MeitY circulates draft DEPA-for-AI training framework with pre-deployment checks for models

The Ministry of Electronics and IT has put out a draft framework, described in coverage as DEPA for AI Training, setting out how AI models could be trained on data held by institutions such as hospitals, banks and government agencies. Reports place its release for consultation in the final days of October, with NewsBytes covering it on 4 November.

Under the design reported by ThePrint, data holders are termed Training Data Providers, while startups and researchers are Training Data Consumers. Training takes place inside a secure environment rather than on copies of the data, builders submit jobs and receive only the trained model, and every step is logged for audit.

The draft also contemplates checks before deployment. Coverage says models would be validated for fairness and bias and verified as fit for use before approval. NewsBytes reports participation is voluntary for now, and legal experts quoted by The420.in raise the possibility that such verification could eventually resemble licensing.

We have not seen the draft text itself, so details on the approval process and penalties rest on press coverage and may change.

Why it matters

Access to high-quality Indian data is a bottleneck for domain models in health and finance. A consent-aware, auditable route to that data could unlock training that is currently blocked by privacy concerns, but unclear approval rules could deter smaller builders.

Indicops view

Data-holding institutions should begin mapping which datasets could be offered under such a scheme and what logging they can support. Model builders should prepare bias and fitness documentation now, and respond to the consultation to keep approval criteria specific and proportionate.