A wave of rapid technology creation from the 1970s through the 1990s lifted the college wage premium by roughly 30% by privileging workers who learn new technologies faster; as technologies aged and standardized the effect faded, explaining the premium's flattening after 2010.

Main Finding

A faster pace of new-technology creation raises the skill (college) premium because college-educated workers learn to use newly invented technologies faster. Using a vintage-technology model calibrated with novel text-based measures of technology births and diffusion, the authors show that the acceleration in technology creation from roughly the 1970s through 2000 generated a sustained 28 log‑point (≈32%) rise in the US college wage premium between 1980 and 2010 (with flattening and partial reversal after 2010). Changes in the pace of technology creation account for about one third of the overall rise in college demand since 1980; the rest reflects residual structural changes (skill-biased technical change, capital deepening, automation).

Key Points

• Mechanism: Skilled workers (proxied by college education) have a comparative advantage in learning and using newly created technologies. That advantage fades as technologies age and become standardized and accessible to less-skilled workers.
• Dynamics:
  • If the arrival rate of new technologies (m(b)) increases, the economy temporarily shifts weight toward younger, more skill-intensive vintages, raising the skill premium.
  • The effect is transitory in theory (long-run premium depends on standardization and productivity life-cycles), but an extended period of rapid technology creation can produce long-lasting elevation in the premium.
• Quantitative magnitudes:
  • Model generates ~28 log‑point (≈32%) increase in college premium, 1980–2010.
  • Total demand for college workers rose ~100 log points since 1980; ~1/3 attributable to faster technology creation.
  • Geography: the mechanism explains 6.2 of an 8.7 log‑point larger increase in the premium in high‑density vs low‑density areas (1980–2005).
  • Age: accounting for younger workers’ faster uptake explains about half of observed age gaps in premium increases.
• Empirical patterns used to discipline the model:
  • New technologies are initially more college‑intensive and become less so as they age.
  • The pace of new technology creation rose starting in the 1970s, accelerated in the 1980s, and tapered in the 2000s.
  • Diffusion gaps across places: modal technology age ≈34 years in the top 1% densest locations (e.g., NYC, SF) versus ≈48 years in bottom 50% density locations.

Data & Methods

• Data sources:
  • Novel text-based measures from Kalyani et al. (2025): identify distinct technologies using Wikipedia technology pages, link to patents via technical bigrams, and trace diffusion in job postings.
  • Labor-market evidence: Current Population Survey (CPS) for the college premium and age profiles; CPS computer-usage data by age to discipline age-diffusion parameters.
  • Regional diffusion estimated from technology ages observed in job postings across locations of varying population density.
• Model:
  • Multi‑vintage framework: technologies born over time, each with an age profile for productivity z(u) and skill intensity α(u) (α high for new tech, declines with age).
  • Two worker types (high-skill = college, low-skill = non-college); skilled learn new vintages faster (higher α for small u).
  • Aggregate output is a CES of vintage outputs; equilibrium determines wages and quantities given the technology-age distribution m_t(u).
  • Analytical results:
    • Balanced growth path (BGP) when technology arrival rate is constant: stable skill premium independent of level of m.
    • Shock/increase in arrival rate raises skill premium transiently by shifting mass to newer vintages.
• Calibration and quantification:
  • Calibrate α(u), z(u), and m(b) using text-based measures of technology birth rates and measured age profiles of skill demand.
  • Assume economy on BGP in 1970 and simulate the observed acceleration in technology arrival (1970–2000) to generate the college premium time series.
  • Decomposition: total change in college demand split into supply changes, change in pace of tech creation (authors’ mechanism), and residual structural changes.
• Key identification assumptions/limitations:
  • College education used as proxy for the ability to learn new technologies.
  • Technology waves assumed identical in intrinsic characteristics (α(u) and z(u) functions are age-dependent but time-invariant across cohorts) — isolates pace rather than structural bias of successive waves.
  • Arrival rate m(b) treated as exogenous.
  • Measurement via text matching (Wikipedia–patent–job postings) can misclassify or miss some technologies; analysis focused on 1976–2007 technology tracing window.

Implications for AI Economics

• Pace matters: If AI accelerates the creation of new technologies (many AI-enabled tools, applications, and sub‑technologies), it will tend to raise the premium for workers who can rapidly learn and apply these new tools. That effect can be large and persistent even if each individual technology eventually becomes standardized.
• Standardization and learning-costs are crucial: The net long-run effect depends on how quickly AI or AI-enabled technologies become standardized and easy to use by non-experts. If AI lowers the learning costs for non-college workers (reducing α(u) differences or shortening the window where α is high), the premium rise will be smaller or shorter-lived.
• Urban concentration and diffusion lags: Slow spatial diffusion of new technologies implies faster, larger wage gains in dense cities where new vintages arrive earlier. Expect AI-driven inequality to be geographically concentrated unless diffusion and adoption outside top tech hubs are accelerated.
• Age and cohort effects: Younger workers (and those with greater capacity to learn/adapt) will benefit earlier from AI-driven waves. Policies aimed at reskilling older workers and accelerating their adoption could reduce cohort gaps.
• Policy levers:
  • Speed up diffusion and lower effective learning costs (training, user-friendly interfaces, accessible tooling, public digital infrastructure) to shrink the window of skill‑biased adoption.
  • Support for regions lagging in adoption (subsidies, digital infrastructure, remote training) to limit urban divergence.
  • Monitor not only which technologies are being created but also their diffusion and standardization rates—text-based patent/job‑post indicators are a useful early-warning toolkit.
• Complementarity with other mechanisms: The paper finds that faster tech creation explains about one third of the historical rise in the college premium; residual structural factors (automation, capital deepening, inherent skill bias of successive waves) remain important. For AI, both the pace of creation and the intrinsic nature of AI (automation vs augmentation, capital intensity) will jointly determine outcomes.

Short takeaway: beyond debates about whether AI is inherently skill‑biased, this paper highlights that the speed at which new AI-enabled technologies are produced and the pace at which they are standardized and diffused are central — accelerating creation raises the premium for fast learners (often college-educated and urban), while faster standardization/diffusion and tools that lower adoption costs by non-experts can blunt that effect.