Interview
Will AI cause job loss, lower incomes and higher inequality — or the opposite? | Michael Webb
- Predicts AI will not achieve the capability to write an undergraduate-level essay on Michel Foucault by the 2040s.
- Anticipates the rate of AI progress will settle somewhere between continued acceleration and slowing, with improvements similar to GPT-3 to GPT-4 likely within the next year or two.
- Forecasts an "explosive economic growth" regime driven by automated innovation and easier idea discovery, though GDP impacts may be delayed for decades due to a productivity paradox similar to the 30-year adoption timeline for computers.
- Expects inequality to follow a historical cycle where capital owners initially benefit, with workers' income share catching up over a 50 to 100-year period as capital becomes less scarce.
- Projects current AI exposure will dominate upper-middle-skill jobs like law and accounting, dipping for top executives, while a "hollowing out" of the software engineering middle will occur as lower-skilled workers perform high-training tasks and higher-skilled roles focus on orchestration.
- Envisions a shift in labor value toward "orchestration" (directing many superior AI assistants), a blend of high technical and social skills, and roles relying on personal networks and trust, while those failing to leverage AI for 10x productivity gains will be out-competed.
- Anticipates that mass unemployment will not occur even if 90% of cognitive tasks are automated, as surplus spending will flow to human-labor-intensive services like caregiving, education, and live human arts performance, which will become scarce and valuable.
- Predicts the average work week could decline toward 15 hours as workers trade income for leisure amid rising productivity, while the primary mechanism for labor displacement will be natural wastage over 10 to 20-year timeframes rather than mass firings.
- Warns that severe negative economic impacts will be geographically concentrated, causing enduring wage declines of around 25% for older workers in small towns dependent on a single automated employer.
- Forecasts AI adoption will be faster than previous general-purpose technologies due to near-zero marginal costs and general-purpose nature, but significantly slower than technological capability allows due to regulatory barriers and interest groups.
- Warns that the US government could classify all AI research overnight via executive order or legislation if safety concerns become sufficiently alarming, effectively shutting down development.
- Anticipates government regulation could cap AI service prices similar to utility companies, transforming providers into regulated entities, while market structure may evolve from a near-monopoly to a larger number of regulated providers.
- Identifies private context obtained through human interviews as a scarce asset, and predicts a critical talent bottleneck in AI safety where university systems cannot expand capacity rapidly enough due to governance and faculty resistance.