## The Macroeconomics of Artificial Intelligence

## Source details

**Canonical URL:** [The Macroeconomics of Artificial Intelligence](https://www.imf.org/-/media/files/publications/fandd/article/2023/december/20-25-brynjolfsson-final.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/fandd/article/2023/december/20-25-brynjolfsson-final.pdf.md)
- [Structured JSON version](/-/media/files/publications/fandd/article/2023/december/20-25-brynjolfsson-final.pdf.json)

---

### Overview
- AI systems exhibit intelligent behavior, such as learning, reasoning, and problem-solving, and have the potential to transform productivity growth, the labor market, and industrial concentration.
- The future of AI is not predetermined; it depends on technological and policy decisions made today.
- Better outcomes require:
  - creative policy experiments;
  - a set of positive goals for what society wants from AI (not just negative outcomes to be avoided);
  - flexibility to adapt as technological possibilities evolve rapidly.

### First fork: Productivity growth
- Context:
  - The US economy has experienced disturbingly low productivity growth for most of the past 50 years, except for a brief resurgence in the late 1990s and early 2000s (Brynjolfsson, Syverson, and Chad 2019).
  - Productivity—output per unit of input—largely determines national wealth and living standards.

- Low-productivity future (risks and mechanisms)
  - AI adoption may remain slow and confined to large firms (Zolas and others 2021).
  - AI could be a narrow labor-saving technology (a “so-so technology”) rather than enabling novel, powerful worker capabilities.
  - Displaced workers might end up in less productive jobs, muting aggregate productivity gains.
  - Economic gains may appear decades after initial technological promise (Solow-style paradox).
  - Organizational and managerial failures could blunt benefits if firms fail to adapt.
  - Legal and regulatory regimes may hinder development, for example through uncertain intellectual property implications of training on massive datasets, potential creation of “patent thickets,” strict regulations, or even outright bans by some actors.

- High-productivity future (opportunities and mechanisms)
  - AI might be applied to a substantial share of tasks performed by most workers (Eloundou and others 2023), massively boosting productivity in those tasks.
  - AI could complement workers, freeing them for nonroutine, creative, and inventive tasks.
  - AI can capture and embody tacit knowledge by drawing on vast digitized data, enabling more workers to tackle novel problems.
  - Outcome: an economy at a permanently higher productivity growth rate, not just a higher level.
  - Integration of AI with robots could expand the amenable parts of the economy.
  - AI-backed research could enable radical advances in medicine, biology, drug design, and even accelerate scientific discovery and AI development itself (recursive self-improvement).

### Second fork: Income inequality
- Context:
  - Over the past 40 years, computers and information technology have contributed to rising income inequality by automating routine middle-income jobs and polarizing the labor force (Autor, Levy, and Murnane 2003).

- Higher-inequality future (risks and mechanisms)
  - AI is designed and implemented primarily as a substitute for human labor, pushing down wages for many workers.
  - Generative AI produces words, images, and sounds—tasks once considered nonroutine and creative—expanding the set of jobs under threat.
  - Entire industries could be upended; displacement could be extensive but not necessarily result in mass unemployment.
  - Many workers could be relegated to low-paying service jobs (hospital orderlies, nannies, doormen) where human presence is valued but pay is low.
  - Labor market polarization intensifies: a small high-skilled elite and a large underclass of poorly paid service workers.

- Lower-inequality future (opportunities and mechanisms)
  - AI primarily augments the least experienced or least knowledgeable workers, raising their productivity (example: Copilot for software coders).
  - Evidence: a study of 5,000 complex customer assistance workers found the least skilled or newest workers using AI assistants showed the greatest productivity gains (Brynjolfsson, Li, and Raymond 2023).
  - If employers shared productivity gains with workers, income distribution could become more equal.
  - Removing tedious routine work via AI could complement creative tasks, improving the psychological experience of work.
  - The call center study also found reduced worker turnover and increased customer satisfaction for workers using AI assistants.

### Third fork: Industrial concentration
- Context:
  - Since the early 1980s, industrial concentration has risen dramatically in the United States and other advanced economies; large “superstar” firms are often more capital-intensive and technologically sophisticated.

- Higher-concentration future (risks and mechanisms)
  - Only the largest firms would intensively use AI in core business functions, becoming more productive, profitable, and larger.
  - AI models may be extremely expensive to develop and require massive datasets that large firms already possess.
  - Example operational costs cited:
    - The GPT-4 model cost more than $100 million to train during its initial development and requires about $700,000 a day to run.
    - The typical cost of developing a large AI model may soon be in the billions of dollars.
  - Executives predict scaling laws that relate increased training costs to improved performance, favoring firms with the biggest budgets and datasets.
  - Even if proprietary AI does not create insurmountable fixed costs, AI may help large firms better internally coordinate complex operations, strengthening the “visible hand” of top executives and challenging decentralization advantages of small firms.

- Lower-concentration future (opportunities and mechanisms)
  - Open-source AI models (examples: Meta’s LLaMA, Berkeley’s Koala) become widely available, fostering a vibrant open-source AI ecosystem across for-profit companies, nonprofits, academics, and individual coders.
  - Broad access to developed AI models gives small businesses access to industry-leading production technologies.
  - A leaked Google internal memo suggested “open-source models are faster, more customizable, more private, and pound-for-pound more capable” than proprietary models and could be trained more cheaply; open-source processes can be repeated by many and may outcompete slow large-team iteration.
  - AI may encourage decentralized innovation; more AI-backed innovators might prefer ownership of small firms over employment in large firms, potentially reversing the long rise in industrial concentration.

### Toward a policy agenda
- Central diagnosis:
  - The path of least resistance leads to low productivity growth, higher income inequality, and higher industrial concentration; reaching the better path requires smart policy interventions.

- Policy framing and key questions
  - Move beyond a hydraulic “more AI or less AI” debate to understanding how policies shape the direction of AI development.
  - Key policy questions include:
    - How to encourage AI that complements human labor rather than simply replacing it?
    - What choices encourage development of AI accessible to firms of all sizes rather than only the largest?
    - What kind of open-source ecosystem is required and how can policymakers support it?
    - How should AI labs and firms approach model development and implementation?
    - How to promote AI that unleashes radical innovation instead of marginal tweaks?

- Actors with power to influence AI’s direction
  - Major corporations (decisions on AI integration and in-house development)
  - AI/computer science labs at universities (open-source development)
  - Federal legislators and regulators (and local regulators)
  - Voters
  - Labor unions (deciding their relationship with AI and demands)

- Research and investment needs
  - There is an imbalance between research advancing AI technology and research on its economic and social consequences.
  - Given that AI effects may be measured in trillions of dollars, far greater investment should be made in research on the economics of AI.
  - Society needs innovations in economic and policy understanding that match the scale and scope of AI breakthroughs to reorient research priorities and develop a smart policy agenda for sustained and inclusive growth.

### Supporting evidence and referenced studies
- Major themes referenced:
  - Labor market impact potential of Large Language Models (LLMs).
  - Advanced technologies adoption and use by U.S. firms.

- Cited studies in the section:
  - Eloundou, Tyna, Sam Manning, Panels Mishkin, and Daniel Rock. 2023. “GPTs Are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models.” arXiv pre-print arXiv:2303.10130.
  - Zolas, Nicholas, Zachary Kroff, Erik Brynjolfsson, Kristina McElheran, David N. Beede, Cathy Buffington, Nathan Goldschlag, Lucia Foster, and Emin Dinlersoz. 2021. “Advanced Technologies Adoption and Use by U.S. Firms: Evidence from the Annual Business Survey.” NBER Working Paper 28290, National Bureau of Economic Research, Cambridge, MA. https://www.nber.org/papers/w28290.

*Erik Brynjolfsson and Gabriel Unger, “The Macroeconomics of Artificial Intelligence,” F&D, December 2023.*

### Section 1

### The Macroeconomics of Artificial Intelligence

### Overview
- AI systems exhibit intelligent behavior, such as learning, reasoning, and problem-solving, and have the potential to transform productivity growth, the labor market, and industrial concentration.
- The future of AI is not predetermined; it depends on technological and policy decisions made today.
- Better outcomes require: creative policy experiments; a set of positive goals for what society wants from AI (not just negative outcomes to be avoided); and flexibility to adapt as technological possibilities evolve rapidly.

### First fork: Productivity growth
- Context:
  - The US economy has experienced disturbingly low productivity growth for most of the past 50 years, except for a brief resurgence in the late 1990s and early 2000s (Brynjolfsson, Syverson, and Chad 2019).
  - Productivity—output per unit of input—largely determines national wealth and living standards.

- Low-productivity future
  - AI adoption may remain slow and confined to large firms (Zolas and others 2021).
  - AI could be a narrow labor-saving technology (a “so-so technology”) rather than enabling novel, powerful worker capabilities.
  - Displaced workers might end up in less productive jobs, muting aggregate productivity gains.
  - Economic gains may appear decades after initial technological promise (Solow-style paradox).
  - Organizational and managerial failures could blunt benefits if firms fail to adapt.
  - Legal and regulatory regimes may hinder development, for example through uncertain intellectual property implications of training on massive datasets, potential creation of “patent thickets,” strict regulations, or even outright bans by some actors.

- High-productivity future
  - AI might be applied to a substantial share of tasks performed by most workers (Eloundou and others 2023), massively boosting productivity in those tasks.
  - AI could complement workers, freeing them for nonroutine, creative, and inventive tasks.
  - AI can capture and embody tacit knowledge by drawing on vast digitized data, enabling more workers to tackle novel problems.
  - Outcome: an economy at a permanently higher productivity growth rate, not just a higher level.
  - Integration of AI with robots could expand the amenable parts of the economy.
  - AI-backed research could enable radical advances in medicine, biology, drug design, and even accelerate scientific discovery and AI development itself (recursive self-improvement).

### Second fork: Income inequality
- Context:
  - Over the past 40 years, computers and information technology have contributed to rising income inequality by automating routine middle-income jobs and polarizing the labor force (Autor, Levy, and Murnane 2003).

- Higher-inequality future
  - AI is designed and implemented primarily as a substitute for human labor, pushing down wages for many workers.
  - Generative AI produces words, images, and sounds—tasks once considered nonroutine and creative—expanding the set of jobs under threat.
  - Entire industries could be upended; displacement could be extensive but not necessarily result in mass unemployment.
  - Many workers could be relegated to low-paying service jobs (hospital orderlies, nannies, doormen) where human presence is valued but pay is low.
  - Labor market polarization intensifies: a small high-skilled elite and a large underclass of poorly paid service workers.

- Lower-inequality future
  - AI primarily augments the least experienced or least knowledgeable workers, raising their productivity (example: Copilot for software coders).
  - Evidence: a study of 5,000 complex customer assistance workers found the least skilled or newest workers using AI assistants showed the greatest productivity gains (Brynjolfsson, Li, and Raymond 2023).
  - If employers shared productivity gains with workers, income distribution could become more equal.
  - Removing tedious routine work via AI could complement creative tasks, improving the psychological experience of work.
  - The call center study also found reduced worker turnover and increased customer satisfaction for workers using AI assistants.

### Third fork: Industrial concentration
- Context:
  - Since the early 1980s, industrial concentration has risen dramatically in the United States and other advanced economies; large “superstar” firms are often more capital-intensive and technologically sophisticated.

- Higher-concentration future
  - Only the largest firms would intensively use AI in core business functions, becoming more productive, profitable, and larger.
  - AI models may be extremely expensive to develop and require massive datasets that large firms already possess.
  - Example operational costs cited:
    - The GPT-4 model cost more than $100 million to train during its initial development and requires about $700,000 a day to run.
    - The typical cost of developing a large AI model may soon be in the billions of dollars.
  - Executives predict scaling laws that relate increased training costs to improved performance, favoring firms with the biggest budgets and datasets.
  - Even if proprietary AI does not create insurmountable fixed costs, AI may help large firms better internally coordinate complex operations, strengthening the “visible hand” of top executives and challenging decentralization advantages of small firms.

- Lower-concentration future
  - Open-source AI models (examples: Meta’s LLaMA, Berkeley’s Koala) become widely available, fostering a vibrant open-source AI ecosystem across for-profit companies, nonprofits, academics, and individual coders.
  - Broad access to developed AI models gives small businesses access to industry-leading production technologies.
  - A leaked Google internal memo suggested “open-source models are faster, more customizable, more private, and pound-for-pound more capable” than proprietary models and could be trained more cheaply; open-source processes can be repeated by many and may outcompete slow large-team iteration.
  - AI may encourage decentralized innovation; more AI-backed innovators might prefer ownership of small firms over employment in large firms, potentially reversing the long rise in industrial concentration.

### Toward a policy agenda
- The path of least resistance leads to low productivity growth, higher income inequality, and higher industrial concentration; reaching the better path requires smart policy interventions.
- Policy framing:
  - Move beyond a hydraulic “more AI or less AI” debate to understanding how policies shape the direction of AI development.
  - Key policy questions include:
    - How to encourage AI that complements human labor rather than simply replacing it?
    - What choices encourage development of AI accessible to firms of all sizes rather than only the largest?
    - What kind of open-source ecosystem is required and how can policymakers support it?
    - How should AI labs and firms approach model development and implementation?
    - How to promote AI that unleashes radical innovation instead of marginal tweaks?
- Actors with power to influence AI’s direction:
  - Major corporations (decisions on AI integration and in-house development)
  - AI/computer science labs at universities (open-source development)
  - Federal legislators and regulators (and local regulators)
  - Voters
  - Labor unions (deciding their relationship with AI and demands)
- Research and investment needs:
  - There is an imbalance between research advancing AI technology and research on its economic and social consequences.
  - Given that AI effects may be measured in trillions of dollars, far greater investment should be made in research on the economics of AI.
  - Society needs innovations in economic and policy understanding that match the scale and scope of AI breakthroughs to reorient research priorities and develop a smart policy agenda for sustained and inclusive growth.

*Erik Brynjolfsson and Gabriel Unger, “The Macroeconomics of Artificial Intelligence,” F&D, December 2023.*

### Section 2

### 20-25-brynjolfsson-final - Section 2

### Major themes referenced
- Labor market impact potential of Large Language Models (LLMs).
- Advanced technologies adoption and use by U.S. firms.

### Cited studies in this section
- Eloundou, Tyna, Sam Manning, Panels Mishkin, and Daniel Rock. 2023.  “GPTs Are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models.” arXiv pre-print arXiv:2303.10130.
- Zolas, Nicholas, Zachary Kroff, Erik Brynjolfsson, Kristina McElheran, David N. Beede, Cathy Buffington, Nathan Goldschlag, Lucia Foster, and Emin Dinlersoz. 2021. “Advanced Technologies Adoption and Use by U.S. Firms: Evidence from the Annual Business Survey.” NBER Working Paper 28290, National Bureau of Economic Research, Cambridge, MA. https://www.nber.org/papers/w28290.

*20-25-brynjolfsson-final - Section 2*

---


_Source: https://www.imf.org/-/media/files/publications/fandd/article/2023/december/20-25-brynjolfsson-final.pdf_