Public records and identity

What “records and identity” actually contains

The public-records side of the Nepali state is best thought of as four interconnected stacks:

Civil identity. Citizenship certificate, vital registration (birth, death, marriage), national ID. The country has a national ID rollout in progress that, if completed, would give every resident a single durable identifier — and would make most of the other stacks coherent for the first time.

Property. Land records held in malpot offices, the cadastral mapping at the Department of Survey, urban property records held by municipalities.

Public-service touchpoints. Voter rolls, drivers’ licences, PAN, social security registration, foreign employment records, education transcripts.

Justice and dispute. Court filings and judgments at every level of the judiciary, police records of complaints, transitional-justice records.

Each of these stacks is uneven in its digitisation. Some are mostly paper (court records). Some are partly digitised but trapped in silos (land records). A few are well-digitised (PAN, voter rolls). None of them talks to the others, which means that the state, as an information system, knows you four times over and has to reconcile its own knowledge by phone.

What AI can usefully do here

Three patterns of AI work produce real public value in this domain.

OCR and structuring of legacy paper. The single largest productivity gain available is turning the country’s accumulated paper records into structured, searchable data. Devanagari OCR (Chapter 2) plus careful manual schema design can take, say, fifty years of Supreme Court judgments and make them searchable in seconds. This is unsexy, multi-year work; the leverage is enormous.

Entity resolution and deduplication. The same person appears in the citizenship register as राम बहादुर श्रेष्ठ, in the PAN system as Ram Bahadur Shrestha, in the voter roll as रामबहादुर श्रेष्ठ, and in their school transcript with their father’s name as the last surname. Matching these is exactly what modern entity-resolution models do well. The work has obvious privacy implications and must be done carefully — but done well, it removes one of the largest sources of friction in Nepali public services.

Plain-language interfaces to public information. A citizen who wants to know how to renew their driving licence does not want to read a 60-page traffic regulation. An LLM-powered service that reads the underlying rules and answers in plain Nepali — and links back to the original documents — is genuinely useful, and is technically straightforward to ship. Several Nepali developers have prototyped versions of this.

The privacy frame that has to come with it

Nepal does not yet have a comprehensive data-protection regime in the way many of its trade partners do. The 2018 Individual Privacy Act exists but is thin on operational specifics for AI use. The 2074 (2018 BS) Electronic Transactions Act predates LLMs entirely.

This is not a complaint; many countries are at the same place. It is, however, the single most important policy gap to close before extensive AI is deployed against state records. A workable Nepali framework would need to address:

• Purpose limitation — what is each record set allowed to be used for, by whom, with what consent?
• Algorithmic decision rights — when an automated system makes a decision affecting a citizen, what right does the citizen have to know, contest, and seek a human review?
• Cross-stack matching consent — what consent is required to link a citizen’s records across the four stacks above?
• Vendor accountability — when a foreign vendor’s model misclassifies a citizen, who is liable?
• Data localisation — what citizen data can leave Nepali jurisdiction, under what conditions?

These are not technical questions. They are political ones. The country will be better off facing them deliberately than letting them be decided implicitly by whichever vendor’s procurement contract gets signed first.

The sovereignty side

There is a related question that is not strictly about privacy but is related: when AI processes Nepali citizen records, where does the computation happen, and who owns it?

The cheapest path today is to use foreign cloud-hosted AI services. This is operationally fine for many things; for citizen records, it raises the same kinds of questions defence ministries have asked about cloud for decades. A future where most decisions affecting Nepali citizens are computed on infrastructure owned by US-listed firms, subject to US legal process, is a future that is worth thinking about now rather than later.

There is no easy answer. Building domestic compute is expensive. Using domestic-but-mediocre models is a real cost in service quality. The most honest position is that this is a trade-off the country has to negotiate consciously, rather than drift into. We return to it in Chapter 6.

Check your understanding

Quick check

—

A team is proposing to use AI to resolve and deduplicate citizen records across the citizenship register, PAN system, and voter roll. What is the most important condition for this work to be net beneficial to citizens?

1. AUse the fastest available model regardless of vendor.
2. BA clear legal framework — covering consent, purpose limitation, algorithmic decision rights, and accountability — must be in place before the cross-stack matching is operationalised at scale.
3. CMake the work as cheap as possible by using only foreign cloud services.
4. DAvoid telling citizens that the work is happening.

Check answer→

What comes next

Health and identity are two of the three areas where AI sits closest to public good. The last is education — both the long-running literacy gap that has shaped half the country’s adult lives, and the much fresher question of what classrooms should do now that any student can summon a competent essay-writer on their phone.