## AI’S PROMISE FOR THE GLOBAL ECONOMY

## Source details

**Canonical URL:** [AI’S PROMISE FOR THE GLOBAL ECONOMY](https://www.imf.org/-/media/files/publications/fandd/article/2024/09/spence.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/fandd/article/2024/09/spence.pdf.md)
- [Structured JSON version](/-/media/files/publications/fandd/article/2024/09/spence.pdf.json)

---

### Context and central thesis
- AI has the potential to reverse the downward productivity trend and, over time, produce a major sustained surge in productivity.
- Michael Spence’s best guess is that meaningful impacts in labor productivity may start to appear "by the end of this decade."
- Roy Amara’s law: overestimate short-run impacts, underestimate long-term ones.

### Macroeconomic headwinds and structural shifts
- Colliding forces constraining supply and raising costs:
  - Shocks: war, pandemic, climate change, geopolitical tensions, resurgent nationalism, and national security-focused economic policy.
  - Secular trends: declining productivity, aging populations in economies accounting for "more than 75 percent of global output," declining fertility, increasing longevity, and rising sovereign debt.
  - Technological revolution: digital transformation accelerated by AI, biomedical/life sciences revolution, and technologies for sustainable energy transition.
- Supply-chain fragmentation and resilience vs. cost trade-offs:
  - Apple now produces "15 percent of iPhones" in India as part of diversification.
  - Only South Korea and Taiwan Province of China currently make the most advanced semiconductors (as opposed to design), creating national-security concerns.
  - Result: a rapid postpandemic fragmentation of global supply networks; impossible to maximize resilience and minimize costs simultaneously.

### Productivity trends and fiscal context (exact figures preserved)
- US labor productivity growth:
  - 1998 to 2007: "1.68 percent" average.
  - 2010 to 2019: "0.38 percent" average.
- Sectoral productivity changes in the US:
  - Tradable goods and services sectors: from "4.27 percent" to "1.23 percent."
  - Large, less productive nontradable services sectors: from "0.73 percent" to effectively zero.
- Sovereign debt context:
  - Global sovereign debt now exceeds global gross domestic product.
  - United States sovereign debt ratio is "120 percent."
  - Europe’s ratio is "88.6 percent," with Greece, Italy, Spain, France, Belgium, and Portugal above the average (Greece and Italy by a lot).
  - China’s sovereign debt appears lower unless debt of state-owned enterprises is counted.
- Pandemic-era effects:
  - Measured productivity edged up during the pandemic, largely because less productive industries were partially shuttered while higher-productivity sectors shifted to remote work.
  - The balance-sheet damage during the pandemic was much less than during the global financial crisis, helping demand remain resilient as interest rates rose.

### Nature and promise of AI
- Generative AI:
  - First AI with a humanlike capacity to operate in multiple domains and detect/switch domains based on conversational prompts.
  - Capable of talking about inflation, writing computer code, and doing some mathematics (work in progress).
  - Better model: machine-human collaboration (“augmentation”) rather than full automation.
- Examples and spillovers:
  - AI systems like AlphaFold predict protein 3-D structures, showing domain-specific power.
  - AI can magnify productivity gains especially for less experienced workers (example: customer service and medical practice).
  - Potential to impact science and technology research across biology, physics, and materials science, and to play a key role in the energy transition.
- Constraints and enablers:
  - Main barriers to building powerful generative AI models: talent, computing power, and rapidly expanding electricity demand.
  - Availability of data is not a major constraint; the internet has ample training data.
  - Systems powerful enough to train models with billions of parameters reside largely in private-sector cloud computing in the United States and China.
  - Mega-platform business models rely on personal data and precise targeting, but training large language models does not require personalized and sensitive data.

### Distribution, diffusion, and competition risks
- Access and diffusion challenges:
  - For full economic impact, AI must be accessible across all sectors and to companies large and small.
  - Risk that market forces alone will produce divergence in adoption; broad diffusion is not guaranteed.
  - Sectors that tend to lag: government, health care, construction, and hospitality.
  - Small and medium enterprises require attention to adopt AI effectively.
- Geographic and policy implications:
  - Europe risks falling behind the United States and China for three reasons:
    - Relative underfunding of basic research.
    - Lack of computing power to support research.
    - Failure to fully leverage the large scale of the European economy due to fragmented capital markets, incomplete service market integration, and fragmented national regulation.
  - China is described as an AI powerhouse.
  - India is likely to be a growing force due to strong digital roots, a large and growing internal market, and deep engineering human capital.
  - Emerging market economies will mainly be consumers of advanced AI technology from the US and China for the next few years.
- Regulatory and market-policy balance:
  - Current policy emphasis is stronger on risk mitigation and misuse than on accessibility, diffusion, and skills.
  - Important to expand policies that promote diffusion and skills without abandoning risk mitigation.
  - Competition policy should be part of the package; government should avoid "picking winners or national champions."
  - Democratizing computing infrastructure and expanding access for researchers and innovators is critical to support widespread diffusion.

### Labor-market impacts and skills
- Nature of job changes:
  - AI-driven structural change and disruption will be long-lived and large-scale.
  - Workers in the middle of the skill distribution will be most impacted—jobs may not vanish but will change.
  - Retraining and new skills acquisition deserve priority attention.
  - Both private and public sectors have roles in smoothing transitions.
- Automation bias and the “Turing Trap”:
  - Risk that AI will be treated as full automation and replacement for humans, which can drive pessimistic employment narratives.

### Potential macroeconomic effects if AI diffuses broadly
- With policy support to accelerate diffusion, AI could:
  - Significantly accelerate economic growth.
  - Help productivity growth rebound.
  - Relax supply-side constraints that contribute to inflation, potentially lowering real interest rates and the cost of capital over time.
  - Facilitate the trillions of dollars of investment needed for energy efficiency and the green transition.
  - Help younger working populations support older populations in aging economies without undue sacrifice.

### Policy priorities (enumerated)
- Strengthen policies for accessibility, diffusion, and skills to realize AI’s full potential.
- Maintain and implement regulation to prevent misuse of the technology and data (risk-mitigation agenda).
- Expand computing infrastructure access beyond mega-platforms to democratize model-building and research.
- Promote competition policy to avoid excessive concentration and ensure broad-based benefits.
- Focus on adoption in lagging sectors (government, health care, construction, hospitality) and support small and medium enterprises.
- Prioritize retraining and new skills acquisition to manage labor-market transitions.

*Source: Michael Spence, F&D, September 2024.*

---


_Source: https://www.imf.org/-/media/files/publications/fandd/article/2024/09/spence.pdf_