The future of the labor force: higher cognition and more skills

Automation and digitization are reshaping occupations by both displacing and transforming tasks. Some traditional jobs (e.g., insurance underwriters, data entry keyers) face obsolescence, while new roles (e.g., blockchain engineers, digital forensics analysts) expand. Automation tends to augment human labor by reallocating routine tasks to machines and creating non-routine tasks leveraging human comparative advantages (e.g., interpersonal interaction, adaptability, problem solving). Conventional human capital measures (e.g., education) are criticized for assuming static skills or using wages as coarse proxies. A task-content approach better reflects skills. Prior work classifies tasks into routine, manual, and abstract, explaining polarization as computers substitute routine tasks and push middle-skill workers toward service roles. Building on skill clustering that separates social-cognitive and sensory-physical skill sets, this study investigates how U.S. workers’ skills are shifting amid technological change, and which skill attributes make jobs less susceptible to automation.

The labor effects of technology are multifaceted, spanning deskilling, upskilling, and reskilling. Decomposition of production tasks can simultaneously simplify (deskilling of routines) and elevate (upskilling of non-routines) job content. Prior clustering work identifies two skill sets: sensory-physical (e.g., physical interaction, sensory sensitivity, precision control) and social-cognitive (e.g., interpersonal interaction, problem solving, resource management, learning ability). The study posits differential automation effects: - H1: Jobs requiring higher sensory-physical skills are more likely to be replaced by machines, due to technical feasibility, trust in machines for rule-based tasks, and cost incentives as sensor costs fall. - H2: Jobs requiring higher social-cognitive skills are less likely to be replaced, due to Polanyi’s paradox (tacit knowledge), algorithm aversion (lower trust in machines for subjective tasks), and the need for empathy, communication, and interpretive judgment. - H3: Workers with a greater number of skills are less susceptible to replacement, since machines face high setup costs for task switching while versatile workers adapt to new tasks, aligning with reskilling imperatives highlighted by future-of-work forecasts.

Data sources: The study integrates O*NET and OEWS under the SOC system, harmonizing multiple SOC versions (SOC2000, SOC2010, SOC2018) via crosswalks to a unified SOC2010d for longitudinal analysis. O*NET provides annual measures of the importance and level of workplace skills, knowledge, and abilities for 923+ occupations; data from 2003–2022 are used (post transitional period). OEWS provides annual employment and wages for ~830 occupations. After mapping O*NET-SOC to SOC versions and crosswalking OEWS and O*NET to SOC2010d, the authors construct occupation-level panel data (2003–2022). Empirical model: Using a decade lag to capture long-run effects (following Autor and Dorn, 2013), the dependent variable is the Box–Cox transformed growth of employment share for occupation j over t−10 to t. The main specification is: ΔG_E,j = α0 + α1 P_{j,t−10} + α2 C_{j,t−10} + α3 N_{j,t−10} + α4 O_{j,t−10} + α5 W_{j,t−10} + ε_j, with occupation and year fixed effects and heteroskedasticity-robust standard errors. Key regressors: P (sensory-physical skill score), C (social-cognitive skill score), N (number/diversity of important skills), O (offshorability index), and W (change in the 10th percentile annual wage over the decade). Variable construction and measurement: - Dependent variable: Growth in employment share over a decade, Box–Cox transformed to stabilize variance and reduce outlier influence. λ is estimated via MLE. The text reports λ = −1.092 (Variable measurements) and λ = −1.029 (Table 4 note). The relationship between transformed and original scales is discussed in an appendix. - Skill scores: From O*NET importance ratings (1–5 scale). To adjust for ubiquitous/overrepresented skills, the revealed comparative advantage (RCA) for skill s in occupation j is computed: RCA(j,s) = [ onet(j,s) / Σ_{j′} onet(j′,s) ] / [ Σ_{s′} onet(j,s′) / Σ_{j′,s′} onet(j′,s′) ]. Based on Abdulvakhitova et al. (2018), skills are grouped into sensory-physical and social-cognitive sets. The study uses 18 selected skills/knowledge/abilities (excluding some generalized work activities used in constructing offshorability) to build Phy (sensory-physical) and Cog (social-cognitive) scores, then log-transforms these scores. - Number of skills (NS): Counts skills with importance > 1 for each occupation; occupations above the average count are coded 1, otherwise 0, as a proxy for worker employability/versatility. - Offshorability (OI): Following Firpo et al. (2011) and Autor and Dorn (2013), constructs an index from face-to-face interaction and on-site job requirements (reversing sign to measure offshorability), then log-transforms it. - Wages: W is the decade change in the 10th percentile annual wage. A robustness exercise also addresses missing 90th percentile wages via imputation by major occupation p90/p50 ratios. Controls and design: Occupation fixed effects and year fixed effects are included. Standard errors are heteroskedasticity-robust. Robustness tests vary periods (excluding GFC and COVID years), skill taxonomies (O*NET ability taxonomy), and wage imputation strategies.

Main regression (Table 4, dependent variable: 100× Box–Cox transformed annual growth in employment share, 2003–2022): - PhyLJ_-10 (sensory-physical skill score, lagged 10 years): −3.480, t = −2.54, significant (***). Interpretation: Higher sensory-physical intensity predicts lower growth in employment share over the subsequent decade. - NSJ_-10 (above-average number of important skills): 6.361, t = 4.36, significant (***). Interpretation: More diversified skill portfolios are associated with higher employment share growth. - OI_-10 (offshorability index): 0.509, t = 2.25, significant (**). Interpretation: More offshorable occupations tend to see higher growth in employment share under the specification, conditional on other controls. - ΔWageJ_-10 (change in 10th percentile wage): −3.205, t = −6.41, significant (***). - Constant: 26.676, t = 4.30, significant (***). - Observations: 6,234. Year FE: Yes. (Occupation FE noted; reported value 0.140; additional fit statistics in text/appendix.) Emerging occupations (2000–2023): Analysis of 150 new occupations shows most cluster in quadrants with higher social-cognitive scores; fewer in high sensory-physical/low cognitive regions; very few with low scores in both. Distribution by major groups (Table 5): - Computer, Engineering, & Science: 54 (36%) - Healthcare Practitioners and Technical: 38 (25.33%) - Management, Business, & Financial: 34 (22.67%) - Natural Resources, Construction, & Maintenance: 7 (4.67%) - Service: 7 (4.67%) - Production, Transportation, & Material Moving: 5 (3.33%) - Education, Legal, Community Service, & Arts: 3 (2%) - Sales & Office: 2 (1.33%) Robustness tests (Table 6): - Limiting to 2003–2007 and 2013–2017: Phy_-10 strongly negative (−9.708***), Cog_-10 positive (6.042***), NS_-10 positive (0.795***), ΔWage negative (−4.462***), OI positive (3.059***), consistent with hypotheses. - Using O*NET abilities classification for Phy/Cog: Phy_-10 negative (−3.755***), Cog_-10 positive (3.943***); NS not significant; ΔWage negative (−4.094***); OI not significant (3.048). - Handling missing 90th percentile wage via imputation: Phy_-10 negative (−3.310*), Cog_-10 positive (6.120***), NS_-10 positive (0.660**), ΔWage negative (−3.197**), OI positive (2.910*). Overall, results support: (H1) sensory-physical intensity predicts declines; (H2) social-cognitive intensity predicts gains; (H3) broader skill sets predict gains in employment share.

Findings align with the theoretical distinction between skills that are rule-based, transparent, and programmable (sensory-physical) versus those that are tacit, context-dependent, and interpretive (social-cognitive). Technical feasibility, declining sensor costs, and human trust in machine execution for objective tasks incentivize automation of sensory-physical functions. Conversely, Polanyi’s paradox and algorithm aversion limit acceptance and effectiveness of automation in social-cognitive domains requiring empathy, interpretive judgment, and nuanced communication (e.g., patient interactions, service and education). The evidence that diversified skill portfolios correlate with employment growth underscores the value of worker flexibility and adaptability—capabilities that complement machines and facilitate role transitions as tasks evolve. Emerging occupations concentrate in domains demanding social-cognitive strengths (e.g., computing, data, health, management), reinforcing the trajectory toward higher cognition and broader skill sets. These patterns indicate that technology predominantly augments human roles where cognitive and interpersonal capabilities are critical, while substituting tasks heavy in sensory-physical components.

The study provides the first empirical evidence that automation affects sensory-physical and social-cognitive skills differently, identifying traits of less automatable labor: higher social-cognitive skill intensity and broader, more diverse skill portfolios. Occupations emphasizing sensory-physical skills face greater replacement risks. Policy and practice implications include: promoting development of cognitive skills (creativity, critical thinking, problem solving), encouraging diversified skill accumulation across cognitive and physical domains, and supporting lifelong learning and reskilling pathways. Education systems should emphasize cognitive skills and interdisciplinary learning, and provide ongoing retraining to cultivate flexible workers who can adapt to evolving technological demands. Future research can extend to objective skill measurement, additional task-taxonomy refinements, and causal identification strategies linking specific technologies to occupational dynamics.

A key limitation is the reliance on O*NET measures derived from job incumbents, occupational experts, job data, and other sources, which introduces subjectivity. Developing more objective skill measurement approaches would strengthen future analyses.