Policymakers and economists have long debated why college-educated workers earn so much more than those without degrees — and whether that gap will keep growing. The question has taken on renewed urgency as artificial intelligence promises yet another wave of technological disruption. Will AI widen inequality further?

The standard explanation for the rising college premium centres on skill-biased technical change — the idea that successive waves of technology structurally favour educated workers (Katz and Murphy 1992, Krusell et al. 2000). As Acemoglu and Autor (2011) document, the wage gap between college and non-college workers roughly doubled between 1980 and 2010. More recently, Acemoglu and Restrepo (2022) have highlighted the role of automation in displacing non-college workers from tasks they previously performed. Acemoglu et al. (2023) have argued on VoxEU that the balance between automation and new task creation shifted around 1980, with digital technologies increasingly displacing workers without generating enough offsetting opportunities.

In a new paper (Hassan et al. 2026), we offer a complementary explanation. We argue that it is not just what new technologies do, but how fast they arrive that matters for wage inequality. Our account builds on a simple and intuitive idea, first articulated by Schultz (1975): educated workers are better at learning to use new, unfamiliar technologies. As technologies age and become standardised, this advantage fades, and all workers can use them. The implication is simple: even if successive technology waves are identical in their characteristics, a faster pace of technology creation shifts the economy toward newer, more skill-intensive technologies, raising the college premium.

A burst of technology creation in the 1980s and 1990s

To quantify this mechanism, we need data on when new technologies emerge and how they reshape labour markets as they age. Building on Bloom et al. (2021) and Kalyani et al. (2025), we identify new technical phrases — two-word combinations that appear frequently in US patents but are absent from historical English usage — as markers of technological innovation. We then link these phrases to distinct technologies documented in Wikipedia, giving us a catalogue of 6,259 technologies emerging between 1940 and 2006. Once we have the relevant phrases in hand, we track their use across patents and job postings. In patent filings, we trace each technology’s emergence year by identifying when its associated phrases first surge in use. In job postings, we track which jobs mention these phrases, allowing us to measure how each technology’s demand for skills evolves as it ages. This approach yields a detailed picture of technology creation and diffusion — from pulse generators in the 1950s to tablet computers in 2006.

Figure 1 Number of new technologies per year, 1940–2006

Notes: Time series for the pace of new technology creation, m(b). The figure plots the number of new technologies created each year. The dashed line shows the raw data, and the solid line is a smoothed version that isolates low-frequency variation.

The data reveal a dramatic acceleration. Before 1970, roughly 25–30 new technologies emerged each year. This pace rose sharply during the 1980s, peaking at around 250 per year, before declining to about 100 per year during 2000–2007. This acceleration was not exclusive to information technology — it spanned life sciences and other fields.

New technologies start skill-intensive, then become accessible to all

We then trace these technologies into the labour market using 300 million job postings from 2010 to 2023. A key empirical finding supports our mechanism: new technologies initially demand far more college-educated workers, but this skill intensity declines steadily as technologies age. Young technologies are mainly mentioned in job postings that require a college degree, with 57% of these postings demanding one. By the time a technology is 80 years old, that share falls to 34%. This pattern — observed consistently across ICT and non-ICT technologies, and across different time periods — provides direct evidence that skilled workers’ comparative advantage in new technologies fades as knowledge about their use spreads.

The model matches the rise and plateau of the college premium

We build a model incorporating these features and calibrate it using our text-based data. The model generates quantitative predictions for the college premium that closely match the US experience.

Figure 2 College premium, model versus data

Notes: Skill premium over time: Model versus data. The figure compares the college wage premium from the model (black line) with the data (red line). The model series shows the predicted path of the college premium in response to the rapid pace of technology creation m(b) between 1970 and 2007, holding all other factors constant. The data series is computed from the Current Population Survey and shows the log wage differential between college and non-college workers. Both series are measured in log points.

In response to the rapid pace of technology creation during 1980–1995, the model generates a 28 log-point (32%) increase in the college premium, aligning with both the timing and the magnitude observed in the data. Crucially, the model also captures the flattening after 2010, as technologies introduced decades earlier became standardised and accessible to less-skilled workers.

We decompose the total shift in demand for college-educated workers since 1970 into three components: changes in the supply of college graduates, changes in the pace of technology creation (our mechanism), and residual structural changes in production such as automation and capital deepening. The rapid pace of technology creation accounts for roughly one-third of the total increase in relative demand for college workers, with residual forms of skill-biased technical change explaining the remaining two-thirds.

Looking forward, if no further acceleration occurs, the model predicts the college premium will gradually revert — falling about halfway back to its pre-1980 level by 2080. Even temporary bursts of technology creation generate elevated inequality that persists for generations.

Why inequality rose more in cities

Our framework also sheds light on why the college premium is higher and rose faster in dense cities — a well-documented pattern (Autor 2019, Eckert et al. 2022). Eeckhout et al. (2021) have argued on VoxEU that IT adoption is strongest in big cities where wages are highest, driving urban polarisation. We propose a complementary channel: new technologies arrive in dense cities first and only later diffuse to less dense areas.

Figure 3 The skill premium across space, 1980 versus 2005–2007

Notes: The skill premium across space: comparing model and data. The figure plots the college wage premium by population density. The dashed lines plot data for the 1980 Census (in red) and the 2005–2007 American Community Survey (in blue). The model predicted levels for the BGP and 2006 are shown by the solid connected dots.

Our job-posting data confirm sizable diffusion lags: the modal technology in the densest US cities is 34 years old, compared with 48 years old in the least dense regions. When we extend the model to incorporate spatial diffusion, it accounts for the bulk of the cross-regional gap in the college premium and for most of the differential increase between high- and low-density areas over 1980–2005.

Why inequality rose first for young workers

A final piece of the puzzle concerns worker age. Card and Lemieux (2001) documented that the college premium rose first among young workers. Our framework offers a demand-side explanation: younger workers are better at learning new technologies, so an acceleration in technology creation raises their skill premium first. Only as technologies diffuse to older workers does the premium rise for them.

Calibrating the model using Current Population Survey (CPS) data on computer use by age, we account for about half of the observed age-specific differences in the growth of the college premium.

Implications for the AI era

Our findings may offer lessons for AI and inequality. If AI represents a new burst of technology creation, our framework predicts a renewed, potentially decades-long increase in the college premium. The magnitude would depend on how fast genuinely new AI-enabled technologies emerge and how quickly they become accessible to non-specialist workers.

However, our results also caution against extrapolating from early stages of a technology wave. The same forces that drive inequality up also bring it back down as technologies mature. The current plateau in the college premium may reflect exactly this: the technologies of the 1980s and 1990s have aged, and their skill demands have broadened. Whether AI will restart the cycle depends on the volume and speed of new technology creation it unleashes.

References

Acemoglu, D and D H Autor (2011), “Skills, tasks and technologies: Implications for employment and earnings”, Handbook of Labor Economics 4B: 1043–1171.

Acemoglu, D and P Restrepo (2022), “Tasks, automation, and the rise in U.S. wage inequality”, Econometrica 90(5): 1973–2016.

Acemoglu, D, D Autor and S Johnson (2023), “How AI can become pro-worker”, VoxEU.org, 4 October.

Autor, D H (2019), “Work of the past, work of the future”, AEA Papers and Proceedings 109: 1–32.

Bloom, N, T A Hassan, A Kalyani, J Lerner and A Tahoun (2021), “How disruptive technologies diffuse”, VoxEU.org, 10 August.

Card, D and T Lemieux (2001), “Can falling supply explain the rising return to college for younger men? A cohort-based analysis”, Quarterly Journal of Economics 116(2): 705–746.

Eckert, F, S Ganapati and C Walsh (2022), “Urban-biased growth: A macroeconomic analysis”, NBER Working Paper.

Eeckhout, J, C Hedtrich and R Pinheiro (2021), “Inequality is an urban affair, and it’s due to new tech”, VoxEU.org, 16 October.

Hassan, T, A Kalyani, and P Restrepo (2026), “The skill premium in times of rapid technological change”, Working Paper.

Kalyani, A, N Bloom, M Carvalho, T A Hassan, J Lerner and A Tahoun (2025), “The diffusion of new technologies”, NBER Working Paper.

Katz, L F and K M Murphy (1992), “Changes in relative wages, 1963–1987: Supply and demand factors”, Quarterly Journal of Economics 107(1): 35–78.

Krusell, P, L E Ohanian, J-V Ríos-Rull and G L Violante (2000), “Capital-skill complementarity and inequality: A macroeconomic analysis”, Econometrica 68(5): 1029–1053.

Schultz, T W (1975), “The value of the ability to deal with disequilibria”, Journal of Economic Literature 13(3): 827–846.