Kristina McElheran on The Effects of AI on Workers and Firms

Episode Summary

Executive Summary: Professor Christina McElhern argues that automation and AI affect workers unevenly, with outcomes shaped less by the technology itself than by firm strategy, management practices, and worker experience. She emphasizes complementarities, reskilling limits, firm inequality, and the need for experimentation before drawing sweeping policy conclusions about AI’s impact.

Main Topics: How the researcher came to study technology and work (Priority: 4/5): McElhern traces her interest to Silicon Valley during the dot-com era, where she saw both the promise and organizational challenges of internet-enabled automation, shaping a research agenda on digitization, strategy, and the future of work. Older workers, automation, and tacit knowledge (Priority: 5/5): Her research finds that major software/automation investments tend to benefit mid-career workers with firm-specific knowledge while older workers are relatively disadvantaged, often leaving firms sooner or moving to less favorable jobs. Younger workers and sorting, not direct harm (Priority: 4/5): The apparent early-career penalty mostly disappears once sorting into occupations, firms, and jobs is accounted for; younger workers are not inherently harmed as much as initial data suggest. Firm inequality and organizational context (Priority: 5/5): Technology’s impact depends heavily on the firm: decision rights, bargaining power, management practices, and productivity orientation determine whether AI boosts wages and productivity or has muted/misaligned effects. Current AI as part of a longer technology continuum (Priority: 5/5): McElhern argues that ChatGPT-style AI is not a clean discontinuity from prior analytics and machine learning; firms have used algorithmic tools for years, but integration into workflows will take time and be uneven. Policy, retraining, and experimentation (Priority: 5/5): She favors rethinking credentialing, retraining, and compensation, but warns against overconfident policy based on narratives or thin data. Her preferred approach is experimentation and evidence generation. Technology hype, science fiction, and realistic expectations (Priority: 4/5): The conversation warns against Skynet-style narratives and unrealistic productivity expectations; historical adoption has often been slow, messy, and organization-dependent rather than instantly transformative.

Key Arguments: Technological change is heterogeneous: averages hide major differences across firms, occupations, ages, and tasks. Mid-career workers with tacit, firm-specific knowledge often benefit most from automation complements, while older workers can be relatively disadvantaged. Younger workers’ apparent losses are often due to sorting into different jobs/firms rather than direct technological harm. The firm is a crucial unit of analysis; workplace design, management practices, and decision rights shape whether AI creates value. Productivity gains from predictive analytics or automation appear when technology matches the firm’s operating model (e.g., stable, efficiency-focused settings). AI should be viewed as an extension of a longer algorithmic/predictive analytics continuum, not a totally new category that instantly rewrites work. Reskilling is constrained by time, money, and motivation; policies should be more creative and user-informed rather than purely top-down. If the policy goal is to help displaced workers, experimentation is better than assumptions: test retraining, compensation, and commercialization models and measure outcomes. Socially, sidelining workers may preserve comfort but reduce leverage, participation, and societal value creation. Big technology shifts usually take longer to show up in productivity statistics than advocates expect, because organizations adapt linearly while technology advances quickly.

Data Points: Date of episode: Tuesday, March 28th, 2023 - Opening introduction by host Scott Walston Survey sample size: 800,000 firms - McElhern described a Census Bureau survey on AI-related technologies Adoption rate of AI-related technologies in 2017 survey: Single digits - Firms reported very low use of machine learning and related tools Coauthor timeline at Harvard Business School: 6 years - Bio given by host for Christina McElhern Research publication outlet: Journal of Econometrics - Paper on software automation and worker outcomes was published the prior year Time lag for factory electrification/productivity: Long, unspecified - McElhern referenced historical diffusion of steam/electricity as a slow process Telephone switching automation at AT&T: 80 years - Example cited to show long adoption lags before productivity effects appear

Pivotal Quotes: "the two great tastes that taste great together" — Christina McElhern: Used to explain complementarities between technology, workers, and firm context "I think we should think hard about how do we help people who need to reskill or can't reskill" — Christina McElhern: Discussion of older workers and the limits of retraining "when in doubt, go find out" — Christina McElhern: Her closing recommendation for policy experimentation and data collection

Implications: Listeners should expect AI’s effects to be uneven and organization-specific, not universally disruptive or instantly beneficial. Firms and policymakers should focus on complementary skills, experimentation, and realistic retraining rather than hype or fear.