Skill-Biased Technical Change

The basic fact — the college wage premium tripled

The most-studied fact in modern US labor economics is the rise in the college wage premium. In 1980, a college graduate earned on average roughly 30 percent more than a high-school graduate. By 1990 the premium was around 50 percent; by 2000 around 70 percent; by 2020 over 80 percent. The premium has stagnated or even slightly declined since around 2010, but the long-run picture is unambiguous: relative skilled wages roughly tripled over 40 years.

The same pattern shows up across advanced economies, though less dramatically: UK, Germany, France, and Canada all saw college premia rise, but typically by half the US magnitude. Latin America and East Asia saw similar rises during their education-expansion decades. The phenomenon is global; the US is just unusually pronounced.

Why did this happen? The leading explanation, dominant in labor economics since the early 1990s, is skill-biased technical change.

The Katz-Murphy framework

Lawrence Katz and Kevin Murphy's 1992 paper "Changes in Relative Wages 1963-1987: Supply and Demand Factors" laid out the canonical analytical machinery. The framework is austere: treat skilled labor (college graduates) and unskilled labor (everyone else) as two inputs in an aggregate production function with constant elasticity of substitution σ.

log(w_s / w_u)  =  (1/σ) · [ log(D_t) − log(L_s / L_u) ]

where w_s/w_u is the skilled-unskilled wage ratio, L_s/L_u is the skilled-unskilled employment ratio, and D_t is a relative-demand shifter. The log skill premium equals (1/σ) times the gap between log relative demand and log relative supply. Two forces, one outcome.

Katz and Murphy used reasonable estimates of σ around 1.4 and observed relative skill supply to back out the implied path of relative demand. They found that to fit the data, relative demand had to grow at roughly 3.3 percent per year — a stable, large, persistent skill bias in technical change. Trend supply growth had slowed (the baby-boomer surge of the 1970s subsided), and trend demand growth was high, so the premium rose.

A worked numerical example

Suppose σ = 1.4, the relative supply L_s/L_u grows at 2 percent per year, and the relative demand shifter D_t grows at 3.3 percent per year. Then over a decade:

Δ log(w_s/w_u)  =  (1/1.4) · [ 0.033 − 0.020 ] · 10
                 =  (1/1.4) · 0.13
                 ≈  0.093
i.e. about 9.3% rise in the wage ratio per decade

Over four decades, that compounds to roughly 37 percent. Add the late-1980s acceleration in technology, the slowing of relative supply in some periods, and the gender-specific patterns, and you get the observed rough tripling of the gap.

Now flip the calculation. If demand growth slows to 2.5 percent — as Beaudry-Green-Sand (2014) argue happened around 2000 — and supply growth continues at 2 percent, the wage premium grows at just (1/1.4) · 0.005 = 0.36 percent per year, which compounds to about 4 percent per decade. That is roughly the stagnant trend observed since 2010.

Routine-biased technical change — the polarisation refinement

The pure SBTC story is too simple. It predicts uniform skill upgrading — high-skill wages and employment grow, low-skill stagnate. But the data shows polarisation: high-skill jobs grow, middle-skill jobs decline, and low-skill in-person service jobs also grow.

Autor, Levy, and Murnane (2003) proposed the routine-biased technical change (RBTC) refinement. Their insight: technology — especially computers — substitutes for routine tasks, the kinds of explicit-rule-following work that can be codified into algorithms. Routine tasks are concentrated in middle-skill occupations: clerical work, accounting, simple machine operation, much of manufacturing. Non-routine cognitive tasks (analysis, management, creativity) and non-routine manual tasks (cleaning, food prep, personal care, security) are harder to automate. So computers hollow out the middle, leaving growth at both ends.

Autor and Dorn (2013) document the spatial dimension: cities with more routine-intensive employment in 1980 saw more polarisation by 2005, with displaced routine workers absorbed into local service occupations. The empirical pattern is dramatic — U-shaped employment and wage growth across the wage distribution, repeated across the US, the UK, Germany, and most of Western Europe.

Comparison — SBTC variants

Hypothesis	Mechanism	Predicts	Evidence (post-1980)
Classical SBTC (Katz-Murphy 1992)	Technology raises skilled-labor marginal product	Uniform skill upgrading	Right direction, wrong granularity
Capital-skill complementarity (Krusell et al. 2000)	IT capital complements skilled labor	Capital deepening raises premium	IT boom 1985-2007 fits well
Routine-biased TC (Autor-Levy-Murnane 2003)	Computers substitute routine tasks	Polarisation, hollowing of middle	Strong: 1980-2010 employment U-shape
Directed technical change (Acemoglu 2002)	Innovation is profit-directed toward larger markets	Skill bias endogenous to supply	Theoretically tight; identification hard
Trade and offshoring	Trade with low-skill-abundant countries hurts US unskilled	Premium rises in skill-scarce activities	Some, especially manufacturing-belt towns
Institutions (unions, min wage)	Decline of unions, falling real min wage	Wage compression unwound	Substantial in 1980s; less since

Goldin and Katz — the race between education and technology

Claudia Goldin and Lawrence Katz, in their 2008 book The Race Between Education and Technology, framed the entire 20th-century US wage history through the SBTC lens. Their central narrative: for most of the 20th century, education was winning the race. Public high schools spread from the 1900s through the 1940s, doubling the share of the workforce with a high-school education. After WWII the GI Bill and the rise of public universities expanded college access. Relative skill supply grew rapidly, often faster than relative demand. The skill premium fell from the 1910s through the 1970s.

Then, around 1980, the dynamic reversed. Educational attainment growth slowed as the easy gains had been made — high-school completion rates plateaued, college completion grew but at a slower pace. Meanwhile, technology, especially information technology, was accelerating its skill bias. Demand pulled ahead of supply. The skill premium rose for four decades.

Goldin and Katz's policy implication was direct: if you want to compress the skill premium, you need to revive education. Pumping more college graduates into the workforce is the supply-side lever; the demand-side bias is harder to reverse. The book is the canonical popular narrative of SBTC.

The AI inflection

The 2020s have introduced a possible reversal. Large language models — and generative AI more broadly — appear to target the cognitive non-routine tasks that have historically been a college-graduate domain: writing, analysis, software development, legal research, marketing. If those tasks become heavily automated, the cognitive-vs-manual skill premium may compress at the top, even as it remains high at the very bottom relative to the very top.

Brynjolfsson, Mitchell, and Rock (2018) developed a "suitability for machine learning" rubric and applied it to 2016 occupations; high-skill occupations scored higher than RBTC's routine baseline would suggest, anticipating the LLM era. Acemoglu and Restrepo (2018, 2022) develop a directed-technical-change framework in which AI can either complement or substitute skilled labor, with the bias determined by economic incentives. Early empirical work — Eloundou-Manning-Mishkin-Rock (2023) for OpenAI; Brynjolfsson-Li-Raymond (2023) for call centers — finds substantial AI exposure for cognitive tasks, with productivity gains concentrated in lower-skilled workers (because they have more to learn from the AI than the experts do). The aggregate wage implication is unsettled, but it could be that AI does to college-graduate work what computers did to clerical work in the 1990s.

Critiques and qualifications

• Identification. "Skill-biased technical change" is essentially a residual — whatever isn't supply growth must be demand growth. The economic content of the bias is interpreted, not measured directly. Card-DiNardo (2002) raise this critique forcefully.
• Premium for which skill? The college vs high-school distinction blurs many gradients. Top-decile vs bottom-decile, college vs no-college, college vs graduate — each tells a slightly different story. The "premium" varies by definition.
• Demand growth is not exogenous. Technology is invented and adopted in response to relative factor prices. Acemoglu's directed-technical-change framework makes the bias endogenous to supply, so the simple supply-demand race is incomplete.
• Institutions matter alongside technology. Real minimum wage declined from the 1960s to the 1980s, then partially recovered. Union density fell from 30 percent in the 1950s to under 10 percent today. Both correlate with the inequality patterns SBTC explains.
• Globalisation. Trade with low-skill-abundant countries can shift relative demand without any technological change. The China shock literature (Autor-Dorn-Hanson 2013) documents large local effects on manufacturing.
• Polarisation isn't pure SBTC. A simple two-skill SBTC framework misses the U-shape. RBTC adds tasks; Acemoglu-Autor (2011) generalise to a task-based framework.

Extensions — what the framework still teaches

• Task-based labor markets. Acemoglu-Autor (2011) and subsequent work model production as a continuum of tasks performed by workers or machines. The framework subsumes SBTC and RBTC as special cases and provides a vocabulary for thinking about AI-era reallocation.
• Spatial polarisation. Autor-Dorn-Hanson (2013) and Moretti (2012) extend the framework to local labor markets, where exposure to automation or trade shocks varies sharply by industrial composition.
• Migration and family formation. Cunha-Heckman (2007) and Chetty et al. (2014) document how skill premia interact with intergenerational mobility, family stability, and place. The labor-economics SBTC framework integrates with broader inequality research.
• Cross-country comparison. Acemoglu (2003) and Autor-Houseman (2010) compare countries with different institutional configurations to identify the institutional vs technological components of the premium rise.
• Education policy. If supply growth is the principal lever, what raises supply? Goldin-Katz, Chetty-Hendren-Katz-Lockwood (2018), Hoxby-Avery (2013) on college access and achievement.
• Forecasting. The Bureau of Labor Statistics produces a 10-year projection of educational attainment vs occupational structure for the US workforce; the gap drives implicit forecasts of the future skill premium.