Automation and the Labor Market

When people worry that AI will take their jobs, they are picturing automation at the job level — a machine walks in and replaces a person. Economists have found this framing too coarse to be useful. The more precise and powerful way to think about automation is at the task level. Jobs are bundles of tasks. AI automates some tasks while complementing others, and the net effect on employment depends on exactly which tasks are affected and how the rest of the job evolves around them.

The Task-Based Framework

Economists David Autor, Frank Levy, and Richard Murnane introduced the influential task-based framework in 2003. Their core insight: what matters is not which occupations exist but which tasks are being performed, and whether those tasks are routine or non-routine, cognitive or manual. Routine tasks follow explicit, codifiable rules — a file clerk retrieving records by category, a bookkeeper applying standard accounting entries, an assembly-line worker performing the same physical motion thousands of times a day. These tasks are easy to automate because the rules can be programmed. Non-routine tasks require judgment, pattern recognition in novel situations, or social interaction — a nurse reading a patient's emotional state, a lawyer constructing an argument from ambiguous precedents, a carpenter improvising when a beam is not plumb. These tasks are harder to automate. Automation does not typically eliminate a job entirely — it eliminates specific tasks within a job, forcing the role to evolve. Radiologists no longer hold X-ray film to a light box; digital imaging automated that. But the core diagnostic task of interpreting complex scans in clinical context remained and arguably expanded. Bank tellers' cash-counting tasks were automated by ATMs — yet teller employment did not collapse for two decades because banks opened more branches and tellers shifted toward customer relationship work.

The current wave of AI is different from previous automation waves in one critical respect: it reaches cognitive, non-routine tasks that prior automation could not touch. Industrial robots automated routine manual tasks. Spreadsheets automated routine cognitive tasks. But generative AI and large language models can now perform tasks that require reading, writing, coding, image interpretation, and even some forms of reasoning — tasks that seemed uniquely human just ten years ago. This means the task-based framework must be extended. The new boundary is not routine vs. non-routine — it is whether a task involves pattern recognition over large, learnable domains (which AI now does very well) versus genuine novelty, physical dexterity in unpredictable environments, complex interpersonal trust, or moral judgment (where AI remains limited). Researchers Eloundou and colleagues (2023) classified 1,016 detailed work activities across 923 occupations for their exposure to large language models. They found that roughly 80 percent of U.S. workers are in occupations where at least 10 percent of their tasks are exposed to LLM automation — and 19 percent of workers are in occupations where 50 percent or more of tasks are exposed. This exposure is concentrated in white-collar, higher-wage occupations — a significant reversal from the historical pattern where automation hit lower-wage, routine-manual workers first.

Labor-Market Polarization

One of the most robust empirical findings in labor economics over the past thirty years is labor-market polarization: the job market has hollowed out in the middle. Employment has grown at the top (high-skill, high-wage professional and managerial work) and at the bottom (low-skill, low-wage personal service work), while middle-skill, middle-wage jobs — the occupations that dominated the 20th-century middle class — have declined as a share of employment. This pattern is not a coincidence. Middle-skill jobs tend to be routine — bookkeeping, clerical work, production supervision, repetitive manufacturing — precisely the tasks that prior waves of automation hit hardest. The result: a dumbbell-shaped labor market, with growth at both ends and erosion in the middle. AI introduces a new wrinkle. Prior automation spared high-skill cognitive workers almost entirely. The new AI wave is different — it directly challenges paralegal research, first-pass financial analysis, entry-level coding, basic radiology screening, and other tasks performed by educated workers who occupy the upper-middle of the wage distribution. If AI erodes the upper-middle of the wage distribution in addition to the middle, the shape of polarization could become more extreme — or could shift as new high-skill roles emerge around AI systems themselves. The empirical evidence on this is still developing, and honest economists acknowledge significant uncertainty about the direction and magnitude of the effect.