DEPA for AI Training: Government Handbook on Sharing Hospital, Bank and Government Data to Train AI Models

The Ministry of Electronics and Information Technology (MeitY) released a handbook called Data Empowerment and Protection Architecture (DEPA) for AI Training at a roundtable in Bengaluru at the end of October 2026. It is open for public suggestions.

It sets out how institutions that hold data, such as hospitals, banks and government agencies (called Training Data Providers), can share datasets with startups, researchers and companies (Training Data Consumers) that want to train AI models. Intermediaries called Data Collabs can pool data from many sources for a sector.

In a section titled "Beyond Individual Consent", the handbook says taking each person's consent is "impractical at this scale" for training. People's interests are to be protected through governance rules and technical safeguards instead.

It separates training (building a model from large, combined data) from inference (using a model on one person's data, like checking a loan application). For inference, consent-based systems like the Account Aggregator still apply.

Proposed safeguards: data must be de-identified before sharing, models are trained inside secure data environments where developers cannot see raw data, and differential privacy (adding statistical "noise") is used so individuals cannot be traced from a model.

Ideas by sector: link TB data from the Nikshay portal with e-Sanjeevani telemedicine data for disease early-warning; train credit models on de-identified GST and utility payment data; and build "skills passports" linked to the Academic Bank of Credits. A government authority would certify models as "fit for use". Pilots, regulatory sandboxes and royalty or access-fee models for data providers are planned.