Will AI cause job loss, lower incomes and higher inequality — or the opposite? | Michael Webb

AI Exposure Patterns and Labor Market Impacts

• Differential exposure curves by technology type:
  • Robots: Expose low-skilled, low-wage jobs (e.g., manual labor) while leaving high-skilled roles largely untouched.
  • Software (legacy): Expose middle-skilled jobs most heavily, with low and high-skilled roles being less vulnerable.
  • Generative AI (current pattern): Exposes "upper-middle skill" jobs most severely; the curve starts low for low-wage work, peaks around the 88th salary percentile, and drops for the very top earners (e.g., CEOs), leaving lawyers and accountants highly exposed.
• High-exposure occupations (pre-GPT-4 baseline): Clinical lab technicians, chemical engineers, optometrists, power plant operators, and taxi dispatchers were identified as most exposed to automation based on patent data.
• High-exposure occupations (GPT-4 era): Interpreters, translators, journalists, poets, writers, mathematicians, and court reporters are now identified as highly exposed to generative AI capabilities.
• Productivity impacts by skill level:
  • Low-skill workers: Experience the largest relative productivity gains (e.g., customer support agents resolving 2x more issues per hour) as AI helps them reach the performance ceiling.
  • High-skill workers: Experience minimal relative gains in tasks with a performance ceiling, as they were already near optimal; however, they see massive absolute productivity boosts (e.g., 10x) when AI acts as a multiplier for complex, open-ended tasks.

Historical Analogies and Adoption Dynamics

• Speed of adoption: General-purpose technologies like electricity and computers typically took ~30 years to reach 50% adoption and ~90 years for full integration; however, AI may adopt faster due to lower switching costs and immediate utility.
• Lower friction for AI adoption:
  • Unlike electricity, which required massive physical infrastructure changes (re-wiring factories, moving from vertical to horizontal layouts), AI software can be integrated with minimal physical disruption.
  • AI models can interact with legacy code (e.g., COBOL) and convert it, removing the massive fixed costs and "switching costs" associated with traditional IT upgrades.
  • Generative AI is general-purpose from day one, unlike early computers which required specific "killer apps" (like VisiCalc) before widespread adoption.
• Decomposition of tasks:
  • Automation historically allows for the "reimagining" of processes (e.g., assembly lines replacing steam-driven vertical factories); AI similarly enables rethinking business structures by removing information bottlenecks for CEOs and reducing middle-manager control over information flow.

Economic Forces and Inequality Trends

• The "Iron Law" of automation: When productivity increases in one sector (e.g., food production), costs drop, freeing up consumer income to be spent on human-intensive services (e.g., dining out, therapy, housing), which creates new demand and jobs in those sectors.
• Historical wage shifts:
  • During early industrial revolutions, inequality often rises initially as capital owners capture the gains from automation.
  • Over decades (50–100 years), inequality often falls as the supply of capital stabilizes and labor becomes scarce relative to new technologies, driving up wages for workers.
• Geographic concentration of harm: Mass unemployment impacts are geographically concentrated; workers in single-employer towns with limited mobility face enduring wage declines (up to 25%) if their primary industry is automated, whereas urban, mobile workers adapt more easily.
• Natural wastage as a buffer: Automation often results in "natural wastage" where firms stop hiring for automated roles, allowing employment to decline through attrition rather than mass layoffs, particularly affecting younger cohorts who avoid entering shrinking industries.

Future Trajectories and Constraints

• Explosive growth vs. stagnation:
  • Optimists (e.g., Tom Davidson): Predict AI could accelerate innovation itself, leading to explosive economic growth by speeding up R&D and idea generation.
  • Skeptics (e.g., academic economists): Argue that historical data suggests productivity gains take decades to manifest in GDP, and AI will follow a similar slow diffusion curve without immediate statistical impact.
• Regulatory and Interest Group Barriers:
  • Professional bodies (e.g., AMA for doctors, Bar associations for lawyers) act as gatekeepers that can legally block or slow AI deployment even if the technology is superior.
  • Regulation may cap profits for dominant providers (e.g., capping token prices), potentially increasing access but possibly reducing incentives for frontier R&D unless structured carefully.
• Market structure: Currently dominated by a few providers (monopolies/oligopolies); broader benefits depend on making access cheap (like bandwidth) to allow widespread adoption across the economy.
• Future labor demand:
  • Human-intensive sectors: Demand is likely to surge for caregiving, teaching, and creative roles where human presence and empathy are valued, creating a potential labor shortage in these areas due to demographic shifts (more time in education/retirement).
  • Work hours: Historical trends suggest a long-term reduction in work hours (e.g., from 70 hours/week in 1870 to 35 hours/week today) as productivity rises, potentially leading to a 15-hour work week equilibrium in the long run.

Career Advice and Strategic Skills

• High-value skills for the AI era:
  • High technical + High social skills: Only roles combining technical proficiency with advanced interpersonal skills (charisma, negotiation, trust-building) have seen exceptional wage growth.
  • Context acquisition: Scarcity will lie in gathering private, non-public context (e.g., secret interviews, insider knowledge) that AI cannot access.
  • Orchestration: The ability to manage and direct swarms of AI agents to solve novel, unstructured problems will be a critical new managerial skill.
• Sectors to avoid: Jobs consisting purely of text-in/text-out tasks with low context requirements (e.g., minimum-wage content factories) are immediately at risk of total automation.
• Sectors to pursue: Roles in highly regulated industries (for stability) or industries with high AI transformation potential (for growth), provided the individual can add value via empathy, trust, and context.

Michael Webb's New Initiative

• Problem: Global talent shortages in AI safety and emerging technologies; universities are constrained by governance, outdated curricula, and competition with corporate labs for top researchers.
• Solution: Webb is launching a new organization (Quantum Leap Education) to rapidly train experts in AI safety and related fields, bypassing traditional academic bottlenecks.
• Status: The project is currently in stealth mode but will be hiring soon, focusing on accelerating the supply of qualified researchers in critical safety areas.