Artificial Intelligence, Algorithms and Administrative Conscience

A welfare algorithm can calculate an eligibility score. It cannot be summoned to a departmental inquiry, and it cannot feel the weight of having been wrong. That gap is the entire ethics of this chapter.

Why this is an ethics question and not only a technology question

Algorithmic tools are already inside Indian governance, not on some future horizon: eligibility scoring for welfare schemes, Aadhaar-based deduplication of ration cards, facial recognition at public events, predictive-policing pilots that flag neighbourhoods for patrol, chatbot triage of citizen grievances. The ethical question is not whether these tools are accurate — that is an engineering question — but where moral responsibility sits once a machine’s output shapes what happens to a specific citizen.

Where conscience and code diverge

An algorithm optimises for the metric it was given. Conscience answers to the person standing in front of the counter, not to the metric. When a ration-eligibility system wrongly excludes a genuine beneficiary because a fingerprint failed to match or a name was spelled differently across two databases, the system is not malfunctioning by its own standard — it did exactly what it was built to do. What it lacks is the officer’s obligation to ask whether this particular exclusion is just, a question no scoring function is designed to ask.

This is the core administrative principle: automation may assist discretion, but it may not replace the duty of individualised justice. Natural justice — the right to be heard before an adverse action — does not become optional because the adverse action was recommended by software rather than typed by a clerk.

Algorithmic bias and the opacity problem

Bias typically enters through training data, not through any coder’s intent. A predictive-policing model trained on historical arrest records will, without correction, send more patrols to the neighbourhoods that were already over-policed — reproducing yesterday’s bias as tomorrow’s forecast and calling it evidence. Many such models are also functionally a black box: even the officials who deploy them cannot always explain why a particular case was flagged, which puts them in direct tension with a citizen’s right to reasons, a settled requirement of Indian administrative law.

This raises the accountability question sharply. When an algorithm errs, who answers for it — the vendor who built the model, the department that procured it, or the officer who signed off on its output? Departmental inquiries and courts can summon officers. They cannot cross-examine a model. Which means an administration that deploys algorithmic tools must, as a matter of design and not afterthought, keep a named human accountability anchor for every automated recommendation that touches a citizen’s entitlement, liberty or dignity.

Working principles for an administrator

Treat algorithmic output as a recommendation, never a determination, wherever entitlements, liberty or dignity are at stake. Preserve a genuine right of appeal and explanation, so that “the system flagged you” is always followed by a human who can say why and can be argued with. Insist on periodic bias audits and disclosure of the categories of data a model was trained on, rather than accepting a vendor’s assurance of neutrality on faith. And hold the line on the principle Chapter 15 already establishes for the illegal order: an officer cannot outsource conscience to a superior’s command, and cannot outsource it to a machine’s output either. “The system decided” is not a defence; it is an admission that no one was actually deciding.