Quick answer: The Ministry of Electronics and Information Technology (MeitY) has released a draft framework for using institutional datasets to train AI models while controlling access to the underlying data. Its DEPA for AI Training handbook proposes electronic agreements, data exchange platforms, secure computing environments and privacy safeguards. It is a draft for discussion, not a new statute or a blanket permission to use personal data.
MeitY listed the draft DEPA for AI Training handbook on 29 October 2026 and announced its launch on 31 October. The proposal matters to researchers, AI developers, hospitals, banks and other institutions that hold useful data but face privacy, confidentiality and intellectual property concerns when sharing it.
What is DEPA for AI Training?
DEPA stands for Data Empowerment and Protection Architecture. The new handbook addresses the training stage of AI: the process in which a model learns patterns from a collection of examples. It distinguishes that task from inference, where an already trained model processes a particular input to produce a result.
The earlier DEPA inference model, illustrated by India's Account Aggregator system, focuses on sharing an individual's information for a specified service with consent. Training a model on large, varied datasets raises different questions: who can use the data, for what purpose, within which computing environment, and under whose oversight? The handbook proposes a structure for those institutional arrangements.
How would the proposed framework work?
| Participant or tool | Proposed role |
|---|---|
| Training Data Provider (TDP) | An institution, such as a hospital, bank or public agency, that holds a dataset and defines conditions for its use. |
| Training Data Consumer (TDC) | A researcher, start-up or enterprise seeking to train or test a model using that data. |
| Data Exchange Platform | A place to discover and request datasets and, in some designs, to provide computing and data protection services. |
| Electronic contract | A machine-readable agreement intended to check and enforce the permitted purpose and conditions of access. |
| Secure Data Environment (SDE) | A controlled setting for training and testing where access to sensitive data can be restricted. |
The draft describes a sequence: prepare and register a dataset; let a consumer make a purpose-specific request; check that request against the provider's terms; train in an approved environment; then validate and monitor the resulting model. It also proposes techniques such as encryption, provenance records and differential privacy to reduce disclosure and misuse risks.
Important distinction: The handbook describes two possible data-handling arrangements. In one, the provider anonymises data and the consumer may download that dataset or use an SDE. In the other, the platform handles anonymisation and encryption of raw data and the consumer works within the SDE. It would be inaccurate to say that every version of the proposal forbids all dataset downloads.
What is the legal position today?
The document calls itself “Draft for Discussion” and presents a proposed architecture. It does not, by itself, amend the Digital Personal Data Protection Act, 2023, the Information Technology Act, 2000, or sector-specific rules. Its suggested controls should therefore be read as the framework's design, not as obligations that have already been imposed on every AI developer in India.
That distinction also applies to model review. The draft discusses validation for bias, security and fitness for use within the proposed network, and suggests a role for a competent government authority. It does not establish a general statutory pre-approval requirement for all AI models. Any actual processing or sharing must be assessed under applicable law and the relevant data rights and contractual terms; an electronic contract in this architecture cannot, on its own, override those requirements.
The draft envisages an oversight body to set privacy, ethics and security standards, register eligible participants, supervise data-sharing procedures and help resolve disputes. Its implementation plan discusses initial pilots in areas such as finance and healthcare. Those are proposals whose details may change as the framework develops.
Why does this matter for tech law?
The legal challenge is not simply whether an AI model can be trained. Institutions must identify the source and permitted use of the data, protect confidential material and intellectual property, and ensure that a training arrangement respects privacy law. The proposed electronic contract and audit trail aim to make those limits visible and enforceable in a technical workflow. Whether the design succeeds will depend on its implementation, the safeguards actually adopted and the applicable legal regime.
For students, the point to remember is the difference between a government draft architecture and an enacted legal duty. The DEPA handbook is a useful example of “techno-legal” governance, in which contractual rules and technical controls are designed together. It should not be cited as though it were a notification creating an AI licensing regime.
Key takeaways
- MeitY's DEPA for AI Training document is a draft framework for controlled data collaboration in model training.
- It describes TDPs, TDCs, electronic contracts, exchange platforms and secure data environments.
- The draft offers more than one arrangement for anonymisation and access; its safeguards must be described with that distinction in mind.
- It proposes oversight and model validation within the framework, without creating a general AI approval law.
- Existing legal duties must be evaluated independently of the handbook.
Frequently asked questions
Is DEPA for AI Training a new law?
No. The MeitY document is marked “Draft for Discussion.” It describes a proposed architecture, not an enacted statute or rule.
Does it allow unrestricted use of personal data to train AI?
No. The framework proposes controls on access and use. It does not replace the need to comply with applicable data protection law.
Must all AI models be cleared by the government before deployment?
The draft discusses validation of models trained within its proposed system. It does not create a general statutory clearance requirement for every AI model.
What is the difference between a TDP and a TDC?
A TDP holds and provides data under defined conditions. A TDC requests access for training, testing or evaluating an AI model.
Can an AI developer download a dataset under the proposal?
One option in the draft permits access to provider-anonymised data for download, with an SDE also available. A second option places anonymisation and access inside the platform's secure environment. The permitted use would depend on the arrangement and applicable terms.
This update reflects the draft and official announcement available on 9 November 2026. Readers should consult the linked MeitY handbook for the full proposal.
