Overview
- MeitY unveiled the draft DEPA for AI Training handbook on Friday in Bengaluru, presenting a non-binding proposal developed primarily with the iSPIRT think‑tank.
- The design routes model training into confidential clean rooms so AI developers never see underlying datasets, with access enforced by electronic contracts, zero‑trust security and differential privacy.
- Early pilots were shown in finance and healthcare, including a bank test that handled over two million requests in two weeks and a tumour‑detection model trained on MRI scans the developer never viewed.
- The handbook requires independent checks of trained models for bias, fairness, security, safety and explainability and proposes Samyog to handle day‑to‑day participant registration and contracts.
- Formal legal adoption and liability rules are unresolved, so the next steps are multi‑stakeholder consultations, regulatory sandboxes and alignment with the DPDP Act before any binding regime is set.