India’s IT ministry has released a draft handbook called DEPA for AI Training, a plan that lets AI developers train models on hospital and bank data without ever seeing the raw records. The idea is simple: the model travels to the data, not the other way around. It is a draft, so nothing is legally binding yet.
Here is the short version before we get into how it works.
- Who: India’s Ministry of Electronics and IT (MeitY), prepared with the iSPIRT Foundation.
- When: Unveiled at a Bengaluru roundtable at the end of October 2026 (PIB posted it on 31 October).
- What: A techno-legal framework for training AI on sensitive data held by banks, hospitals and agencies.
- Status: Draft handbook. Consultations are expected in the coming weeks.
If you build, buy or are simply affected by AI in India, this is the early shape of the rules your data may live under.
What DEPA for AI Training actually is
DEPA stands for Data Empowerment and Protection Architecture. The original version, from NITI Aayog in 2020, is the consent system behind India’s Account Aggregator network, which lets you share your financial data with apps you choose.
The new handbook stretches that idea to AI training. According to India’s Press Information Bureau, it combines contractual permissions, controlled computing environments, privacy safeguards and data provenance tracking. Provenance means a record of where data came from, who allowed its use and how it was used.
The problem it targets is what the handbook calls a “data deadlock.” Hospitals and banks hold the best data for AI, but they rarely share it, because a leak can be very costly. Elets CIO notes that penalties under India’s data protection law can reach ₹250 crore.
That fear keeps useful data locked away. A hospital may hold thousands of MRI scans that could train a better cancer-detection tool, yet the safest choice for the hospital is to share nothing. DEPA tries to change that trade-off by making sharing safe enough to be worth the risk.
How the “sealed room” works
Think of a locked lab. A developer sends in a model, the lab trains it on the data inside, and only the result comes out. Elets CIO describes four building blocks behind this setup.
- A data exchange platform where datasets are registered.
- Machine-readable electronic contracts that spell out what is allowed.
- A secure environment that runs only what the contract permits.
- Differential privacy, a technique that adds statistical noise so individuals can’t be picked out.
The point is that the contract is enforced by software, not by trust. Developers get back a trained model or just a decision, such as a loan outcome, and never touch the raw files.
Who plays which role
The framework splits the work between a few parties. The table below shows who does what, based on the Elets CIO report.
| Role | Who it is | What it does |
|---|---|---|
| Data providers | Hospitals, banks, public agencies | Register datasets and keep custody of them |
| Data consumers | Startups, researchers, enterprises | Request access for one specific use case |
| Data Collabs | New intermediaries | Combine and standardise data from several sources |
| Samyog | Industry group powered by iSPIRT | Registers participants, supplies standard contracts, handles disputes |
This structure matters because it gives smaller teams a standard way to ask for data, instead of negotiating a one-off deal with every hospital.
Early pilots, and what’s still unclear
Elets CIO reports two early tests. A bank has run the approach since March, reportedly handling over two million requests in two weeks. In healthcare, hospitals and a health-tech startup are training a tumour-detection model on MRI scans that the startup never sees.
The handbook also sizes the opportunity. It puts the financial services landscape at $550 billion and the healthcare market at more than $600 billion, and it names education, with 250 million students, as another target. Those are the handbook’s own figures as reported, so treat them as pitch numbers, not forecasts.
Several big questions remain open. Liability if data leaks from a sealed room is not specified, and payment for data owners “will be explored.” Observers quoted by Elets expect binding rules to come from sector regulators rather than a new law.
Timing also matters. The main rules of India’s DPDP Act, 2023 are due to take effect in May 2027, so this draft is landing just ahead of them.
What it means for you
If you’re an everyday user, nothing changes today. Longer term, the goal is AI tools trained on real medical and financial data without your records being copied around.
It also sits in a wider debate about who controls training data. Our guide on blocking AI training with robots.txt covers the website side of that, and our note on the SIM binding rule for WhatsApp in India shows how fast Indian digital rules are moving.
Frequently Asked Questions
Is DEPA for AI Training a law?
No. It is a draft handbook and a framework, not legislation. Binding requirements would likely come later from sector regulators.
Does the AI developer see my personal data?
Under the design, no. Raw personal and sensitive data stays inside the secure environment, and developers receive only the trained model or a decision.
Who built the framework?
MeitY prepared it with the iSPIRT Foundation. MeitY Secretary S. Krishnan launched it at a Bengaluru roundtable.
When will it take effect?
There’s no date yet. Consultations are expected in the coming weeks, with finance and healthcare pilots first.
My take: this is the most practical approach to AI training data I’ve seen from a government this year, because it lets data stay where it is. It will only work if the open questions on liability and payment get clear answers. Watch the consultation period, since that is where the details will be settled.
Leave a Reply