How AI reshapes work through task reallocation
Menaka Hampole, Dimitris Papanikolaou, and Bryan Seegmiller, in this paper, developed a comprehensive approach to measure firm-level AI adoption and its impact on workers using novel natural language processing techniques applied to resume and job posting data from 2010-2023.
Methodology
- AI adoption measurement: The AI-implementing employees were identified by analyzing 58 million LinkedIn résumés from Revelio Labs. Using large language models, the authors extracted over 1 million distinct AI use-cases, creating a detailed picture of how different firms apply AI technologies. This approach captures internal AI development that traditional surveys might miss.
- Task-level exposure analysis: The authors used advanced text embeddings (GTE-Large model) to measure semantic similarity between AI applications and worker tasks from O*NET occupational data. Tasks were classified as “AI-exposed” if their similarity with AI use-cases exceeded the 95th percentile threshold, allowing for precise measurement of which job functions face potential automation.
- Comprehensive impact assessment: The authors examined multiple outcome measures including firm performance metrics (sales, profits, productivity, employment), five-year employment growth patterns across occupations, and task-level skill demand changes using job posting data spanning over 1 billion records.
Key findings
- Strong firm-level benefits: AI adoption significantly boosts firm performance, with one standard deviation increase in AI use leading to approximately 9.6 percentage points higher revenue growth, 7.4 percentage points higher total factor productivity, and 6-7 percentage points higher employment growth over five years.
- Task-level substitution with offsetting effects: While AI substitutes for specific tasks (reducing skill demand by ~4.8% for highly exposed tasks), the employment impact depends critically on two factors: mean exposure across an occupation’s tasks (which reduces labor demand) and concentration of exposure in specific tasks (which enables workers to reallocate effort to non-automated tasks, partially offsetting job losses).
- Modest net employment effects: Despite clear task-level substitution, aggregate employment impacts remain limited due to counterbalancing forces. High-exposure occupations experience about 14-15% employment share reduction from direct substitution, but this is substantially offset by productivity gains at AI-adopting firms and within-occupation task reallocation.
- Differential impact across worker groups: AI exposure is highest among high-wage, white-collar occupations (business, finance, engineering), but these groups experience only modest net employment declines (~3%) once offsetting effects are included. Some sectors like legal occupations even see slight gains due to firm productivity benefits.
The research reveals that AI acts as a powerful substitute at the individual task level, but reallocation mechanisms within jobs and productivity-driven expansion at adopting firms substantially moderate the aggregate employment impact.
Artificial Intelligence and the Labor Market
Authors: Menaka Hampole, Dimitris Papanikolaou, Bryan Seegmiller
From: Yale School of Management, Kellogg School of Management, MIT Sloan School of Management
Guiding innovation for inclusive growth
The paper by Anton Korinek and Joseph Stiglitz develops a framework to analyze how society can steer technological innovation—particularly in the age of automation and AI—to increase labor demand and promote equitable economic outcomes, rather than allowing market forces to determine the direction of technology. The model includes heterogeneous agents differing in capital and labor skills, a production function dependent on multiple technology parameters, and a social welfare function that aggregates utility across individuals. The planner in the model does not have perfect tools for redistribution (e.g. taxation incurs distortions or administrative costs). Instead, steering technology itself becomes a second-best policy for affecting how income is distributed.
Findings
- When is steering innovation most valuable?
Steering is especially important when social safety nets are inefficient. The ability to redirect the course of technological innovation—through policies like taxes, subsidies, or R&D nudges—becomes more attractive and effective for supporting workers if governments have limited capacity to redistribute income efficiently. If direct redistribution is costly or leads to distortion, guiding innovation offers a primary means of ensuring equitable outcomes.
- Properties of “worker-friendly” innovation
The desirability of technological innovation is determined by three factors:
– Labor complementarity: Technologies that enhance, rather than replace, human labor are more likely to result in higher wages and employment.
– Labor’s share in output: Worker outcomes are improved when innovations raise productivity in sectors that are already labor-intensive.
– Who benefits: The societal value of innovation is higher if it benefits lower-income or lower-skill workers, as these groups derive greater welfare from increased income.
- Redistribution vs. steering: A dynamic balance
– Redistribution and steering are complementary. When redistribution is costly, steering technological progress directly toward labor-friendly outcomes becomes preferable. But if redistribution is efficient and cheap, the need for steering diminishes.
– Existence of a threshold: As technological progress continues, there is a “critical threshold.” Before reaching it, steering can counteract labor devaluation. Beyond that point—if automation or AI makes human labor almost obsolete—steering alone is ineffective, and the focus must return to robust direct redistribution and social safety nets.
- Role of robot taxes and innovation subsidies
Robot (automation) taxes are justified when automation is likely to displace low-income, high-need workers. Optimal tax rates depend not just on technological impact but on which groups are most harmed or benefited. Policies should actively support R&D and deployment of innovations that make workers more productive or create new labor-complementary tasks.
- Task-based analysis: Beyond jobs lost
Automation eliminates some tasks but can create new, human-centered ones. Effective policy isn’t just about “protecting jobs” but also about “creating new types of work” where human skills are unique and valued. The net result for workers depends on the economy’s ability to facilitate transitions into these new areas.
- Practical policy recommendations
– Nudges and information: Governments can influence innovation direction through information campaigns, sharing best practices, or highlighting the social value of certain research agendas.
– Sector-specific strategies: Because impacts differ by industry, some sectors may require stronger steering than others—especially those with high labor intensity or greater vulnerability to automation.
– Integration with labor market policies: Worker retraining, wage subsidies, and employment supports should complement technology steering to buffer workers from disruption and help them transition to new roles.
Steering Technological Progress
Authors: Anton Korinek, Joseph Stiglitz
From: University of Virginia, Columbia University
