The Dependency Question That Development AI Can't Escape

In 1970, the economist Herbert Myint wrote that technology transfer from developed to developing economies tended to produce what he called “export enclaves” — pockets of modern productivity embedded in traditional economies, generating output that flowed to the developed world and generating local employment without transforming the broader economic structure. The metaphor was for physical infrastructure: mines, plantations, manufacturing facilities owned by foreign companies. The output left. The benefits to the host economy were real but bounded.

The enclave metaphor has been applied, with varying degrees of precision, to successive waves of technology transfer: the green revolution’s seed and fertilizer systems, the IMF’s structural adjustment programs, the mobile banking infrastructure that runs on networks owned by Vodafone and Airtel. It is being applied now, by a small number of development economists and an even smaller number of AI policy researchers, to AI. The question they are asking is: who controls the infrastructure, who captures the value, and what does the host country get besides the use of someone else’s tools?

The strongest version of the dependency argument applied to AI goes like this. The AI tools deployed in developing countries — whether by American companies (Google, Microsoft, Meta) or Chinese ones (Alibaba, Tencent, Huawei) — are built on infrastructure that the deploying company controls: the model weights, the training data, the hardware, the software stack, and the terms of service that govern access. The host country gets access to the tool’s outputs. It does not get access to the underlying capability. If the tool is discontinued, or if the terms change, or if the geopolitical relationship between the host country and the technology supplier changes, the access goes away.

This is a real concern. The history of platform dependency is populated with communities and businesses that built significant value on platforms that then changed the terms — app store fee increases, API restrictions, algorithm changes that eliminated organic traffic. These are not catastrophic for individual developers who can adapt. At the country scale, for critical services like health, finance, and government administration, dependency on a single external provider is a structural vulnerability that is worth taking seriously.

The weaker versions of the dependency argument are less compelling. The version that says “no AI at all is better than AI you don’t control” is contradicted by the experience of mobile banking: M-Pesa gave Kenyans access to financial services they did not have before, even though Safaricom is partly owned by Vodafone. The counterfactual that matters is not “AI you own” versus “AI someone else owns” but “AI someone else owns” versus “no AI, or AI that works much more poorly.” In the near term, for most developing countries, those are the actual choices.

The middle ground — which is where the most serious development economists actually operate — acknowledges both the real value of AI access and the real costs of unmanaged dependency, and looks for governance frameworks that can capture more of the former while reducing the latter.

The most practical version of this framework involves several components. First, investing in the capability to understand and evaluate AI tools — the “AI literacy” infrastructure of technical staff, regulatory capacity, and procurement expertise that allows a government or institution to make informed choices about which AI tools to adopt, on what terms. Second, negotiating terms of service that include data portability, transparency requirements, and transition provisions that reduce lock-in risk. Third, supporting the development of domestic AI capabilities that can, over time, provide alternatives for the most critical applications.

None of these is rocket science. They are the same toolkit that development economists have recommended for technology transfer for decades. The specific application to AI requires updating the toolkit for the ways that AI dependency is different from, say, telecommunications infrastructure dependency.

The ways AI dependency is different are worth being specific about. AI tools that improve with use create a feedback dynamic that is different from static technology transfer. When you use a medical AI tool, your usage data can (depending on the terms) improve the model for future users. If the data and the model improvements accrue to the technology supplier rather than the host country, the host country is not just using the tool — it is providing training data that makes the tool better, at no compensation, for users everywhere including in the supplier’s home market.

This is not hypothetical. Google’s medical AI programs in East and Southeast Africa have involved partnerships with local hospitals in which patient data was used to train and evaluate AI models. The terms of those partnerships, and the degree to which the resulting model improvements are accessible to the partner hospitals versus retained by Google, have been sources of tension that have been resolved in some cases and not in others. The broader principle — that AI deployment creates data flows that can be structured to benefit the host country or to benefit the supplier — is one that few host countries have the negotiating sophistication to address effectively.

The intellectual property dimension is related. If a domestic research institution builds a model specifically designed for a local disease burden or a local language or a local agricultural system, using local data and local expertise, the question of who owns that model and can commercialize it globally is one that intellectual property frameworks have not resolved cleanly. Development institutions, like the Wellcome Trust and the Bill & Melinda Gates Foundation, have experimented with IP frameworks for global health research that attempt to preserve low-income country access to research outputs. Adapting those frameworks for AI requires addressing the specific features of AI IP — model weights, training data, fine-tuning processes — that are not clearly analogous to pharmaceutical patents.

The counterargument to dependency concerns is the counterfactual argument applied to infrastructure rather than just to access. Countries that have the most beneficial relationships with AI technology are not necessarily the ones that built the most independence — they are the ones that built the most effective local capacity to use, adapt, and build on available tools. South Korea’s technology industry grew by mastering and then improving Japanese technology before developing independent capabilities. Taiwan’s semiconductor industry grew by manufacturing chips designed by Americans before developing its own design capabilities. The path to independence runs through intelligent use of dependency, not through premature rejection of available tools.

This argument is correct in its broad strokes and requires careful qualification in the AI context. The manufacturing economy cases (Korea, Taiwan) involved building physical capabilities — factories, engineers, process knowledge — that were inherently local and could not be transferred back to the original supplier. AI capabilities are different: a model trained in Nigeria on Nigerian data is code that can be hosted anywhere. The barrier to foreign capture of locally developed AI capabilities is governance, not physics.

The dependency question is not going to be resolved by waiting for perfect conditions. AI tools are being deployed in developing countries now, by both American and Chinese companies, at a pace that reflects real demand from real users who benefit from the services. The question is not whether to engage with this deployment but how to engage in ways that capture more of the long-term value while managing the risks.

That requires governments with the technical capacity to negotiate intelligently, development institutions willing to fund the governance infrastructure that makes intelligent negotiation possible, and AI companies — on both the American and Chinese sides — who accept that responsible deployment in developing markets includes terms that benefit the host country rather than just extracting value from it.

The enclave metaphor is not destiny. But avoiding it requires conscious choices by all the parties involved, and those choices are not being made with sufficient intentionality anywhere in the current ecosystem.