Medha Thapliyal

Sheet 03 Writing Revised October 2026

Sheet 03/07. 4 May 2026. 3 minute read.

ABHA before algorithms

India's bet is that identity, registries, and consent come before models. I think the consent manager, not the model, is the real AI policy.

03/07 · Essay

By the government’s own count, the Ayushman Bharat Digital Mission has issued health identifiers to over ninety crore people, linked more than a hundred crore records, and registered over five lakh facilities. Whatever one thinks of the figures, the architecture behind them is the most consequential AI policy decision India has made, and it does not mention AI.

The decision is sequencing. The ministry’s stated position, repeated at a regional digital health summit last year, is that identifiers, registries, and consent frameworks are prerequisites for AI at scale, not afterthoughts. Build the rails first. Decide who can see what, and on whose say-so, before anyone trains a model on it.

In the hepatitis work I contributed to, the single largest improvement in program monitoring did not come from a better analytical tool. It came from redesigning the forms so that a screening result, a referral, and a treatment initiation were recorded as linked events rather than three separate tallies. The data flow was the intervention. The dashboard was a view of it.

ABDM’s consent manager is the same idea at national scale. It separates the question of whether data exists from the question of who may use it, and it makes the second question answerable by the person the data describes. For AI, this is the whole game. Every proposal to train a model on health records eventually runs into the secondary-use problem: the data was collected for care, and the model is not care. A consent architecture that can express “my records may be used to treat me but not to train your model” is a policy instrument. A model governance framework that assumes the data is already available is a press release.

What is still missing

I do not want to be naive about this. The consent manager is a specification, and specifications are not the same as practice. Consent in an overcrowded outpatient department is often a thumbprint on a form nobody read. The ministry is rolling out a clinical decision support system built at a national institute to tens of thousands of hospitals, and I have not seen a public evaluation plan, a model-version policy, or an account of how a wrong suggestion in a district hospital gets reported and fixed. The WHO’s guidance on large multimodal models asks for exactly those things, including mandatory post-release audits, and it notes something that rarely appears in Indian policy discussion: the compute and water costs of these models fall disproportionately on countries that did not build them.

There is also no linkage, in the published literature, between ABDM and disease surveillance in the way people sometimes assume. Hepatitis A and E outbreaks are still detected the way they were before: through the integrated disease surveillance program’s water-outbreak reports. The rails exist. The trains, for surveillance at least, have not been built.

The argument

My argument is narrow. If you want to know whether a country is serious about AI in health, do not read its AI strategy. Read its consent architecture, its registry quality, and its facility-level data flows. India has made an unusual and defensible bet that those come first. The test, over the next few years, is whether the models that arrive on the rails are held to the same standard of accountability that the rails themselves were designed to make possible.

Sources

03 · Adjoining

Keep reading.

  1. Apr 2026 Five thousand alerts, zero outcome studies India's media-scanning surveillance system has flagged thousands of potential outbreaks. Nobody has yet shown it changed the course of one. Surveillance AI needs the discipline we demand of drugs.
  2. Aug 2026 The red alert and the ASHA A clinical copilot in Nairobi cut diagnostic errors most when it interrupted. What would an interruption mean for a community health worker with no clinical backstop?
  3. Mar 2026 Consulting without selling Most AI readiness tools for non-profits are written by vendors. Responsible advice often means recommending no AI at all, or a spreadsheet.