Why AI readiness rankings are misleading India

Every year, global AI readiness rankings make their predictable rounds through policy circles. Countries are scored, ranked, and colour-coded. Governments cite them in speeches. Ministries use them into strategy documents. Headlines celebrate small climbs up the table, as if intelligence itself were a league competition.

The underlying assumption is rarely questioned: artificial intelligence is something Governments acquire. That assumption is wrong. AI is not an asset a State accumulates, like foreign exchange reserves or computing infrastructure.

It is a force that changes how decisions are made, who makes them, and who can be held responsible when things go wrong. Any framework that treats AI readiness as growth — more models, more pilots, more data, more compute — is not measuring readiness. It is measuring exposure.

This is not a philosophical argument. It is an administrative one. AI systems are not static infrastructure. They are continuously operating decision filters. Once embedded in welfare delivery, taxation, compliance, policing, credit allocation, or regulatory enforcement, they do more than “assist” officials.

They alter workflows. They set the pace of decisions. They determine who is flagged, delayed, denied, or approved — often without anyone actively choosing each outcome.The real question, then, is not whether a country can deploy AI. Many can. The harder question is whether a state can retain control once automated decisions operate at scale.Control, in institutional terms, is concrete. Can automated systems be paused during crises without collapsing service delivery? Can frontline officials override individual decisions quickly, without fear of audit penalties or disciplinary action? When an automated decision causes harm, does responsibility land clearly on a named authority — or dissolve into technical explanations and vendor contracts?

On these questions, India, the United States, and the European Union are all far less prepared than their rankings suggest — each in a different way.The global fixation on AI readiness rankings — exemplified by instruments such as the Stanford AI Index — rests on an outdated mental model. These frameworks reward visible capacity: research output, talent concentration, compute access, startup ecosystems, venture capital flows, and public funding announcements. Progress is treated as additive. More of everything is assumed to mean greater readiness. What these rankings do not measure is reversibility. They do not ask how many automated decisions a Government can realistically review or overturn in a day. They do not track appeal capacity relative to automated volume. They do not measure how long it takes to correct a wrong decision, or whether correction is even possible once a system is fully embedded. Deployment counts as success. Reversal barely registers.

This blind spot is not accidental. It shows political incentives.Automation is attractive because it scatters accountability. When decisions are mediated by models, responsibility fragments. Officials defer to systems. Ministries defer to vendors. Vendors defer to technical complexity. Outcomes occur, but ownership becomes unclear. Failures rarely arrive as dramatic crashes. They accumulate quietly — as delays, denials, queues, and appeals that never quite resolve.

Readiness rankings normalise this condition by treating deployment itself as progress.The failure mode is predictable. As AI systems scale, sensing capacity expands far faster than decision authority. Automated systems can flag millions of cases, but bureaucracies cannot review millions of appeals. Humans remain legally responsible, but operationally constrained. Overrides exist on paper, but are difficult to use. Appeals exist, but move slowly. Explanations arrive, if at all, after harm has already occurred.

India encounters this problem first because it governs at population scale. On paper, India performs well across AI readiness frameworks. It has a large technical workforce, expanding digital public infrastructure, and ambitious national AI programmes. Systems linked to Aadhaar now mediate welfare delivery, subsidies, and access to public services for hundreds of millions of people. At that scale, even small error rates become large human problems.

A 2022 audit by the Comptroller and Auditor General of India documented widespread welfare denials caused not by ineligibility, but by authentication failures, data mismatches, and grievance mechanisms that frontline officials could not override in real time. Parliamentary committees have repeatedly flagged the same pattern: automated checks trigger exclusion, while correction pathways remain slow, opaque, and weakly integrated into everyday administration.

Many of these failures predate widespread AI-driven risk scoring. India’s experience shows not reckless use of technology, but rigid institutions. When automation is added to inflexible systems, errors spread faster and responsibility gets pushed across ministries, states, and private vendors. Rules are followed. Justice becomes harder to obtain.

If AI systems now expand into taxation, credit allocation, predictive compliance, or fraud detection without a parallel expansion of appeal and override capacity, exclusion will not merely scale. It will become routine.The United States presents a different failure mode. The US leads global rankings on research output, frontier models, venture capital, and compute access. Yet authority is fragmented across federal agencies, states, courts, and private contractors. AI systems increasingly influence credit decisions, hiring, healthcare triage, policing, and benefits eligibility, while responsibility for outcomes remains dispersed.

When automated systems fail, correction often arrives through litigation rather than administration. Courts become the mechanism of reversal. This may protect individual rights, but it does not scale. Legal redress grows slowly. Automated harm grows quickly.

The result is familiar: Rising legal backlogs, inconsistent rulings, delayed policy correction, and governance backed by lawsuits rather than design. Readiness rankings treat this legal capacity as strength. In practice, fragility is masked as resilience. Litigation is not controlled. It is what happens after control has already failed. Europe represents a third path. The European Union has responded to AI by regulating it in advance.

Its framework emphasises risk classification, compliance obligations, documentation, and safeguards. On paper, this appears cautious and humane. But regulation does not automatically restore control.

Once AI systems are embedded in administrative workflows, reversibility remains limited. Oversight bodies are often under-resourced. Appeals move slowly across member states. Accountability is split between national regulators, EU institutions, and private firms. Compliance may be formalised, but the ability to halt systems quickly or correct outcomes at scale remains weak.

Europe has built guardrails. It has not yet solved the actuation. Across all three cases, readiness rankings miss the same core issue: the capacity to stop systems cleanly, reverse decisions quickly, and assign responsibility clearly when automation fails. That is why the concept of “AI readiness” itself is unstable. A state that deploys AI faster than it can reverse automated outcomes is not technologically advanced. It is accumulating institutional debt. This debt compounds over time. Once automated systems restructure accountability pathways, rolling them back becomes politically costly and administratively disruptive. At that point, even known failures are tolerated because correction itself threatens the appearance of control.

For India, the implications are immediate and unavoidable. India does not yet lack AI capability. It lacks institutional braking power.

A system that fails one per cent of the time does not produce inconvenience at India’s scale. It produces exclusion. When automated decisions cannot be paused during crises, when frontline officials cannot override systems without fear of audit penalties, and when responsibility for failure cannot be clearly named, the result is not efficiency. It is administrative injustice delivered at speed.

The question India must answer is not whether it can deploy AI faster than others. It already can. The real test is whether it can stop automated systems when they begin to fail—cleanly, quickly, and without chaos.

The next phase of AI power will not belong to the countries that deploy fastest or score highest. It will belong to those that can halt systems without paralysis, reverse outcomes without litigation explosions, and govern automation without surrendering authority to it. Until rankings measure that capacity, they are not indicators of preparedness. They are indicators of how comfortable a State has become with losing control.

Author is an theoretical physicist at the University of North Carolina at Chapel Hill, an AI advisor, and the author of the forthcoming book The Last Equation Before Silence; views are personal