समाचार Samāchār · Tech & AI · 31 October 2026

MeitY drafts a consent-led data pipe for AI training, built on the UPI playbook

The DEPA for AI Training framework, prepared with iSPIRT, would let startups and researchers train models on bank, hospital and public data without ever seeing the raw records.

India has proposed a way to unlock its data for AI without giving up privacy. The IT ministry released a draft framework called DEPA for AI Training on 30 October, prepared with the iSPIRT Foundation. MeitY Secretary S Krishnan unveiled a companion handbook at a roundtable in Bengaluru. Principal Scientific Adviser Ajay Kumar Sood and Infosys co-founder Kris Gopalakrishnan were among the speakers, with HDFC Bank, Sarvam and Emergent AI also represented.

The design extends the Data Empowerment and Protection Architecture that already runs consented sharing through Account Aggregators. That model moves one person’s data, with consent, for one use. This one is for bulk, de-identified data used to build models. Banks, hospitals and public agencies register datasets as Training Data Providers. Startups, researchers and firms request them as Training Data Consumers for a stated purpose. Data Collabs aggregate and standardise data across sources.

The safeguards are technical, not only legal. Training runs inside a secure environment, so consumers never see raw records. Differential privacy protects individuals in the aggregate. Machine-readable electronic contracts check every request against the permitted use. Trained models must pass checks for bias, security and explainability before deployment. Business Standard reports that the document warns simple anonymisation will not work because of re-identification risk, and puts the prize at $500 billion of AI-driven value for India’s GDP by 2030. Pilots are planned in finance and healthcare, with regulatory sandboxes alongside.

This is the right instinct. India built UPI and Aadhaar as public rails and let thousands of firms build on top. AI at population scale needs the same treatment for data. Indic models trained on Indian credit, health and education records will serve Indians better than models trained elsewhere. Credit goes to the MeitY and iSPIRT teams who did the design work. The test now is speed: pilots that run, a consortium that governs and rules simple enough for a ten-person startup to use.

What a founder takes from it: Compute and models get the headlines. Clean, lawful, Indian data is the harder input, and this is the first serious design to supply it at scale. Founders building for Indian credit, health and education should read the draft now and help shape it, because the rules written in the pilots will become the rails.