India’s Ministry of Electronics and Information Technology (MeitY) launched a handbook on October 31 for “DEPA for AI Training,” a plan to let developers train AI models on sensitive datasets without ever receiving a copy of the data. In the words of the country’s Principal Scientific Adviser, Ajay Kumar Sood, “the model travels to the data, rather than the data travelling to the model.” The framework is a proposal for discussion rather than a law, and the government has not said when any platform will be running.
What DEPA for AI Training is
DEPA stands for Data Empowerment and Protection Architecture. It builds on India’s earlier consent-based data-sharing system for personal financial data, the Account Aggregator, which lets individuals approve sharing a specific record for a specific purpose. MeitY’s document says that approach does not scale to AI training, which needs large collective datasets rather than one person’s data for one transaction. DEPA Training is meant to fill that gap, and the document describes it as a “techno-legal” design: part software, part contract.
The government’s PIB press release says the framework combines contractual permissions, controlled computing environments and privacy-enhancing safeguards, and adds data provenance so the origin, permissions and use of data can be traced during training.
How the architecture works
The MeitY document names a handful of roles and building blocks:
- Training Data Providers (TDPs): organizations that hold data, such as hospitals, banks and government agencies.
- Training Data Consumers (TDCs): developers, researchers, start-ups and regulators who want to train or test models.
- Data Exchange Platforms: intermediaries that connect the two sides and enforce the rules.
- Secure Data Environments: isolated computing spaces where training happens, so providers’ raw data stays protected and is not exposed to consumers. The document says all model training must take place inside approved environments and that resulting models are subject to validation.
- Electronic contracts: machine-readable agreements that spell out permitted uses and are enforced technically.
- Differential privacy: a mathematical technique, with audited privacy budgets, intended to stop a trained model from revealing information about individuals.
The framework describes two options for anonymization. In one, providers anonymize data themselves; in the other, the exchange platform handles anonymization and encryption inside its secure environment. The document says an appropriate regulatory organization would oversee the arrangement.
Why India is proposing it
The Register’s account says India wants models trained on local data because it expects offshore model makers to be less motivated to address local needs or to let data owners stay in control. The MeitY foreword by the Secretary of the Department of Science and Technology talks about a “data chasm”: Indian researchers have talent but often lack large, high-quality, India-specific datasets.
The government points to health, finance and education as early targets. Examples in the framework include unlocking fragmented health data for diagnostics and drug discovery, and using shared financial data to extend credit to an estimated 400 million underserved individuals and small firms. The Register notes the optimism is considerable and that it is unclear how much work remains before any of this runs.
How it compares with the usual approaches
Most AI training today works by copying data into the developer’s own infrastructure, which is where disputes over consent and copyright arise. We explain those fights in our guide to AI copyright and training-data lawsuits. DEPA flips the flow: data providers keep custody, set conditions and can audit what happened. It is closer in spirit to a data clean room than to a public dataset release.
The trade-off is practical. Training inside a controlled environment adds cost, limits which tools can be used, and requires providers to trust the platform operator. Differential privacy also reduces accuracy as privacy budgets tighten. Whether providers will take part depends on incentives and liability rules that the framework leaves to later detail. For the broader idea of keeping data and compute under national control, see our explainer on sovereign AI, and for how smaller economies are approaching AI, our report on the World Bank’s small-AI guidance.
What to watch next
- Public comments on the handbook and whether MeitY turns it into binding rules or standards.
- Which regulator would oversee data exchange platforms and certify secure environments.
- The first pilot sectors, most likely health, finance or education, and which providers join.
- Whether foreign developers can become data consumers, and on what terms.
- Technical details such as hardware requirements and the privacy budgets allowed.
Key takeaways
- MeitY launched a handbook on DEPA for AI Training on October 31; it is a framework for discussion, not a law.
- Models would be trained inside secure environments, so data providers never hand over raw data.
- The design uses data exchange platforms, electronic contracts, secure environments and differential privacy.
- India says the aim is models built on local data, with early focus on health, finance and education.
- It is unclear when platforms will run or how many providers will participate.
Frequently asked questions
What does DEPA stand for?
DEPA stands for Data Empowerment and Protection Architecture, an Indian government framework for sharing data under consent and technical safeguards. The new handbook extends it to AI model training.
Does DEPA let AI developers download Indian datasets?
Not in the main design. Developers would train models inside approved secure environments, where the raw data stays protected and is not exposed to them. One option in the document lets consumers download anonymized data.
Is DEPA for AI Training already in force?
No. MeitY launched a handbook and framework document for discussion, and it is unclear how much work remains before platforms exist.
