Artificial Intelligence Needs Intelligent Societies
Cheap AI, human capital, and the political economy of collective intelligence
The United States is pouring extraordinary amounts of capital into semiconductor fabrication, data centers, cloud infrastructure, electrical generation, and artificial intelligence research. The scale of this buildout reflects a familiar theory of technological power: the firms and countries that control the most advanced models will control a scarce and indispensable economic resource. If frontier intelligence remains expensive, proprietary, and concentrated, then ownership of computational capacity can generate durable rents.
Yet the economics of artificial intelligence may be moving in the opposite direction. Open-weight models continue to improve, engineering techniques diffuse quickly, and the cost of deploying capable systems keeps falling. DeepSeek, Moonshot AI, and other challengers matter not only because they intensify competition with American firms, but because they suggest that useful machine intelligence may become increasingly difficult to monopolize. The frontier will continue to advance, but most organizations do not need the single best model in the world. They need a sufficiently capable system that can be deployed reliably and cheaply.
If computational intelligence becomes abundant, the central source of value shifts. General-purpose technologies rarely transform production through ownership alone. Electricity became economically consequential when factories were redesigned around distributed power. The internet created value through new systems of logistics, communication, software, and organizational coordination. As one input becomes cheaper, the complementary capacities required to use it become more valuable. Cheap AI therefore does not make human capital obsolete. It increases the value of the people and institutions capable of directing, evaluating, and incorporating machine intelligence into collective activity.
political economy · public institutions · research systems
coordination · authority · memory · learning
judgment · expertise · interpretation
models · prediction · generation · retrieval
As computational intelligence becomes less scarce, the capacities above it become the limiting factors.
This creates a serious contradiction for the United States. During roughly the past forty years, American political economy has systematically weakened many of the institutions that produce human, organizational, and collective intelligence. Public universities have become more dependent on tuition and private revenue. Students increasingly bear the cost of education through debt. Federal research support has become more uncertain and more closely tied to commercial or strategic priorities. Corporate research laboratories devoted to long-horizon inquiry have declined, while firms increasingly acquire innovation through start-ups, mergers, contractors, and consultants.
The same pattern appears inside employment. Internal labor markets and stable career ladders have eroded. Firms invest less in long-term training because workers are treated as mobile inputs rather than repositories of developing expertise. Outsourcing converts organizational capabilities into contracts. Consultants replace internal knowledge. Contingent workers perform tasks without being integrated into systems of institutional learning. Managerial strategies aimed at eliminating slack and redundancy remove the spare capacity through which organizations experiment, recover from disruptions, and preserve knowledge across personnel changes.
These developments are normally treated as separate problems involving higher education, labor markets, corporate governance, innovation policy, or financialization. Taken together, they describe a broader process of social divestment from human capital. American capitalism has repeatedly attempted to substitute market coordination for institutional capacity: outsourcing for internal expertise, software for workers, consultants for organizations, acquisition for research, and financial discipline for industrial strategy. Artificial intelligence appears as the latest and most ambitious version of this substitution, promising that expertise itself can be detached from the institutions that cultivate it.
This promise confuses the generation of outputs with the organization of intelligence. Large language models can write, summarize, classify, retrieve, and propose. They do not independently decide which problems deserve attention, which standards should govern an answer, how authority should be allocated, or which tradeoffs a collective should accept. They do not preserve tacit knowledge accumulated through shared practice, create legitimate goals, or secure cooperation from people whose interests conflict. Those are organizational and political accomplishments.
| Institution-building model | Externalization model |
|---|---|
| Train workers and build careers | Hire contingently and purchase skills |
| Maintain internal research capacity | Acquire start-ups and license technology |
| Fund universities as public infrastructure | Shift costs to tuition and student debt |
| Preserve slack and organizational memory | Optimize headcount and outsource expertise |
| Develop collective capacity | Purchase computational outputs |
The distinction can be expressed as a hierarchy. Computational intelligence generates outputs. Human intelligence supplies judgment and interpretation. Organizational intelligence coordinates specialized knowledge through roles, routines, authority, and memory. Collective intelligence connects organizations through universities, professions, governments, research systems, and productive institutions. The current AI debate concentrates overwhelmingly on the first layer while assuming the others will either adapt automatically or become less important. In practice, cheap computation increases the pressure placed on every higher layer.
An organization with weak internal expertise may use AI to produce more reports, presentations, messages, and analyses without becoming better at deciding. Abundant generated material can intensify confusion when no one possesses the authority or judgment required to distinguish what matters. An institution that has lost experienced workers may be able to retrieve formal documentation while lacking the tacit understanding needed to apply it. A university system weakened by declining public investment cannot indefinitely supply the scientific knowledge and trained personnel upon which model development depends. AI can accelerate the use of accumulated knowledge, but it cannot independently reproduce the institutions that accumulate it.
This is why the substitution model of AI adoption is so destructive. As argued in The AI Substitution Trap, organizations often treat AI as a reason to remove workers before determining whether they have preserved the capacity to evaluate, correct, or replace the system. The resulting dependence is obscured because labor costs appear to fall while vendor dependence, error correction, and institutional fragility increase. The Token Budget Problem similarly showed why replacing payroll with compute does not eliminate costs; it changes their form and often hides the continuing need for judgment. The infrastructure risks described in The AI Stranded Asset Problem arise from the same mistake at a larger scale: assuming that more computation automatically creates a productive system capable of absorbing it.
The international competition over AI should therefore be understood less narrowly than a race between model developers. China’s potential advantage does not rest simply on whether a Chinese laboratory produces the highest benchmark score. It may lie in the relationship between inexpensive models and a larger productive system encompassing manufacturing, technical education, logistics, infrastructure, state coordination, and industrial policy. These institutions are neither uniformly effective nor free of political distortion, but they create channels through which machine intelligence can be embedded into production rather than sold primarily as a high-margin service.
The United States retains extraordinary universities, laboratories, firms, and technical workers. The problem is not an absence of intelligence but the political economy through which those capacities are maintained and coordinated. A strategy centered on data centers, semiconductor subsidies, proprietary models, and venture capital cannot by itself reverse decades of institutional erosion. Computational infrastructure is necessary, but it cannot substitute for scientific careers, public research, vocational education, stable organizations, or public agencies capable of planning and implementation.
Artificial intelligence initially appeared to confirm the dominant trajectory of American capitalism. Markets could replace planning, software could replace workers, consultants could replace internal expertise, and AI could replace the institutions that produce knowledge. The falling cost of computation reveals the limit of that vision. When capable models are widely available, the scarce factor is no longer access to generated intelligence. It is the capacity to organize intelligence toward durable purposes.
The societies that benefit most from AI will not simply own the largest models or construct the most data centers. They will sustain the human expertise, organizational memory, public research, educational systems, and coordinating institutions that allow computational abundance to become productive. The central challenge of the AI era is therefore inseparable from the reconstruction of human capital. Artificial intelligence needs intelligent societies because computation, however advanced, cannot supply the collective capacity that decades of social divestment have allowed to deteriorate.