Tech-driven skill race is raising inequality and social pressure

A new study suggests that the race to build smarter AI systems mirrors a deeper and more troubling economic dynamic: an escalating competition in which returns rise exponentially at the top, compelling both humans and firms to spend beyond socially optimal levels.

Titled Janus-Faced Technological Progress and the Arms Race in the Education of Humans and Chatbots, the working paper models how technological progress and lognormal wage distributions combine to create powerful incentives for overinvestment in both human education and AI development.

Exponential returns and the educational arms race

Under the hood, the model is a technology parameter that magnifies the productivity of skills. As this parameter increases, both mean income and inequality rise. Expected wages grow exponentially with the average level of skill and with the variance of the skill distribution. The model demonstrates that technological progress does not merely raise incomes; it amplifies the payoff to being at the top.

This amplification generates strong incentives for individuals to invest heavily in education. Unlike traditional human capital models that assume diminishing returns and interior solutions, the author’s framework shows that when income rises exponentially with skill, educational investment can exhibit corner solutions. Individuals are pushed to invest as much as possible to gain marginal advantages.

Under certain conditions, the ex-ante optimal level of educational investment can exceed median income. This means that individuals rationally choose high levels of schooling before knowing their exact place in the income distribution, yet more than half of them may later earn wages insufficient to cover those costs. The result is a form of inefficient competition: everyone feels compelled to run faster, but many end up worse off after the race.

The paper further examines how preferences over risk shape attitudes toward technological progress. Using constant relative risk aversion preferences, the author shows that technological growth always increases expected utility if individuals are relatively risk tolerant. However, once the coefficient of relative risk aversion exceeds one, there exists an optimal technological plateau. Beyond that level, further technological advances increase income risk and inequality enough to reduce expected utility.

In this framework, GDP can rise monotonically while welfare declines for sufficiently risk-averse individuals. Technological progress is therefore Janus-faced: it enhances productivity and average income while simultaneously intensifying inequality and risk.

The analysis also extends to sectors characterized by network effects and scale economies. In industries such as operating systems, search engines and social media platforms, small performance differences can yield large monopoly rents. the author introduces a parameter capturing the share of income driven by competitive advantage rather than pure productivity. As monopoly rents increase, individuals overinvest in education to capture these outsized rewards.

From a social planner’s perspective, this investment is inefficient. Individuals take the average skill level of the population as given and invest to outcompete others. Yet when everyone does so, aggregate productivity gains do not match the scale of investment. The model shows that as long as some share of income derives from competitive rent capture rather than productive output, private educational investment exceeds the socially optimal level.

Calibration: Rising GDP, rising risk

To test the model’s implications, the author calibrates it using U.S. household income data for 1975 and 2024, combined with IQ as a proxy for skill. The calibration reveals that the technology parameter amplifying skill has increased by more than 50 percent over the period. While both mean and median incomes have risen substantially, the baseline productivity parameter has fallen, indicating that earnings have become more dependent on skill and more concentrated in the upper tail.

The ratio of mean to median income has increased, reflecting growing inequality. The model interprets this as a shift toward a more meritocratic but riskier distribution. Returns to skill have become steeper, and earnings dispersion has widened.

The calibration suggests that individuals with a coefficient of relative risk aversion above approximately 2.5 would, ex-ante, prefer the income distribution of 1975 to that of 2024, despite the near doubling of per capita GDP. For these individuals, the increased inequality and income volatility outweigh the benefits of higher average earnings.

The paper also calculates the marginal value of skill by examining the expected income gain from one additional IQ point. Between 1975 and 2024, the annual income increase associated with a single IQ point has multiplied several times over, especially for individuals above the average skill level. Over a 40-year work life, the lifetime earnings effect becomes substantial.

This explosive growth in the value of marginal skill intensifies the pressure to invest in education. The model predicts that as long as current trends continue, the incentive to accumulate additional skill will keep rising. Extrapolating forward suggests even steeper rewards for high-skill individuals if technological amplification persists.

The model highlights broader social consequences. The rising opportunity cost of time devoted to non-market activities becomes particularly salient. If additional schooling yields disproportionately large lifetime income gains, time spent on family formation or childbearing carries increasing economic trade-offs. The analysis suggests that technologically driven increases in the return to skill may have dramatically raised the opportunity cost of having children, potentially contributing to declining fertility in advanced economies.

From human capital to AI arms race

Technology firms are currently allocating vast sums to train large-scale AI models and chatbots. The model interprets this as a parallel competition in which returns to AI capability also follow a lognormal pattern.

If AI performance yields exponentially increasing returns and if scale and network effects generate winner-takes-most markets, firms have strong incentives to invest aggressively. Planned capital expenditures by major technology companies run into hundreds of billions of dollars. Even firms with substantial cash flows are borrowing to finance AI development.

In this environment, investment decisions resemble corner solutions. Firms push spending to the limits of their financial capacity in an effort to move their AI systems into the extreme upper tail of the performance distribution. Yet ex-post, many firms may find that their investments fail to generate commensurate returns, mirroring the dynamic faced by students whose educational outlays exceed their realized earnings.

Network effects further intensify competition. If a slightly superior AI model captures a disproportionate share of users and data, the incentive to overinvest increases. The same mechanism that drives individuals to overextend in pursuit of educational advantages drives firms to escalate AI spending.