The Real AI Divide Is Not Intelligence, but Agency

What if the most important question about artificial intelligence is not whether it can make decisions, but whether more people can still make meaningful choices after it arrives?

This question matters everywhere, but it matters especially in countries where institutions are under strain, infrastructure is uneven, and a large share of the population has never been well served by existing systems. In such places, AI is often presented as a shortcut to development: a way to improve medical diagnosis, increase agricultural yields, personalize education, strengthen public administration, and create new industries without first reproducing every stage of industrial history.

That promise is real. But it hides a danger. A technology can expand the number of choices available to a society while quietly narrowing the number of choices available to individuals. It can make a ministry more efficient while making citizens less able to contest its decisions. It can give a farmer a better forecast while making her dependent on a platform she cannot inspect or leave. It can make education more accessible while reducing learning to the optimization of measurable answers.

The central issue, then, is not simply whether AI produces better outcomes. It is whether AI enlarges human agency, understood as the ability to generate options, choose among them, and pursue a chosen path.

The best use of AI is not to replace human judgment with machine judgment. It is to give more people the practical power to exercise judgment at all.

Development Is Really the Expansion of Options

Economic development is commonly measured through income, productivity, infrastructure, or access to services. These measures are important, but they describe outcomes more readily than freedom. A deeper definition of development is the expansion of a person’s real options: the ability to choose where to work, how to learn, when to seek medical care, whether to remain in a community, and how to respond when institutions fail.

Consider a rural patient who must travel six hours to see a specialist. An AI diagnostic tool connected to a local clinic could increase her options dramatically. It might help a nurse identify a dangerous condition, recommend an urgent referral, or distinguish a treatable infection from an illness that requires advanced care. The value is not merely that the software is accurate. The value is that the patient now has a viable path that did not previously exist.

The same logic applies to agriculture. A smallholder farmer often makes decisions with incomplete information: when to plant, how much fertilizer to use, whether a pest outbreak is approaching, and whether a changing weather pattern justifies taking on debt. A low cost advisory service using satellite data, local weather records, and crop information could turn one uncertain choice into several visible alternatives. Plant now, wait two weeks, switch crops, irrigate less, seek a cooperative loan, or insure part of the harvest.

This is a more useful way to think about AI in developing economies. The goal is not to import an abstract capability called intelligence. The goal is to create option generating infrastructure.

A school with an AI tutor is valuable not because the tutor can produce fluent explanations. It is valuable if a child who cannot afford private instruction can ask questions without embarrassment, receive feedback at her own pace, and discover subjects that were previously inaccessible. A public data system is valuable not because it makes government look modern. It is valuable if it helps officials identify neglected communities and gives citizens evidence with which to demand better services.

The distinction is crucial. AI can increase the capacity of institutions without increasing the agency of the people those institutions serve. These are not the same achievement.

The Infrastructure Paradox: AI Can Democratize Access, but Centralize Dependence

The first tension appears in infrastructure. Many countries lack reliable electricity, broadband access, computing resources, high quality data, and technical expertise. Yet these constraints can also create an opportunity to build systems that are more focused and locally appropriate than those found in richer countries.

A region does not necessarily need a large hospital in every district if it can combine community health workers, diagnostic support, mobile connectivity, and a reliable referral network. A school system does not need to wait for every village to have a specialist teacher if it can use offline learning tools that synchronize when a connection becomes available. A farmer does not need a sophisticated office to benefit from an agricultural model if advice can be delivered through a basic phone in a local language.

This is sometimes called leapfrogging, but the phrase can be misleading. Leaping over old infrastructure is not the same as eliminating dependence. A country that bypasses desktop computers by adopting mobile platforms may still become dependent on foreign operating systems, cloud providers, payment networks, and proprietary models. The visible technology becomes cheaper, while the invisible control moves elsewhere.

That is why infrastructure must be evaluated at two levels:

1. Operational infrastructure: electricity, connectivity, devices, data systems, and computing capacity.
2. Agency infrastructure: the ability to understand, audit, modify, contest, and replace the systems on which society depends.

The first enables AI to function. The second determines who governs it.

Imagine a government that adopts an automated system to decide which households qualify for food assistance. The system may process applications quickly and reduce administrative costs. But if citizens cannot learn which information affected their ranking, correct inaccurate records, or appeal a decision, the system has not simply improved administration. It has changed the relationship between citizen and state.

Efficiency has become a gatekeeper.

This is the danger of treating infrastructure as politically neutral. Roads, electrical grids, identity systems, and databases all distribute power. AI systems do the same, only more opaquely and at greater speed. A poorly designed database can make an entire population invisible. A biased model can convert historical exclusion into an apparently objective score. A platform can provide essential services while collecting information that people cannot meaningfully refuse to share.

The appropriate question is not only, “Can this system deliver the service?” It is also, “Can the people affected by this system remain capable of acting without it?”

The Agency Test: Does a System Expand, Substitute, or Collapse Choice?

A useful framework for judging AI projects is to classify them according to their effect on human agency. There are three categories.

1. Agency expanding systems

These systems increase the number or quality of options available to people while preserving their ability to decide. An agricultural adviser that presents several planting strategies, explains the assumptions behind each, and allows farmers to combine the advice with local knowledge is agency expanding.

A medical tool that highlights possible diagnoses, identifies missing information, and helps a clinician explain alternatives to a patient can also be agency expanding. It makes expertise more available without pretending that uncertainty has disappeared.

2. Agency substituting systems

These systems make decisions on behalf of people in circumstances where delegation may be useful, but where human oversight still matters. Automatic credit assessment, school placement, hiring filters, and medical triage often fall into this category.

Substitution is not automatically harmful. People delegate decisions every day. The problem arises when delegation is presented as necessity, when no meaningful appeal exists, or when the system’s objective is narrower than the person’s actual interests. A model may optimize repayment probability while ignoring whether a loan would allow a family to survive a temporary shock and become more secure over time.

3. Agency collapsing systems

These systems remove practical alternatives, conceal the rules by which decisions are made, or make individuals dependent on a single institution or platform. A digital identity system that is required for access to food, healthcare, or employment but cannot be corrected easily is an example. So is an educational platform that penalizes students for asking questions outside its prescribed curriculum while presenting its outputs as neutral guidance.

The difference between substitution and collapse is often not the algorithm itself. It is the surrounding institutional design. A decision can be automated and still remain contestable. A recommendation can be imperfect and still be useful. What turns technology into a threat is the disappearance of alternatives, explanations, and routes of appeal.

This yields a practical rule:

A system is humane when it makes action easier without making refusal impossible.

That rule should guide investment, regulation, and design. Before funding an AI project, decision makers should ask whether users can opt out, whether they can receive a reason, whether they can correct the data, and whether a human institution remains responsible for the outcome.

Local Knowledge Is Not Noise to Be Optimized Away

AI systems are often introduced into environments where formal data is scarce and informal knowledge is abundant. This creates a temptation to treat local knowledge as anecdotal, inefficient, or irrelevant because it does not fit neatly into a database.

That would be a serious mistake. In many communities, people possess finely tuned knowledge of soil, weather, disease, migration, family networks, and social trust. It may not be complete, and it may contain errors, but it often captures variables that distant systems cannot see.

A crop model might recommend planting based on regional rainfall patterns. A farmer may know that a particular field floods after a nearby road is repaired, or that a local seed variety survives a pest that does not appear in national records. A public health model may identify a high risk district. A community health worker may know that the apparent decline in clinic visits reflects a rumor about treatment rather than an improvement in health.

The most effective systems therefore treat AI as a translator between forms of knowledge, not as a machine that replaces one form with another. The model can reveal patterns. Local users can interpret context. Each corrects the blind spots of the other.

This principle also matters for language. A tool that technically supports a language but misunderstands its idioms, social roles, or culturally specific ways of expressing distress may be less useful than a simpler system designed with local participation. Inclusion is not achieved by adding a language option to a menu. It requires adapting the categories through which the system understands people.

Participation must therefore happen before deployment, not merely after complaints begin. Teachers, nurses, farmers, civil servants, and patients should help define what the system is for, what errors are unacceptable, and what kind of explanation is useful. They should have authority to reject a design that is technically impressive but socially unworkable.

This may appear slower than importing a ready made product. In practice, it is often faster than repairing a system that has lost public trust.

The New Development Agenda: Build Capability, Not Just Consumption

The usual AI investment agenda emphasizes access to tools. Give people chatbots, automate offices, deploy diagnostic software, and connect services to platforms. These steps can help, but access alone does not create agency. A population can consume advanced technology while remaining unable to shape it.

A more durable agenda has four layers.

First, build basic reliability. Electricity, connectivity, device access, secure data storage, and maintenance are not peripheral concerns. An AI service that works only in capital cities or fails during a power interruption is not a national solution.

Second, build human capability. This means training not only elite engineers, but also teachers who can evaluate educational tools, nurses who can question clinical recommendations, public servants who can inspect procurement decisions, and citizens who understand how automated systems affect them.

Third, build institutional contestability. Every high impact system needs clear responsibility, documentation, independent evaluation, data correction procedures, privacy safeguards, and accessible appeals. Trust should be earned through recourse, not requested through branding.

Fourth, build economic pathways around ownership. If AI creates value in a country but the data, models, infrastructure, and profits are controlled elsewhere, the country may gain services without gaining productive power. Local firms, universities, cooperatives, and public institutions need opportunities to develop, adapt, and govern the tools they use.

These layers reinforce one another. Skills without infrastructure produce frustration. Infrastructure without institutions produces dependence. Institutions without economic ownership produce rules that others implement. The objective is not technological self sufficiency in every component. No country needs to build every chip or train every model. The objective is sufficient capability to make informed choices about what to adopt, what to modify, and what to refuse.

Key Takeaways

1. Evaluate AI by the options it creates, not only by the tasks it automates. Ask whether people gain meaningful alternatives, better information, and greater capacity to act.

2. Separate operational infrastructure from agency infrastructure. Connectivity and computing matter, but so do transparency, appeal rights, local expertise, and the ability to replace a system.

3. Use AI to support judgment where context matters. In healthcare, agriculture, education, and governance, systems should surface possibilities and uncertainty rather than conceal both behind a single answer.

4. Protect the right to refuse and the right to correct. No essential service should depend on an unchallengeable automated decision.

5. Invest in local capability, not just imported access. Train the people who will interpret, govern, maintain, and improve these systems, and ensure that value creation is not permanently outsourced.

The most consequential AI question for developing countries may not be whether they can catch up with the technological frontier. It may be whether they can use the frontier to create a different relationship between expertise and ordinary people.

For centuries, scarce expertise has limited human possibility. A specialist was far away, a teacher was unavailable, a government office was opaque, and a market forecast arrived too late. AI can loosen those constraints. But it can also create a new scarcity: the scarcity of independent judgment in a world where recommendations surround every decision.

The choice is not between accepting AI and rejecting it. The choice is between systems that make people more capable and systems that make people more compliant. The first kind of technology gives users more ways to understand, decide, and act. The second gives institutions more ways to classify, predict, and control.

Development should be judged by which direction a society moves. The future will not be equitable merely because intelligent tools are widely available. It will be equitable when those tools help more people generate possibilities for themselves, choose among them, and pursue a life they had a genuine role in designing.