Digital public infrastructure built the world's best plumbing for permission. The missing record is comprehension — and India is the only country positioned to standardise it.
This is not an edge case. At India's scale — twenty-two scheduled languages, a Language Diversity Index of 0.93, six meaningfully different literacy bands — the gap between captured consent and actual comprehension is the median case.
The consent artefact gave permission a standard form: open, revocable, granular, auditable. The Account Aggregator framework operationalised it. The data protection law made consent managers a regulated category. No other country has plumbing this good for permission.
Legal scholars reviewing consent managers have already named the risk: a “consent mirage” — opt-ins accumulating without understanding. The law gestures at the problem (notices must be available in scheduled languages), but translation is not comprehension. A perfectly translated paragraph of legalese is still legalese.
Informed-consent law requires consent in a language the patient understands — and clinical practice built the instrument to verify it: the patient explains the procedure back in their own words before signing. Cheap, validated across hundreds of studies, and framed as a test of the explanation, not the patient.
Created because loan disclosures were technically complete and functionally unread. It compresses a loan's terms into a page a borrower can actually parse — regulation moving from disclosure toward comprehension.
Consent requests must be in “clear and plain language” — unintelligible consent is legally void. The burden of being understood sits with the system, not the citizen.
AI systems now carry explicit obligations to be understandable to the people they affect. The same wound, twenty-four languages, no instrument.
Call it the comprehension artefact — the human-readable counterpart to the consent artefact.
Persona-based design collapses at a billion users. What scales is classifying the materiality of each step, so friction lands only where the stakes do.
Browsing a screen. The system stays invisible.
Choosing an option. Light guidance, an undo path.
Submitting, paying. A plain-language account of what happens next.
Irreversible acts — biometric sharing, loan agreements. A verified checkpoint.
A failed check routes to help, not to a wall — a family member, a banking correspondent, a support line. This honours a reality every Indian designer knows and no Western framework models: comprehension here is often a social act. The person tapping the screen is not always the person deciding. The artefact can record that too — who assisted, in what role — turning the most uncomfortable fact about consent at scale into structured, auditable data.
Every checkpoint outcome is a labeled data point: this explanation, in this language, at this reading level, did or did not land with this kind of user. Aggregated, that is the largest corpus of verified Indic comprehension data ever assembled — exactly the alignment-layer data Indian language models lack, generated as a by-product of protecting users.
The first wave of DPI exported permission infrastructure; UPI and the consent artefact are studied worldwide. The comprehension artefact is the natural successor — Europe's AI transparency rules, mis-selling settlements everywhere, every consent banner ever clicked unread. Whoever defines the metric defines the standard.
Verification slows the funnel. A recorded failure obligates the platform to act on it. Institutions will resist creating a record that can pause their own conversion.
The liability already exists — it is simply unpriced. Comprehension failures currently surface downstream as fraud claims, complaints and fines. The artefact moves the cost upstream, where it is smaller, and gives the institution something it has never had: proof that it explained itself.
The models are ready — sovereign, multilingual, voice-first. The missing piece is a ten-page specification: a schema, a materiality taxonomy, a verification method, conformance tests. A working proof of concept already exists.