Stanford Graduate School of BusinessWorld Development Report 2026: The Promise of AI - AI as a General-Purpose Technology
Stanford Graduate School of Business
Why the Developing World Is Optimistic 0:00
The speaker opens by noting a puzzle: rich countries, especially the United States, tend to be pessimistic about AI, while poorer countries are far more optimistic. The explanation lies in the starting point. When people struggle to feed their children, sell their crops, or get basic medical and teaching information, new technology that fills those gaps looks like a clear gain rather than a threat. At the same time, developing countries are far from their fully resourced potential, so the same investment can produce a much larger percentage improvement than it would in a wealthy country.
Facts That Should Anchor the Conversation 4:31
Before applying any framework, the report insists on grounding it in real facts about developing economies. These countries have far fewer nurses and teachers per person, so automation there fills a scarcity rather than replacing surplus workers. About 90 percent of people work in firms with fewer than 10 employees, shifting the focus toward building capability rather than fears of mass layoffs. Many of these countries also import their key technology inputs, meaning profits can flow out through foreign firms rather than staying in the local economy, which complicates the usual productivity story.
What Makes AI a General-Purpose Technology 8:01
A general-purpose technology is not just big and widely used, it triggers reinforcing changes across sectors as industries adjust to new input prices and copy each other's organizational practices, a process that unfolds slowly and under uncertainty. Some firms hold their position for non-economic reasons, such as political protection or entry barriers, so markets alone cannot explain the pace of change. AI also spawns other general-purpose technologies, like affordable software coding tools, which let developing countries build low-cost applications suited to their own resource-limited settings.
Boosting General-Purpose Capabilities 13:31
Most workers in developing economies run small shops, restaurants, or farms with little expert information or software support, so added expertise acts as a complement to their existing work rather than a replacement. Small productivity gains, like making farmers five percent more productive or cutting travel time to get correct medical treatment, translate directly into real output. Improving sectors like healthcare, education, and agricultural extension lifts everyone's productivity. The scale of the gap is stark: sub-Saharan Africa has 2.9 physicians per 10,000 people against a global average of 16, and low-income countries average 40 primary students per teacher.
Meeting people where they are 17:31
The report looks at how people will actually consume AI, noting that only about 20 percent of people in low-income countries own smartphones. That does not rule out AI reaching them, since tools can be built into messaging apps shopkeepers already use, or delivered as two-way voice messages over basic feature phones, as already tested in agricultural advice projects. The key idea is that adaptation enables adoption: AI only helps when it is reshaped to fit low-resource settings, and developing countries can take part in that adaptation themselves.
Readiness, bottlenecks, and outsourcing risk 20:30
Surveys show 90 percent of small businesses already use messaging apps, and GitHub activity reveals real AI development happening in India, Pakistan, Bangladesh, and Nigeria, not just rich countries. Yet outsourcing hubs are vulnerable, since a multinational could shift all its business process work from the Philippines to Bangalore almost overnight, then automate much of it. This raises the question of how freed-up human capital gets redirected productively, which depends on bottlenecks, the scarce pieces that limit how far AI's gains can spread.
Modeling AI's macro impact 25:00
Chris explains research with Chad Jones aiming to ground predictions of AI-driven growth in historical patterns of automation, from plows to tractors to GPS-guided equipment. Their model centers on weak links, the idea that output depends on whichever task or input is scarcest, much like a cake needs eggs, sugar, and flour together, or a chain is only as strong as its weakest link.
Why scarcity, not abundance, sets limits 29:31
Using a harmonic mean formula, Chris shows that once one input is abundant, adding more of it barely raises total output, because the scarce input still constrains everything. Since software is only about 2 percent of U.S. spending, making it infinitely productive would raise GDP by roughly 2 percent, and even automating half the economy's tasks at infinite productivity would lift output by only about 19 percent. Growth therefore hinges on how strong these weak links are and how quickly automation spreads beyond the tasks already being automated.
The measurement framework explained 33:00
The speaker lays out a model where output comes from combining many tasks, and each task can be done by either machines or labor, whichever is cheaper. Over time machine productivity has risen much faster than human productivity, which is why more tasks shift to automation. Two key numbers matter: how binding the weak links are, estimated at a very low substitutability of about 0.2, meaning growth is strongly bottlenecked by the worst-performing tasks, and the automation rate, estimated at roughly 2 percent of tasks shifting from humans to machines each year. The core lesson from history is that machines improving quickly is not enough on its own; productivity gains also depend on letting machines take on an ever-expanding range of tasks, since freezing machine use at 1950s tasks would have wiped out almost all the growth seen since.
Projecting growth scenarios forward 39:31
Looking ahead, three possible futures are considered: machines eventually doing everything, humans always retaining some tasks, or a middle path where human involvement keeps shrinking without disappearing. Under a business as usual scenario, where machine productivity grows about four points faster per year than human productivity, growth only edges up to around 2.6 percent by 2070, putting output just 4 percent above trend by 2050. Under a more extreme Moore's Law everywhere scenario, with machine productivity growing 11 points faster per year, growth rates could reach 7 percent by 2030, with some versions of the model even producing runaway, near-infinite output. The historical takeaway offered is that the US grew by shifting tasks from humans to faster-improving machines, and future growth will depend on how fast and how broadly that shift continues, with widespread moderate-quality AI mattering more for aggregate growth than narrow, extreme improvements. A commentator then raises how weak-link thinking, echoing Michael Kremer's O-ring theory, might apply to developing countries, where weak links could include education, health, and small business capacity, prompting the response that technology adoption barriers like institutions and property rights will shape AI diffusion much as they have shaped older technologies.
Calibrating Models For Developing Regions 48:30
The discussion turns to how the model could be adjusted for developing-world scenarios where different bottlenecks, or weak links, dominate. One speaker notes that even in the United States, many firms are slow to adopt transformative AI technology, and that educated middle managers play a crucial role in helping less-skilled workers use new tools effectively. This management layer is described as a key ingredient for getting value out of innovations that are developed elsewhere rather than locally.
The Silicon Valley Rebuttal 50:00
A recurring pushback from tech industry conversations is the claim that if AI becomes valuable enough, any friction or bottleneck will simply be solved, much like finding a way to bake a cake without eggs if the reward is high enough. In response, the speaker points to real organizational barriers to adoption that are not about prices, including government support and regulatory protection, using the example of a bank that could lobby rather than simply become more efficient. The historical record is cited as a caution: high returns to innovation existed for decades in U.S. history without producing rapid growth, so the real question is whether AI changes the process of innovation and discovery itself in a new way. Physical-world limits, such as how close robots are to replacing humans, and worker pushback from people fearing job loss, are named as further forces slowing adoption.
Why Waiting Can Be Rational 53:30
A business facing a major reorganization around AI might rationally choose to wait, since much of the investment could be matched or surpassed by a competitor within six months. The same logic applies to institutions like a university deciding whether to overhaul staffing around AI. Falling costs reinforce this point: a frontier model that was expensive a year ago can become dirt cheap within months, so a developing country priced out of frontier inference today may find it affordable soon. This creates a real tradeoff between adopting an expensive frontier model or a cheaper, well-tested one. An example from Africa showed smaller models run on gaming computers successfully handling narrow, local tasks like speaking a dialect, without needing a large, slow frontier model.
Weak Links Beyond The Model Itself 56:01
Even a correct AI-driven answer, such as the right medical treatment, is worthless if the resources to deliver that treatment, like drugs reaching a rural village, are missing. This illustrates a weak link outside the model's own accuracy.
Labor Shares Versus Living Standards 57:00
On labor markets, machines doing more tasks over time does not automatically shrink the human share of labor income, a result considered natural across many economic models. Even if labor's share did fall, people experience income levels, not shares, so a smaller share of a much larger pie can still mean higher living standards for everyone.
Weak Links Cut Both Ways 58:00
A forthcoming paper examines how the same weak-link logic that limits growth can also limit safety. On the downside, a small disruption to a narrow set of tasks can cause outsized losses, as in cybersecurity, where it is unclear whether attackers or defenders gain the advantage, especially during a transition period when not everyone has access to frontier defensive tools. Biorisk is named as another serious concern, since AI could help bad actors learn to produce harmful things, lending legitimacy to concerns about pacing frontier AI development that go beyond profit-driven regulatory capture. Open-source models, while useful for lowering costs in developing countries, are seen as leaving biosafety and cyberattack risks open.
Cyberattacks Move At Attacker Speed 59:30
Applying the same micro-economic framework to cybersecurity shows why it deserves urgent attention. Governments care about pace of change, but attackers do not face the same adoption costs or trust-and-safety review that legitimate businesses do, and they only need one attempt among many to succeed. If that attack hits a critical shared component, the damage can be outsized. A forthcoming Journal of Economic Perspectives paper highlights that many firms rely on the same software products or widely used services, such as financial systems, gas pipelines, or prescription software, creating concentrated weak links; one cyberattack caused a third of the United States to lose digital prescription access for a month. Surveys show people in developed countries are worried and distrust institutions to fix this, a distrust the speaker argues is rational given repeated failures to solve the problem.
Getting The Right Risks Right 1:03:00
The concern is that downplaying job-loss fears as a pervasive threat, rather than a sector-specific issue, risks distracting from the cybersecurity and biosecurity dangers that economic analysis shows are genuinely present in the short run. The hope expressed is that readers take away a balanced view: real opportunities exist and should be pursued, real risks exist, and the priority is identifying which risks actually matter most.
Closing Remarks and Coffee Break 1:04:01
The session closes with a short thank you to the research team, noting that solid research, data, facts, and theory help direct resources toward the most productive uses. The event then breaks for coffee.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
