The World Bank warns that AI could make poorer countries dependent on foreign platforms.
The warning sits alongside a useful concession: developing economies have fewer jobs immediately exposed to generative-AI automation than rich ones. The bank estimates that 4.5% of jobs in developing countries are vulnerable to automation, compared with 14.2% in rich countries. Many workers in poorer economies remain employed in agriculture, small enterprises, and occupations where current large language models cannot simply substitute for a person.
That is the true half of the argument. The trouble is that “less exposed to automation” is not the same thing as “safe from extraction.” A worker can keep the job and still lose bargaining power, income, and knowledge about the system governing the work.
The World Bank says relatively simple AI tools could help workers perform tasks now reserved for scarce specialists. A small business could translate documents or prepare accounts without hiring a specialist. A local administrator could use a narrow system to process forms even where electricity is unreliable and internet access is poor. The bank is not claiming that every developing economy needs a frontier model or a warehouse-sized data centre.
Fine. This is the sensible version of the AI argument, and it is much smaller than the one being sold in the capital markets. The tool performs a bounded task. A worker remains responsible for the result. The system is replaceable. The local institution retains the records and can inspect the process. That is automation in service of labour.
The other possibility is automation as a private toll road. The World Bank identifies the relevant fact: the most advanced AI systems are controlled by a small number of companies, mainly in the United States and China. A developing country that adopts those systems may receive useful capabilities while depending on a provider it cannot easily replace. The provider supplies the model and the interface; the local economy supplies the institutional setting, the data, the workers, and eventually the subscription revenue.
That is an analytical analogy to Cory Doctorow’s chokepoint capitalism, not a claim that the World Bank has proved deliberate extraction by AI companies. Doctorow and Rebecca Giblin use the term for an intermediary positioned between producers and the people who need their work, with enough buying or selling power to squeeze both sides. An AI provider can occupy an analogous position across borders: not because every provider is secretly operating a coordinated scheme, but because control over a difficult-to-replace technical layer creates leverage over whoever depends on it.
Interoperability means a customer can switch providers without rebuilding the entire system. In plain language, the door is not welded shut behind the first vendor.
The World Bank recommends buying models, cloud services, and other AI tools from several countries, while ensuring that the systems can work together and be swapped out. That is the correct remedy, although it arrives in the cautious vocabulary of a development report. It is not merely diversification. It is an attempt to preserve exit before the customer becomes captive.
The distinction matters because the bank also calculates that rich economies are likely to receive productivity benefits more than three times larger than developing economies if current adoption trends continue. This is not a mysterious failure of national character. It reflects unequal access to computing power, local data, reliable electricity, skilled workers, and capital.
The source does not establish that rich countries collectively own the cloud layer, most data-centre capacity, or much of the chip supply, and the column should not pretend otherwise. The narrower, documented point is sufficient: the advanced systems are concentrated in companies based mainly in the United States and China, while poorer countries generally lack the infrastructure to reproduce the frontier-AI buildout. They are being invited to improve productivity by renting access to capabilities whose underlying supply chain they do not control.
The physical dependence therefore needs to be stated precisely. A country that relies on a foreign platform may depend on foreign-owned data centres, computing infrastructure, and network services that keep that platform available. That is different from saying it relies on a foreign electricity grid. The electricity may be generated locally; the servers, ownership, maintenance, financing, and technical decisions can still be controlled elsewhere. Software may appear weightless. The infrastructure that runs it is not.
A model is not a universal mind. It is a statistical system trained on particular data, tuned toward particular objectives, and deployed through particular interfaces. If the training data omit local languages, administrative practices, medical conditions, or economic realities, the system may reproduce those omissions at scale. “Customizing AI to the local context” is therefore not a decorative preference. It is a requirement for the system to work.
Local data create a second dependency. The more a country adapts its schools, hospitals, tax offices, banks, and employers to one foreign platform, the more expensive it becomes to leave that platform. The application may be imported cheaply at first. Institutional dependence arrives later, after the records have been formatted, the workers trained, and procedures rewritten around the vendor’s assumptions.
To be fair, the World Bank does not recommend that developing economies imitate the United States and China by spending hundreds of billions of dollars on data centres. It correctly notes that useful systems can run on limited local computing power. A country does not need to build a frontier model to use a narrow tool well.
The trouble is that “narrow tool” must remain a policy condition, not a marketing phase. If a public hospital adopts a translation system, it should be able to export its records, audit the model, compare another provider, and continue operating when the vendor changes its prices. If a ministry adopts a decision-support system, the ministry must retain authority over the decision and citizens must retain a route to challenge it. If a farm uses software to diagnose equipment faults, the owner should not need the manufacturer’s permission to repair the machine afterward.
That requires open standards, local custody of data, and public technical capacity. It also requires procurement rules that treat replaceability as a safety feature rather than a technical nicety. The experts who warned that AI could double GDP growth while fueling underemployment and unrest were pointing at the same distributional problem: productivity is not a benefit until somebody has the power to keep it.
Doctorow’s term enshittification provides another useful analogy, not evidence about any particular AI provider. He uses it to describe a structural decay pattern in which a service first serves users, then business customers, then shareholders at the expense of everyone else. The concept matters here because it directs attention away from the character of an individual engineer and toward the constraints around a system. Competition, regulation, interoperability, and labour power determine how easily a provider can worsen terms without losing its customers or suppliers.
The same question applies to AI procurement. Can the customer move its data? Can another model be substituted? Are the system’s outputs auditable? Does a public institution understand the model’s limits well enough to reject its recommendation? Can workers challenge a decision made with its assistance? These are not abstract questions about technological sovereignty. They are the ordinary engineering conditions for avoiding a system that works only while one vendor permits it to work.
Developing countries should not be asked to choose between dependence on the United States and dependence on China. Nor should they be told that sovereignty requires reproducing the entire frontier-AI supply chain. Use multiple vendors, keep local custody of data, and make every provider replaceable before the contract becomes the infrastructure.
A country can borrow a tool without surrendering the workshop. The contract is where that distinction is decided.