data responsibly and securely, and that the DEPA Training Framework enables trusted data collaboration at scale, unlocking access to large-scale datasets for AI training while safeguarding privacy, commercial confidentiality and institutional control.

Why institutions have held back

The roundtable brought together representatives from government, industry, technology, academia, research and media to discuss trusted data collaboration and its role in advancing AI innovation and public good. That breadth of attendance reflects how wide the data-sharing problem is. Researchers need diverse datasets to develop more robust AI models, while data custodians and contributors need their rights protected. Kris Gopalakrishnan, co-founder of Infosys and chairman of Axilor Ventures, said the framework could help researchers access diverse datasets to develop more robust AI models while protecting the rights of data custodians and contributors.

Sharad Sharma, co-founder of iSPIRT Foundation, described the launch as a significant milestone for India’s data economy, comparing its importance with the announcement of the Unified Payment Interface (UPI) in February 2016. That comparison is about the scale of ambition, not a prediction of outcome: UPI changed how payments work in India by creating a shared, trusted layer that many different institutions could use. DEPA is presented as an attempt to create a comparable shared layer for data, where permission, control and accountability are built into the collaboration itself.

• Launched on October 31, 2026, at a high-level roundtable in Bengaluru by MeitY, with support from the iSPIRT Foundation.
• DEPA stands for Data Empowerment and Protection Architecture, and the handbook covers AI training.
• The framework combines contractual permissions, controlled computing environments and privacy-enhancing safeguards.
• Data provenance tracks the origin, permissions and use of data throughout the AI training process.
• The model travels to the data, so institutions do not have to transfer their underlying information.
• The launch brought together government, industry, technology, academia, research and media representatives.

What this means for AI development

For AI developers, the practical promise of DEPA is access to datasets that are currently locked away, without the need to negotiate a separate, uncertain arrangement for every dataset. For data holders, the promise is the opposite side of the same coin: participation without loss of control. If that balance holds in practice, it could widen the range of data available for training, including data from sectors where privacy and confidentiality have made sharing difficult.

The timing also matters. AI features are reaching more users in more countries, as Apple brings its AI features to iPhones in mainland China, and investment in AI continues at a scale that has unsettled markets. In that environment, the question is no longer only which models are most capable, but which models are trained on data obtained through a process that data holders, regulators and the public can trust. A handbook does not settle that question on its own, but it gives institutions a common language for permission, provenance and control.

What to Watch Next

What to watch next is whether trusted data collaboration at scale follows from the handbook, whether institutions use the framework to share data for AI training while retaining control, and whether the approach of the model travelling to the data facilitates the multi-party and cross-border collaboration described at the launch. The other test is whether privacy, commercial confidentiality and institutional control are safeguarded in practice as access to large-scale datasets widens, and whether researchers gain access to the diverse datasets needed to develop more robust AI models while the rights of data custodians and contributors are protected.