MeitY’s DEPA for AI Training Handbook is an attempt to capitalise on the reputation of India’s innovative DEPA framework. But instead of empowerment, it takes control of data away from data principals and passes off tried-and-tested clean room technology as a novel Indian innovation. This is embarrassing.

Last week, the Ministry of Electronics and Information Technology (MeitY) released a new framework aimed at breaking the “data deadlock” and democratizing data access to build artificial intelligence (AI) models. This new “Data Empowerment and Protection Architecture (DEPA) for AI Training” framework purports to enable responsible data sharing while ensuring privacy, transparency and trust.

I am not convinced it will achieve its aims. Not only is this not a novel architecture, it seems to subvert much of what India’s DEPA was designed to achieve. Decisions around what can be done with personal data, which under DEPA were placed in the hands of the data principal, will now be taken by the institutions that hold that data as fiduciaries—quite literally stripping away DEPA’s primary contribution to India’s data governance discourse.

The Original DEPA Innovation

Europe’s General Data Protection Regulation (GDPR) confers on individuals a right to data portability—to request that their personal data be transferred either to them or to another service provider. Instead of prescribing how this should be implemented, GDPR left it to businesses to work it out for themselves. India, on the other hand, built DEPA, a data portability framework that gives individuals granular control over who their personal data is shared with and for what purpose. As a result, it empowers them with the tools they need to realize the value of their digital exhaust. And when we implemented the Account Aggregator framework in the financial sector, we proved that micro-portability could work at scale.

If there was one mistake we made, it was calling DEPA a “consent management framework” instead of a portability solution. As a result, the term “consent managers” is now part of the Digital Personal Data Protection (DPDP) Act, even though that term has a very different meaning in the data protection context. This is why so many products marketed as consent management solutions bear no resemblance to how the term is used in the DEPA architecture.

But having made that one mistake, we seem determined to make things worse.

This is Not Empowerment

The DEPA for AI Training Handbook starts with the premise that individual consent is impractical at scale. On that basis, it takes the power to determine what is done with personal data away from the data principal and vests it in the hands of training data providers (hospitals, banks and the like), who will now be responsible for sourcing, curating and preparing the datasets for AI training. I have long argued that consent is broken, and even that using clean rooms would strengthen DEPA. But DEPA is the one place where consent works, and when I suggested we use clean rooms, it was to enforce the consent provided, not to replace it.

The handbook also conveniently overlooks all that happens downstream of AI training. Since AI models are, in effect, massive data compression systems, they only need to be trained once on a given dataset to extract all the value they can from it. The handbook does nothing to ensure that the data principals stand to benefit from the AI models they helped create—other than suggesting that training data providers could receive priority access to the resulting models and promising to “explore” royalties for them. In other words, direct benefits will accrue only to model developers and the institutions that hold personal data, not to those whose data was actually used. It is hard to see how this framework can claim to treat the rights of data principals as “the bedrock upon which the entire architecture is built.”

This is not Novel

But perhaps most disappointing of all is the claim that this is a uniquely Indian innovation—what the principal scientific advisor calls in his foreword, “an elegant, India-specific solution.” Secure data environments are nothing new. In 2019, Finland’s law on Secondary Use of Health and Social Data required that individual-level data be analysed only in secure operating environments, and the UK’s 2022 Data Saves Lives strategy made secure data environments the default for accessing National Health Service data. The handbook’s own evidence that the approach works is a 2022 project in which 71 institutions, led by Penn Medicine, trained a brain tumour model using federated learning within a “confidential clean room.” Calling this idea novel in the face of such a long history of use is disingenuous.

As accustomed as I am to techno-optimists masquerading as experts in legal design, the failings evident in this 40-page handbook exceed anything I have seen before. It is a word salad of technological jargon and regulatory buzzwords that promises more than it can hope to deliver. Electronics and Information Technology Minister Ashwini Vaishnaw, in his foreword, assures us that it is “consent-based” and “user-controlled” even though the rest of the document says the contrary.

It was troubling enough when we took a novel and innovative data portability architecture and obfuscated the power of that innovation by mischaracterising it as a consent management framework. But that seems trivial compared to what we are now doing—taking a tried-and-tested clean room system and calling it the next version of the DEPA framework. Having not yet recovered from the damage caused by the first misapplied label, we seem bent on making the same mistake again.

At this rate, the only thing truly Indian about DEPA will be the confusion we will create every time we do something novel with it.