Rebalancing AI — Daron Acemoğlu, Simon Johnson

Source details

AI adoption could boost productivity growth by 1.5 percentage points per year over a 10-year period and raise global GDP by 7 percent ($7 trillion in additional output), according to Goldman Sachs.
Industry estimates include a supposed 10 percent chance of an “explosive growth” scenario, with global output rising more than 30 percent a year.
The authors argue that current techno-optimism relies on the “productivity bandwagon” belief that technological change inevitably drives higher productivity, higher net wages, and shared prosperity — a belief at odds with historical experience when automation primarily replaces workers.
The central policy imperative: direct AI development to complement workers and create new, human-complementary tasks rather than primarily automate and replace labor.

Productivity is defined as average output per worker (total output divided by total employment).
Firms respond to marginal productivity (the additional contribution of one more worker), not average output per worker.
Automation can raise average productivity while leaving marginal productivity unchanged or reduced, undermining incentives to hire.
Over the past four decades, automation raised productivity and corporate profits but did not lead to shared prosperity in industrial countries.

Innovations that boost worker marginal productivity do so by enabling individuals to contribute more to production or by creating new tasks.
Historical examples:
- Mass-production methods and assembly lines in car manufacturing (beginning in the 1910s) introduced new design, technical, machine-operation, and clerical tasks that increased demand for workers.
- Many fast-expanding occupations in recent decades (MRI radiologists, network engineers, computer-assisted machine operators, software programmers, IT security personnel, data analysts) did not exist 80 years ago.
New tasks have been vital to employment and wage growth over the past two centuries and have enabled new products and more efficient production processes.

Automation can increase employment if it yields substantial productivity gains that stimulate new jobs within or across industries (example: early 20th-century car manufacturing expanding oil, steel, chemical industries, and enabling new retail and service activities).
“So-so automation” produces limited productivity benefits and can reduce employment without creating offsetting jobs (example: self-checkout kiosks shifting scanning tasks to customers).
Rapid robot adoption in parts of the American Midwest contributed to mass layoffs and prolonged regional decline.
Surveillance-oriented technologies can extract more effort from workers, producing small productivity gains but worsening worker outcomes.
The authors emphasize choice in directing technological development toward surveillance, automation, or worker-empowering tools.

Historically, automation plus creation of new tasks proceeded in relative balance through much of the 20th century (notably between about 1870 and 1970), producing better-paying and less dangerous work.
After approximately 1970, the offsetting force of new task creation slowed while automation accelerated; this trend persisted over the ensuing five decades.
Workers without four-year college degrees have been particularly affected, increasingly concentrated in low-paying services such as cleaning, food service, and recreation.
AI could either accelerate automation without offsetting job creation or enable introduction of new labor-complementary tasks across diverse skill sets.
AI is likely to automate some tasks previously too flexible or judgment-based for pre-AI automation; the aggregate productivity impact depends on maturity, cost, and reliability of these technologies.
The prevailing intellectual paradigm in the digital tech sector favors achieving human parity across cognitive tasks, encouraging automation over human-complementary technologies.
AI also offers opportunities to complement worker skill and expertise (examples: training and retraining educators, medical personnel, modern craftspeople; creating demand for human oversight, customer communication, and sophisticated services).

Reform business models:
- Establish clear ownership rights for consumers over their data.
- Tax digital ads.
- Enable a more diverse range of business models or require more competition to reduce incentives for attention-grabbing and data expropriation.
Tax system:
- Create a more symmetric tax structure by equalizing marginal tax rates for hiring (and training) labor and for investing in equipment and software to shift incentives toward human-complementary choices.
Labor voice:
- Ensure workers have a voice in AI development.
- Restrict deployment of untested or insufficiently tested AI in applications that could put workers at risk (examples: high-stakes personnel decision-making tasks including hiring and termination, workplace monitoring and surveillance).
Funding for more human-complementary research:
- Increase public support for R&D in technologies that pair AI tools with human expertise in vital social sectors.
- Encourage competition and investment in human-complementary technologies and advise on adoption for publicly funded education and health care programs once sufficient progress exists.
AI expertise within government:
- Develop a consultative AI division within government to assist agencies and regulators for more timely, effective decision-making.

AI could increase global GDP over the next five years, although not as substantially as enthusiasts claim.
AI might modestly raise GDP growth in the medium term.
The likely first-order impact on the current trajectory is increased inequality within industrial countries.
Middle-income countries and many lower-income countries face risks from capital-intensive AI technologies; there is no guarantee AI will generate more jobs than it destroys on its current path.
Redirecting AI toward a human-complementary path and using it to address pressing social problems could allow global benefits; if a “just-automate” approach prevails, shared prosperity will be harder to achieve.

Source: Rebalancing AI — DARON ACEMOĞLU, SIMON JOHNSON; F&D Magazine; December 2023.

F&D: Rebalancing Ai
- F&D: Rebalancing Ai (Markdown version)
- F&D: Rebalancing Ai (PDF)