The Second Door

The essays preceding this one in the edition have argued that artificial intelligence raises a question — what human beings and institutions become as machines grow more capable — that applies wherever the technology reaches. It does not reach everywhere the same way. The consequences of generative AI for a management consultancy in Boston and for a call-center operator in Manila are not the same consequences, and a framework that treats “human flourishing in the age of artificial intelligence” as a single, undifferentiated condition will miss what is, for a large share of the world’s population, the more consequential question: not whether reasoning and communication are being automated, but whether the specific economic path many developing countries were counting on to close the gap with wealthier ones is still open.

The first door was already closing

For most of the twentieth century, industrialization was the standard route by which poorer countries converged with richer ones: build a manufacturing base, absorb agricultural labor into higher-productivity factory work, and grow. Rodrik (2016) documented that this route has been narrowing for several decades, independent of any AI-specific cause. Manufacturing’s share of employment and output now peaks earlier and at substantially lower income levels than it did for the economies that industrialized first — a pattern Rodrik terms premature deindustrialization, driven by a combination of globalization and labor-saving technological progress in manufacturing itself. The effect has fallen unevenly: economies in Asia with strong manufactured-export sectors have been comparatively insulated, while Latin American economies in particular have been hit hard. For much of the developing world, the traditional route to convergence was already less available than it had been for the countries that walked it first.

A second door opened

Baldwin (2019) and Baldwin and Forslid (2020) describe an alternative that digital connectivity has made newly available: services-led development through what Baldwin calls telemigration — workers in lower-wage countries performing remote work for employers in higher-wage ones, without physically relocating. Baldwin and Forslid argue explicitly that this route could extend the “emerging market miracle” geographically, resembling India’s services-led growth more than China’s manufacturing-led growth, and reaching economies that missed the earlier industrialization window Rodrik describes. The pattern is not speculative. Horton, Kerr, and Stanton’s (2017) analysis of remote-work contracts on a major online labor platform found the largest source countries for this kind of work to be the Philippines, India, and Bangladesh — economies for which the services door, unlike the manufacturing one, has been opening rather than closing.

The same key may close it

The technology that makes telemigration possible — digital connectivity paired with software capable of performing analytical, administrative, and communicative work remotely — is not separate from the technology now capable of performing a growing share of that same work directly. Baldwin and Forslid’s own account does not treat this as a settled matter in the services-led path’s favor: they note directly that “white-collar robots may displace some offshore humans,” meaning the same globotics transformation that opens the second door for a labor-cost arbitrage business model can, at the same time, erode the labor-cost advantage that model depends on. A telemigrant customer-service worker, translator, or data processor competes not only against workers in other countries but increasingly against a system that performs the same task at a lower marginal cost than any worker, anywhere, can offer.

This is no longer only a theoretical tension. Brynjolfsson, Chandar, and Chen (2025), using high-frequency payroll data covering millions of U.S. workers, found a 13 percent relative decline in employment for early-career workers in occupations most exposed to generative AI since its widespread adoption, concentrated specifically in roles where AI automates tasks rather than augments them — the pattern the services-led development path depends on not occurring at scale. The evidence for the developing economies this essay is actually about is thinner and less rigorous than that: the International Labour Organization has estimated that roughly one in four Philippine workers, some 12.7 million people, hold jobs in occupations that could be affected by generative AI, and industry reporting from the Philippine and Indian outsourcing sectors describes slowing entry-level hiring alongside continued aggregate growth — a mixed and still-unsettled picture, not a documented collapse. The clearest realized-effects evidence available concerns a wealthy labor market, not the developing ones whose access to the second door is actually in question. That gap in the evidence is itself worth stating plainly, since it means the strongest data currently available speaks to the risk this essay describes without yet confirming or resolving it for the population the essay is actually about.

This is not a claim that the second door is closing. It is a claim that whether it closes, stays open, or opens further is presently unresolved for the developing economies in question, that the clearest available evidence so far concerns a different population than the one at stake, and that the answer will not be determined by the technology’s trajectory alone.

What determines the outcome

The dynamics described above are not only economic. Mohamed, Png, and Isaac (2020) argue that AI development and deployment can reproduce a colonial structure of power even where no colonial intention is present: the technology is designed, trained, and governed overwhelmingly by institutions in wealthy economies, while its consequences — beneficial or displacing — are absorbed by workers and economies with comparatively little voice in the decisions that produce them. On this reading, telemigration is not simply a neutral opportunity that technology happened to open. It is a labor-cost arbitrage whose terms — which tasks are profitable to offshore, which are profitable to automate instead, and on what timeline — are set almost entirely outside the economies whose workers depend on it.

This sharpens what “durable capability” has to mean in Thesis 4 below. The Institute’s Capability-Conversion Problem, argued elsewhere in this edition, offers a relevant distinction here. A country whose workers perform outsourced tasks using AI tools without those tools building durable local capability — expertise, institutional capacity, the ability to move up the value chain rather than compete indefinitely on cost — occupies a fundamentally weaker position than one where the same technology use converts into lasting capacity. Read alongside Mohamed, Png, and Isaac, that capacity cannot be only technical or educational. Durable capability plausibly also requires some degree of control over the terms of the arbitrage itself — the capacity to shape AI governance, data, and labor standards, or to invest in AI development locally, rather than only absorbing decisions made elsewhere. Whether telemigration functions as a genuine development path or a temporary arbitrage that closes as the cost of automation falls is likely to depend less on which countries currently supply the most remote labor and more on which countries convert that period of participation into capability of this fuller kind — technical, institutional, and, in Mohamed, Png, and Isaac’s sense, a degree of agency over the arrangement itself.

Four theses

This essay’s claims are subject to challenge under the terms of the Institute’s Disputation Protocol. Editorial disclosure, including this publication’s AI review process, appears in the closing section of this edition.

References

Baldwin, R. (2019). The Globotics Upheaval: Globalization, Robotics, and the Future of Work. Weidenfeld & Nicolson.

Baldwin, R., & Forslid, R. (2020). Globotics and development: When manufacturing is jobless and services are tradeable. NBER Working Paper No. 26731. Also published in World Trade Review.

Brynjolfsson, E., Chandar, B., & Chen, R. (2025). Canaries in the coal mine? Six facts about the recent employment effects of artificial intelligence. Stanford Digital Economy Lab Working Paper.

Horton, J. J., Kerr, W. R., & Stanton, C. (2017). Digital labor markets and global talent flows. NBER Working Paper No. 23398.

Mohamed, S., Png, M. T., & Isaac, W. (2020). Decolonial AI: Decolonial theory as sociotechnical foresight in artificial intelligence. Philosophy & Technology, 33(4), 659–684.

Rodrik, D. (2016). Premature deindustrialization. Journal of Economic Growth, 21(1), 1–33.