An intelligence inside the state cannot be allowed to roam like a mythological animal merely because it is efficient. Oversight must be proportional to consequence.
In a district court in Uttar Pradesh, somewhere in the old dust of files and adjournments and stamped certainties, a translation engine, fed on three decades of judgments, is silently eating into a backlog which human clerks, with all their fatigue and tea and memory, could not really touch. Elsewhere, in a gram panchayat in rural India, a transcription tool sits through the meeting like an invisible munshi, and the moment voices fall silent, minutes exist; a record which earlier took days, or never took form at all, now congeals almost immediately into governance. No trumpet sounded, no headline trembled. Yet, taken together, these small events probably reveal more about the entry of artificial intelligence into Indian governance than the larger ceremonies of announcement ever could.
This moment belongs less to software manuals and more to civilizational history. It may prove to be an inflection larger than the printing press or the steam engine, because the machines are no longer merely moving things, storing things, accelerating things; they are beginning, in some strange derivative sense, to think. The claim is large, maybe even vertiginous, and it cannot pretend that the shadows are ornamental. Jobs will move, shrink, mutate; data centres will drink energy and water; every intelligence layer has an ecological and human underside. But the more interesting question is not of danger alone, it is of position. On the usual scorecard of technology, data, compute, models, expertise, India does not stand at the summit. What it has, already spread like a nervous system under the skin of the republic, is Digital Public Infrastructure: Aadhaar as identity, payment rails as circulation, consent architecture as connective tissue. If this is the substrate, then AI is not a mountain India has to climb in imitation of others; it becomes an intelligence layer which can be placed upon an already living grid.
This shift of metaphor matters. It explains why the most consequential AI in Indian government does not look like a shiny chatbot, polite and sterile, answering questions in the accent of nowhere. The first barrier between citizen and state has always been language, that old gatekeeper sitting at the counter. Roughly nine in ten Indians do not speak or read English fluently, and therefore governance, however benevolent in design, often arrives in a tongue which is not the citizen’s own. Bhashini, a transcription platform, is infrastructure made for precisely this wound. Gram panchayats using it have already generated records of more than fifty lakh meetings, minutes which once depended on a clerk’s memory, mood, pen, patience, and were more often than not lost into the grey fog of unwritten administration. Voice interfaces, growing from the same assumption, that a person should speak in her own language and in whatever literacy she possesses, are doing quietly what campaigns of inclusion only proclaim. Alongside this, the translation mission of government has expanded its own web, moving nearly 300 million translations every month across four dozen portals; language, that old moat, is being crossed by machines which do not know poetry, but may yet know access.
Beyond the answering machine lies the next creature, the transactional layer, the system which does not merely reply but anticipates, joins fragments, releases action. Crop-loss compensation offers the shape of it. In principle, satellite data, weather records, identity, land records and scheme rules can be stitched into one administrative organism, so that a farmer need not file a claim at all; eligibility is determined, loss is seen, payment moves. Fraud detection has already walked some distance on this path. A Comptroller and Auditor General report found that machine-learning analysis of Aadhaar-linked direct benefit transfer data helped save nearly three lakh crore rupees by identifying leakages which manual audit had missed for years. Haryana’s Parivar Pehchan Patra is another expression of the same instinct, linking each family’s welfare eligibility to birth, death and marriage registers, so that an old-age pension can awaken automatically in the month a resident turns sixty. Predictive governance, then, is not a distant spaceship hovering in policy dreams; it is the next iteration of systems already breathing.
Inside government, where memory often migrates with transfers and dies in old files, the efficiency argument has become concrete. Maharashtra’s information technology department has built Maha AI, trained not on the vague totality of the internet but on the state’s own government resolutions, notings and official files. It allows civil servants to draft cabinet notes, proposals and requests for proposals in Marathi, without depending on whichever officer happens to remember how a similar file was processed three years ago. This is not glamour, not the theatrical dream of a sovereign omniscient model sitting above the nation like a metallic god. It is smaller, and perhaps for that reason wiser. Sovereign AI, in this view, need not be one large language model pretending to do everything; it may be a federation of specific intelligences, a drafting assistant here, a PM-KISAN module there, answering a farmer about pest control or soil health, each tool shaped by people who understand the exact problem, the local grammar of failure, the little bureaucratic demon it is meant to tame.
But an intelligence inside the state cannot be allowed to roam like a mythological animal merely because it is efficient. Oversight must be proportional to consequence. The judiciary’s experiments with AI research tools remain assistive, not decisional, because hallucination in a courtroom is not a philosophical inconvenience; it is injury in the language of law. Systems have already been shown to cite Supreme Court judgments that do not exist, phantom precedents born from statistical confidence. For higher-risk domains, the safer architecture is a two-key mechanism: one AI proposes an action, another AI independently interrogates the first, and a human being makes the final call. This is more demanding than the lazy phrase “human in the loop,” which policy documents love like an amulet. A human pasted at the end of an automated pipe can become a rubber stamp very quickly; scrutiny must be designed into the flow, not sprinkled over it like holy water.
Explainability is not decorative, not a kindness offered by engineers to anxious citizens. It is a legal and democratic necessity. The power asymmetry between citizen and state is already steep; when a government decision cannot explain itself, it becomes not merely opaque but challengeable, and under the Digital Personal Data Protection Act, potentially vulnerable in law. Facial recognition gives the warning in a clean, almost cruel form. American deployments trained overwhelmingly on white male faces have shown measurably poorer accuracy for people of colour. For India, with its densities of caste, region, gender, class, skin tone, age, language and visibility, procurement without asking what a system was trained on, and whom it was trained to see clearly, is not modernization; it is blindness automated.
The nearest danger, however, may not come from sanctioned pilots at all. Bhashini, Maha AI, fraud-detection models on DBT data, these are visible creatures, named and placed inside institutional frames. The more immediate risk is Shadow AI, the unsanctioned, already creeping habit of officials using personal accounts on public tools to draft notes, summarise files, polish letters, or digest government documents, outside any approved system, without the department even knowing that such leakage has begun. This is not an abstract cyber sermon. Uploading an official file into a consumer chatbot can constitute an offence under the Official Secrets Act; if that file contains personal data, the department may also become liable under the DPDP Act. The demon here is banal, almost comic: not a rogue superintelligence, but a tired officer, a pending note, a free tool, and one confidential file poured into a machine whose memory belongs elsewhere.
What comes into view is a country that has built rails for this new force with more discipline than it has yet written the rules for running trains safely upon them. The infrastructure is real; the early use cases are not imaginary ornaments but working devices, clearing backlogs, recording meetings, translating portals, detecting leakages, drafting files, triggering benefits. The judiciary’s insistence that a human remain the final authority shows that caution and usefulness need not be enemies, that restraint can coexist with velocity. Whether the rest of government learns this grammar, or lets Shadow AI, bad procurement and unexplained automation outrun the rails, will decide the memory of this transformation. It may remain another announcement, garlanded and forgotten; or it may become an outcome, absorbed into the ordinary machinery of the republic, like electricity, like language, like a record finally written.
*Brijesh Singh is a senior IPS officer and an author (@brijeshbsingh on X). His latest book on ancient India, “The Cloud Chariot” (Penguin) is out on stands. Views are personal.
