How Could Better AI Shrink Its Economic Share?

TL;DR

Advanced AI could receive a smaller economic share if automated goods become abundant and consumers quickly become satiated, while scarce human involvement retains value. The outcome depends on demand for new automated goods, willingness to pay for human participation, and whether labor remains complementary to capital across production networks.

Transcript

Today I'm chatting with Alex Imas, who  is Director of AGI Economics at Google   DeepMind and Professor of Economics at the  University of Chicago, and Phil Trammell,   who is Head of Economics at Epoch  and research scholar at Stanford.  In general what I want to understand in this  interview is what economics tells us about   what we can expect i... Read More

Key Insights

• Economic value accrues to whatever remains scarce after automation. If machines can produce many goods and services without human input, scarce human time, human participation, land, ownership claims, or unsatisfied forms of demand may determine where income and bargaining power concentrate.
• The relational sector consists of goods and services whose value partly depends on a human remaining involved. A doctor could belong to this sector even if machines handle most administrative tasks, provided consumers pay more for a human to deliver diagnoses and emotional support.
• A separate machine economy could become increasingly self-contained because machines do not demand human-provided experiences. Humans would still purchase automated goods, causing wealth to flow from human services into machine production, while machines would have little corresponding reason to purchase services from humans.
• Economic forecasting about automation has repeatedly missed structural change. David Ricardo correctly anticipated that valuable jobs of his era would be automated, but failed to anticipate how lower prices, greater spending power, expanding services, and newly created jobs could sustain high employment.
• Labor share is the portion of economic output paid to people through wages, while capital share goes to owners of machines, land, buildings, companies, and related claims. The transcript states that labor has received roughly 60% for hundreds of years despite extensive automation.
• Network-adjusted factor shares account for labor and capital throughout an entire supply chain rather than examining only the final production step. A product that appears automated can still embody substantial labor through the research, machinery, components, and supporting processes required to produce it.
• Fully automated supply chains do not necessarily make capital's overall economic share dominant. If automated output expands but consumers become satiated even faster, its marginal utility and value could fall, leaving scarce human-centered services with a larger share of measured economic activity.
• Better evidence requires measuring consumer demand elasticities, job creation and destruction, occupational tasks, and willingness to pay for human involvement. Conjoint analysis could compare machine-only services with alternatives that preserve one human task, revealing which activities genuinely belong to the relational sector.
• More videos from this channel:
  • How did Meiji Westernization upset Asia's balance of power
  • What is the core idea of general relativity explained?
  • What is the true timeline for AGI and automation?
• More from this channel: Dwarkesh Patel

Questions & Answers

Q: What could remain scarce after advanced AI automates production?

Human time and human involvement could remain scarce because the number of humans is limited, even when machines can produce abundant goods and services. Consumers may value a human ballerina, barista, doctor, or other provider precisely because a person participates. Other outcomes depend on whether people remain unsatisfied with automated goods and whether AI continually creates valuable new varieties.

Q: What is the relational sector in an automated economy?

The relational sector includes goods and services for which human participation is part of what consumers value. It is broader than live performance or hospitality. A doctor's administrative tasks might be automated while the human still delivers a diagnosis and provides support. If consumers pay more for that remaining human task, the service qualifies as relational under the proposed framework.

Q: Why might better AI receive a smaller share of the economy?

Better AI could make automated products extraordinarily abundant, but abundance does not guarantee a large economic share. If consumers become satiated and the marginal utility of additional automated output falls faster than production rises, spending may concentrate on scarce human-centered experiences. The quantity produced by machines could therefore soar while its share of economic value becomes smaller.

Q: What are labor share and capital share?

Labor share is the portion of total economic output paid to people through wages. Capital share is the portion paid to owners of machines, land, buildings, companies, and related claims. The transcript says labor has received roughly 60% of the economy for hundreds of years, with the remaining 30–40% going to capital owners.

Q: Why has labor share remained stable despite automation?

Labor may remain valuable because it complements capital across production networks. Even when the final step for a product looks automated, people can contribute through earlier stages that create machines, components, knowledge, and supporting services. The transcript also notes an accounting controversy: when accounting methods are held constant, one cited analysis finds that labor share may not have fallen.

Q: How could complete supply-chain automation change the economy?

Complete supply-chain automation would represent a qualitative break from earlier automation because some goods could reach a network-adjusted capital share of one. No human labor would be required anywhere in their production, and consumers would not intrinsically care about human involvement. Even then, the overall capital share would remain ambiguous because demand and satiation determine the value of automated output.

Q: Why are confident forecasts about AI and employment difficult?

Confident forecasts are difficult because economists lack reliable data about consumer demand elasticities, the jobs being created or destroyed, occupational tasks, and preferences for human participation. Historical predictions also missed structural change and new employment. The interview therefore recommends mapping multiple scenarios and identifying the particular scarcity conditions and evidence that would make each scenario possible.

Q: How can researchers measure demand for human involvement?

Researchers could use conjoint analysis to compare willingness to pay for a fully machine-produced service with willingness to pay for a version that preserves one human task. This approach could estimate how strongly consumers resist removing humans from particular activities. The resulting evidence would help distinguish genuinely relational jobs from roles where human involvement adds little consumer value.

Summary & Key Takeaways

• The economic effects of advanced AI depend on what remains scarce after automation. Human participation may retain value in relational services, while fully automated goods could become abundant. Whether value shifts toward labor or capital depends on consumer demand, satiation, product variety, and preferences for keeping humans involved.

• Labor has historically received about 60% of economic output despite extensive industrial automation. One explanation is that labor continues adding value throughout production networks, even when final production steps appear automated. Advanced AI could create a qualitative change by enabling some entire supply chains to operate without human labor.

• The interview favors scenario analysis over confident forecasts because economists lack crucial evidence about demand elasticities, job creation, job destruction, and human preferences. Researchers could work backward from possible outcomes, identify the scarcity conditions required for each, and collect data through tools such as conjoint analysis and improved occupational tracking.