Highlights: Michael Webb on whether AI will soon cause job loss, lower incomes, & higher inequality

Exposure patterns by job tier and technology:

• Automation by robots predominantly impacts low-skilled, low-paid jobs, with exposure dropping sharply as income rises.
• Software automation historically targets middle-skilled jobs, leaving both the lowest and highest skill tiers relatively less exposed.
• AI exposure follows an inverted "U" curve, peaking at the 88th percentile of income, where upper-middle skilled roles (e.g., lawyers, accountants, doctors) face the highest risk.
• A recent OpenAI paper utilizing GPT-4 and different methodologies replicated the AI exposure curve, validating the finding that high-skill professional roles are most vulnerable.
• Despite high economic exposure, regulated professions (doctors, lawyers, accountants) possess significant political power to erect barriers and mitigate wage reductions or displacement.

Historical precedents for automation and employment:

• Research on ATMs demonstrates that while automation reduced tasks per branch, it lowered the cost of operation, prompting a net increase in the number of branches and total employment in banking.
• The shift to ATM technology allowed banks to expand into smaller towns, shifting human labor from cash handling to higher-value-added customer services.
• This dynamic is driven by demand elasticity and complementarity: as the cost of service drops, demand increases, often requiring more human staff despite higher automation per unit.

Individual-level displacement and "natural wastage":

• In many cases, automation eliminates jobs through "natural wastage" rather than mass firing, as companies stop hiring new staff while existing employees retire or leave for better opportunities.
• This process typically relies on short tenure rates (e.g., six months at McDonald's) to clear positions without immediate displacement shock.
• Displacement becomes severe when geographically concentrated; older workers in towns with a single large employer face enduring wage declines of approximately 25% if their local industry is automated.
• Younger workers in diverse urban labor markets can transition more easily, whereas mid-career workers in single-industry towns often lack mobility due to family, housing, and skill constraints.

Diffusion rates of past general-purpose technologies:

• The adoption curves for electrification (1894–1971) and computers follow nearly identical trajectories, taking roughly 30 years to reach 50% adoption.
• Despite continuous hardware improvements (e.g., Moore's Law), the US capital stock shifted from 0% to 8% software/hardware only between 1970 and 2000, with significant growth delayed until the mid-1990s.
• The telephone industry provides a 90-year case study: manual switchboard operator employment peaked 30 years after full automation technology was invented, with the final switch closed in 1980.
• Adoption delays are attributed to the difficulty of reorganizing corporate processes built around human labor and the high fixed costs of switching infrastructure, particularly in low-volume areas.

Regulatory and collective action as adoption bottlenecks:

• Powerful interest groups, such as the American Medical Association and Bar Councils, act as collective action barriers, often successfully slowing AI adoption to protect existing jobs and wages.
• Professional bodies control entry to professions (accreditation numbers) and standards, allowing them to mandate "human-in-the-loop" requirements.
• In sectors like education, strong unions and the autonomy of teachers make government mandates to adopt AI tools difficult to enforce, as human resistance to changing established workflows is common.
• Government action can move faster than technology; in extreme cases (e.g., nuclear classification), authorities can instantly halt research or deployment via executive order or legislation if perceived existential risks are sufficient.

Long-term economic projections and structural constraints:

• Historical precedent shows economies can adapt to massive task automation; US agriculture saw 90% of jobs automated over 150 years, resulting in a transformed but stable economy.
• Even if 90% of cognitive tasks are automated, non-cognitive tasks involving physical interaction or unique human elements will likely remain in high demand.
• Economic inertia, capital availability constraints, and the need to maintain ongoing production (e.g., "changing the planks while sailing") limit the speed of total workforce transition.
• While AI may eventually reach superhuman capabilities, the integration of such technology into the global economy is expected to occur gradually due to these structural and political frictions.