Govt rolls out data-sharing framework to boost AI training while protecting privacy

DEPA for AI Training is a techno-legal framework launched by MeitY to make data sharing for AI training safer. It matters because institutions often hold useful datasets but hesitate to share them. Their concerns include privacy, confidentiality, intellectual property and control over information.

The framework enables purpose-bound collaboration. Institutions can permit datasets to be used for a defined AI-training purpose without handing over permanent control. Contractual permissions, controlled computing environments and privacy-enhancing safeguards work together. Data-provenance tools record where data came from, what permissions apply and how it is used.

The framework was unveiled by MeitY secretary S Krishnan in Bengaluru, with support from the iSPIRT Foundation. It aims to unlock large-scale datasets while protecting privacy, commercial confidentiality and institutional control. The article presents DEPA as a possible foundation for trusted collaboration among institutions and across national borders.

Institutions can share data for AI training without giving up control of the underlying information. DEPA creates rules for purpose-bound collaboration, so access is connected to an agreed use rather than becoming unrestricted possession. This addresses privacy, confidentiality and intellectual-property concerns.

The key mechanism combines contractual permissions with controlled computing environments and privacy-enhancing safeguards. For example, a hospital or research institution could allow a model to train against its dataset within an approved environment, while the dataset remains under institutional control. Provenance records track the data’s origin, permissions and use during training.

This approach is intended to make trusted collaboration possible at scale. The article says it could support cooperation among multiple institutions and across national borders. It also aims to give researchers access to more diverse datasets while protecting commercial confidentiality and the rights of data custodians and contributors.

The phrase describes a reversed data-sharing model. Instead of copying institutional datasets into a central AI system, the model accesses those datasets where they already reside. The data stays with the institution, which helps preserve control and reduces the need for broad transfers.

In practice, an AI model could run inside a controlled computing environment approved by the data-holding institution. Contractual permissions would define the purpose, while privacy safeguards and provenance records would help protect and track the activity. The model learns from the data without requiring the underlying information to be moved elsewhere.

Principal Scientific Adviser Ajay Kumar Sood used this phrase to explain DEPA’s approach. He said it could support collaboration among multiple institutions and across national borders. The model therefore becomes the moving element, while institutions retain custody of their datasets and can participate in shared AI training with fewer transfer-related risks.

The article does not state a typical training-data volume. Established industry practice is that modern large AI models may use datasets containing billions or even trillions of tokens, alongside large collections of images, audio or other records. The exact amount varies by model and task.

At that scale, data may come from many institutions and sources, each with different permissions, sensitivities and quality levels. Copying everything into one central system can increase exposure, complicate confidentiality and make intellectual-property rights harder to track. It also becomes difficult to confirm how every item was obtained and used.

DEPA addresses this challenge by combining purpose-bound contracts, controlled computing environments, privacy-enhancing safeguards and provenance mechanisms. These tools are intended to make large-scale collaboration more accountable. The article’s framework focuses on enabling access without requiring institutions to transfer their underlying information.

If researchers can use more diverse datasets without transferring the underlying information, AI development could become broader and more useful. Models would have access to perspectives and evidence from multiple institutions while data custodians keep control. Kris Gopalakrishnan said this could help researchers develop more robust AI models.

The practical mechanism is controlled collaboration. A model can work with datasets inside approved environments, under contractual permissions and privacy safeguards. Provenance records can show where the data came from, what permissions governed it and how it was used. This allows institutions to contribute without treating unrestricted data transfer as the only option.

The article also links the approach to faster scientific discovery. Ajay Kumar Sood said DEPA could enable trusted cooperation among institutions and across national borders. He described it as India’s contribution to global science diplomacy and an inclusive, responsible global AI ecosystem.

The article identifies institutions as data holders and refers to the rights of data custodians and contributors. DEPA is designed so participation in AI training does not mean surrendering institutional control. It protects the interests of those who provide or steward the information, although the article does not specify a single ownership rule.

Contracts are the first safeguard. They can grant permission for a defined purpose, setting boundaries around how data may be used. Controlled computing environments and privacy-enhancing measures add technical protection. Provenance records then create an audit trail by recording the data’s origin, governing permissions and use during AI training.

Together, these measures support accountability and transparency. They can help institutions understand whether data is being used as agreed and help protect commercial confidentiality. The framework’s goal is trusted collaboration, not unrestricted pooling of information. Its approach keeps control with the relevant institutions while enabling useful model development.

AI training is the process in which a model learns patterns from examples in data. The model adjusts its internal parameters while processing those examples, so it can later generate predictions, classifications or other outputs. The article focuses on making this training possible through responsible data collaboration.

The amount of data matters because more relevant examples can expose a model to more patterns. Quality matters because inaccurate, incomplete or poorly labelled information can teach unreliable patterns. Diversity matters because datasets representing different people, settings or conditions can help a model work more robustly beyond a narrow source. These are general AI principles; the article specifically highlights access to large-scale and diverse datasets.

DEPA aims to expand that access without requiring institutions to transfer underlying information. Its contracts, safeguards and provenance records are intended to make collaboration safer. Researchers could therefore use broader datasets while institutions protect privacy, confidentiality, intellectual property and control.