Rockefeller Foundation

2025 Journal#2025-10-04-Sat

Preparing for Rockefeller Foundation - Schmidt Sciences Bellagio Convening:

The AI Economy and the American Worker: An Analytical Synthesis for Strategic Action

Executive Summary

The rapid proliferation of generative artificial intelligence (AI) represents a pivotal moment for the American labor market, presenting a complex landscape of unprecedented opportunity and significant risk. This report synthesizes a range of empirical studies, economic theories, and real-world use cases to provide a comprehensive foundation for the ‘Scenarios and Solutions: Preparing for the AI Economy’ conference. The analysis reveals that AI is being adopted at a historically unmatched pace, but its diffusion is highly uneven, concentrating in high-income regions and among skilled professionals, which risks deepening economic inequality. A critical divergence has emerged between AI’s use as a broad personal utility by the public and its deployment as a specialized automation engine by enterprises.

The central tension identified is whether AI will primarily automate human labor, leading to a bifurcated economy, or augment human capabilities, fostering a revitalized middle class. The evidence suggests that while current corporate adoption patterns lean towards automation, the technology’s underlying applicability is broad, spanning a wide range of middle-skill occupations in fields like sales and administrative support. This indicates that the future trajectory is not technologically predetermined but will be shaped by the choices made by employers, workers, and policymakers.

Exploring future scenarios in key sectors such as healthcare, education, and public administration reveals the tangible consequences of these choices. An automation-centric path could de-skill professionals and displace administrative workers, while an augmentation-focused approach could empower a wider set of employees, improve service quality, and enhance job satisfaction. To navigate this transition, this report outlines a framework for action centered on three priorities: shaping employer practices to favor human-complementary AI, strengthening worker voice in technological deployment, and implementing public policies that support equitable transitions. Finally, it presents a research roadmap built around nine grand challenges in the economics of AI, providing a structure for the urgent work required to ensure that the AI economy leads to broad-based prosperity.

Part I: The AI Economy in Practice: Current Impacts on the American Worker

This section addresses the first conference goal: to establish a common understanding of how AI is affecting American workers today. It synthesizes empirical data from multiple sources to paint a detailed, nuanced picture of the present landscape, moving beyond hype to documentable reality.

1.1 The Unprecedented Pace and Uneven Pattern of AI Adoption

The diffusion of AI into the American workplace is occurring at a speed with no historical precedent. Unlike previous general-purpose technologies such as electricity or the personal computer, which took decades to achieve widespread use, Generative AI has reached a significant portion of the workforce in just a few years.1 Data from September 2025 indicates that 40% of U.S. employees now report using AI at work, a figure that has doubled from 20% in 2023.1 This rapid uptake is further evidenced by the massive user base of platforms like ChatGPT, which reached 700 million weekly users globally—approximately 10% of the world’s adult population—less than three years after its launch.3 This velocity is fueled by a relentless pace of innovation, with new, more powerful Models being released continuously, and by staggering capital investments in the underlying computational infrastructure.5

However, this rapid adoption is not uniform. A defining characteristic of the current landscape is its unevenness, which manifests geographically and socioeconomically. The Anthropic Economic index reveals a strong positive correlation between AI usage and economic output. Globally, high-income nations like Israel and Singapore exhibit the highest per-capita AI usage, while emerging economies lag significantly.2 This pattern is mirrored within the United States, where usage is concentrated in wealthier states and regions with knowledge-based industries. Washington D.C. and Utah, for example, show per-capita usage rates that are 3.82 and 3.78 times their share of the population, respectively, far outpacing less affluent states.2 This data provides a strong empirical basis for concerns that, if left unaddressed, AI could amplify existing economic divides rather than bridge them.7

1.2 A Tale of Two AIs: Divergent Usage in Consumer and Enterprise Contexts

A critical divergence is emerging between how AI is used by individuals and how it is deployed by businesses. This split in usage patterns has profound implications for public perception versus economic reality.

On one hand, AI is evolving into a personal utility for the general public. A comprehensive analysis of ChatGPT usage reveals that non-work-related messages have grown to constitute over 70% of all consumer interactions, up from 53% the previous year.3 The most common use cases are broad and personal in nature: “Practical Guidance,” “Seeking Information,” and “Writing”.4 Furthermore, early demographic skews are evening out, with the gender gap in usage having narrowed dramatically, suggesting AI is achieving mainstream acceptance as a general-purpose tool for life management, learning, and communication.3 This broad, collaborative public experience may foster a perception of AI as a helpful, benign assistant.

On the other hand, enterprise deployment of AI appears to be more narrowly focused and heavily skewed towards automation. Analysis of business use of AI through application programming interfaces (APIs) presents a starkly different picture. For enterprise clients, 77% of AI tasks are “directive,” meaning the user delegates the full task to the AI, compared to a 39% directive rate for consumer users.2 The tasks themselves are also more specialized; coding and office/administrative work dominate enterprise use, while the shares of educational and creative tasks are significantly lower than in consumer patterns.6 This indicates a strong business incentive to deploy AI as an “automation engine” for direct task replacement and efficiency gains. The divergence between the public’s experience with a collaborative AI assistant and the corporate implementation of an automation tool may create a societal blind spot, where the true, structural economic impacts on the labor market are not widely understood or debated.

1.3 The Augmentation-Automation Spectrum in the Modern Workplace

Within the workplace, AI is being applied across a spectrum from augmentation (enhancing human capabilities) to automation (substituting for human labor). While current enterprise trends lean towards automation, a powerful counter-narrative and significant evidence suggest a vast potential for augmentation, which holds the key to AI’s long-term impact on job quality and wages.

The strongest quantitative indicator of an automation focus is the finding that businesses delegate 77% of tasks to AI in an end-to-end fashion.2 This aligns with the widespread concerns about mass job loss and technological unemployment.10 However, a more optimistic perspective, articulated by economist David Author, posits that AI’s unique opportunity is to “extend the relevance, reach and value of human expertise”.11 In this view, AI acts as a cognitive tool that can empower a broader set of workers to perform higher-stakes decision-making tasks currently reserved for elite experts.11

Empirical data provides support for both ends of the spectrum. Analysis of Microsoft Bing Copilot usage shows that common user goals are information gathering and writing, with the AI often acting as a “coach, advisor, or teacher” to the human user—a clear augmentation function.14 Similarly, the OpenAI study finds that two-thirds of “writing” tasks at work involve modifying existing text (e.g., editing, summarizing, translating), rather than creating content from Scratch.3 This suggests a common workflow where AI produces the “first draft,” elevating the human role to one of refinement, judgment, and strategic direction. A compendium of 601 real-world use cases provides concrete examples of this spectrum in action. Deutsche Bank’s research tool, which slashes the time to create reports from hours to minutes, is a clear augmentation of expert analysts, while AI-powered predictive ordering systems at fast-food chains could directly substitute for human order-takers.16

1.4 Early Signals from the Labor Market: Wages, Skills, and Opportunity

The initial economic impacts of AI on the American workforce are complex and seemingly contradictory, creating a pivotal tension between the risk of deepening inequality and the opportunity to rebuild the middle class.

On one side, the adoption data strongly suggests that the benefits of AI are currently concentrating among the highly skilled and well-compensated. The correlation of AI use with state-level GDP per capita and the finding that economic gains are accruing to “highly skilled professionals and enterprise clients” point toward a classic case of skill-biased technological change that could widen income disparities.2

However, a more nuanced reality emerges when analyzing the technology’s applicability to tasks, rather than just its current adoption. Research from Microsoft that analyzed the applicability of AI to a wide range of occupations found that the highest potential is in knowledge work, sales, and administrative support—a mix of highand middle-skill roles.14 Crucially, this study found only a “weak correlation between AI applicability scores and educational requirements” and a slightly higher applicability for high-wage, but not the highest-wage, occupations.15 This is a vital finding, as it suggests that AI’s potential is not confined to an elite sliver of the workforce but extends deep into the core of the labor market.

The discrepancy between the adoption data (concentrated at the top) and the applicability data (broad-based potential) highlights a critical “adoption gap.” The technology itself appears capable of augmenting a wide array of jobs, but current market forces, access to capital, and training opportunities are channeling its benefits narrowly. This refutes a purely deterministic view of technology’s impact. The central policy challenge, therefore, is not about stopping technology but about closing this gap and finding ways to translate AI’s broad applicability into broad-based economic benefits.

Metric	Anthropic Economic Index (Claude)	Chatterji et al. (ChatGPT)	Tomlinson et al. (Bing Copilot)
Primary Data Source	User conversations & Enterprise API traffic	Representative sample of consumer user messages	Anonymized user conversations
Top 3 Work-Related Use Cases	1. Coding (36%) 2. Education (12.4%) 3. Science (7.2%)	1. Writing (40%) 2. Seeking Information 3. Practical Guidance	1. Information Gathering 2. Writing 3. Communication
Key Automation Indicator	77% of enterprise API tasks are “directive” (fully delegated)	Not directly measured, but de novo creation is less common than modification	AI often acts in a service role (coach, advisor), suggesting less full automation
Key Augmentation Indicator	Shift from debugging to new code creation	~67% of writing tasks involve modifying existing text (editing, summarizing)	High prevalence of user goals related to research and learning
Occupations w/ Highest Usage/Applicability	Usage correlates with knowledge-based industries (e.g., IT in CA, Finance in FL)	Work usage is more common for educated users in highly-paid professional occupations	Sales, Computer & Mathematical, Office & Administrative Support
Correlation with Wages/Income	Strong positive correlation between usage and GDP per capita	Work usage more common in highly-paid occupations	Weak correlation with educational requirements and wages

Part II: Scenarios for the Future of Work

This section addresses the second conference goal: to explore multiple scenarios for the progress of AI and its impact on the labor market, with a focus on education, healthcare, and the public sector. It moves from description to projection, building out two plausible but divergent futures based on the evidence and theories presented.

2.1 The Bifurcation Scenario: A Trajectory Toward Deeper Inequality

This scenario extrapolates from the current concentration of AI benefits. If present trends of investment and adoption continue unabated, AI-driven productivity gains will accrue primarily to capital owners and a small class of elite knowledge workers who can effectively leverage the technology. For the majority of the workforce, particularly those in middle-skill roles, this future is characterized by wage stagnation, increased job precarity, and direct displacement as the pace of automation outpaces the creation of new, augmentation-focused roles.

This scenario is grounded in the empirical findings of the Anthropic report, which documents a strong link between AI adoption and regional income, the dominance of automation-focused deployments in enterprise settings, and the concentration of benefits among the highly skilled.2 It reflects the widely shared public concern, identified by researchers at Brookings, that AI’s net impact on jobs will be negative.20 In this vision, AI continues the trend of labor market polarization that began with earlier waves of computerization, hollowing out the middle and creating a stark divide between a small group of “AI masters” and a large group of workers in low-paid service jobs that are not easily automated.11

2.2 The Augmentation Scenario: A Pathway to a Rebuilt Middle Class

This scenario posits that AI, if developed and deployed with intention, can reverse the decades-long trend of labor market polarization. By serving as a powerful cognitive tool, AI can augment the skills of non-elite workers, increasing their productivity, decision-making capabilities, and economic value. This could lead to the restoration of stable, well-compensated middle-skill jobs and a more equitable distribution of economic gains.

This scenario is a direct exploration of the thesis advanced by David Author, who argues that AI’s unique potential lies in its ability to “weave information and rules with acquired experience to support decision-making”.11 This capability could allow workers with foundational domain knowledge, but not necessarily elite credentials, to perform high-stakes tasks previously reserved for top-tier experts like doctors, lawyers, and engineers.12 This optimistic view is supported by the empirical finding that AI’s

applicability is broad, with high potential in middle-skill occupations like sales and administrative support, and shows a weak correlation with formal education and wage levels.14 This suggests the technological foundation for broad-based benefits exists. Crucially, this outcome is presented not as a forecast but as an attainable possibility that depends on human agency and collective choice—a direct rebuttal to the idea of technological inevitability.11

2.3 Sectoral Deep Dive: Manifestations in Healthcare, Education, and the Public Sector

The abstract scenarios of Bifurcation and Augmentation become concrete when examined through the lens of specific sectors. By mapping real-world AI use cases onto these sectors, the tangible choices and consequences for workers become clear.

Healthcare

Bifurcation Scenario: AI is deployed primarily for administrative automation, such as billing and scheduling, leading to the displacement of clerical staff. Advanced diagnostic AI systems are controlled by specialists at elite medical centers, de-skilling regional practitioners who are reduced to the role of data-entry technicians. For example, an AI system for radiology screening, such as the one used by Apollo Hospitals, could be implemented in a way that reduces the need for entry-level radiologists rather than augmenting their diagnostic capabilities.16
Augmentation Scenario: AI tools empower nurses, physician assistants, and other mid-level practitioners to perform more advanced diagnostic and patient-management tasks under supervision, effectively extending the reach of medical expertise into rural and underserved communities. AI-powered search across vast clinical and research databases, like the system deployed at the Mayo Clinic, can accelerate the development of treatment plans and democratize access to cutting-edge knowledge for all clinicians, not just top researchers.16 This path requires significant investment in workflow redesign and retraining.

Education

Bifurcation Scenario: AI-driven tutoring platforms and automated grading systems are used to replace human teachers for core instruction, particularly in budget-constrained public schools. This relegates educators to roles as classroom managers and proctors, leading to de-professionalization, wage depression, and a two-tiered education system. This aligns with the observed decline in “education” as a share of enterprise AI use cases, suggesting businesses may not see a profitable path for augmentation-focused tools.2
Augmentation Scenario: AI tutors handle personalized, rote learning and practice, freeing human teachers to focus on fostering higher-order skills like critical thinking, collaboration, and creativity. AI assistants help teachers with administrative burdens like lesson planning and communication, allowing them to devote more time to individual student mentorship.10 This reflects the high share of “tutoring or teaching requests” in consumer AI use, indicating a strong public demand for AI as a learning aid that complements, rather than replaces, formal education.3

Public Sector

Bifurcation Scenario: Government agencies adopt AI primarily for cost-cutting, deploying chatbots and automated systems to handle citizen-facing services and benefits processing, leading to a reduction in the public sector workforce. AI-powered surveillance and performance-monitoring tools are used to increase managerial control over public employees, reducing their autonomy and professional judgment.
Augmentation Scenario: AI agents help social workers, case managers, and other public servants manage large caseloads by summarizing documents, flagging critical information, and handling routine paperwork. This frees up their time to focus on high-touch, empathetic human interaction and complex problem-solving. AI tools could help city planners and analysts process vast datasets to improve public services, augmenting their expertise and leading to better policy outcomes.

The path taken in each sector will depend heavily not just on the technology itself, but on the organizational capital—the investments in training, workflow redesign, and trust-building—that institutions are willing to make. A short-term focus on cost savings is likely to lead toward the Bifurcation Scenario, whereas a long-term focus on capability enhancement and service quality could pave the way for Augmentation.

Dimension	Bifurcation Scenario (Automation-Dominant)	Augmentation Scenario (Human-Centric)
Primary Economic Driver	Cost reduction & efficiency gains	Capability enhancement & quality improvement
Primary Beneficiaries	Owners of capital & elite experts	Broad-based labor, including middle-skill workers
Impact on Wage Distribution	Increased polarization; hollowing out of the middle	Compression; rebuilding of the middle class
Key Skill Requirements	Prompt engineering, data labeling, low-skill service work	Critical judgment, domain expertise, human-AI collaboration
Illustrative Healthcare Example	AI automates radiology readings, displacing junior radiologists.	AI assists nurses with initial diagnoses, expanding their scope of practice.
Illustrative Education Example	AI platforms deliver standardized lessons, replacing teachers.	AI tutors handle drills, freeing teachers for project-based learning.
Illustrative Public Sector Example	Chatbots replace human agents for citizen services.	AI summarizes case files, allowing social workers more client-facing time.
Guiding Philosophy	Technological inevitabilism; focus on substitution	Human agency; focus on complementarity

Part III: A Framework for Action: Policy Levers and a Research Roadmap

This section addresses the third conference goal: to brainstorm and flesh out practical actions and a research roadmap. It synthesizes the forward-looking and prescriptive elements of the readings to provide a structured basis for strategic discussion and planning.

3.1 Forging a Proactive Response: Shaping Practices, Voice, and Policy

Achieving the more optimistic Augmentation Scenario is not an automatic outcome of technological progress; it requires a concerted, multi-stakeholder effort to actively shape the direction of innovation and its deployment. A proactive response must focus on three key areas: influencing employer practices, strengthening worker voice, and implementing smart public policy.10

Priority 1: Employer Practices: The central challenge is to incentivize firms to choose the harder but more broadly beneficial path of augmentation over the simpler path of automation. The difference between these paths often comes down to a firm’s willingness to invest in “organizational capital”—the complementary processes, skills, and structures needed to effectively integrate new technology.2 Proposed actions include the development and promotion of “good jobs” frameworks for AI deployment, which prioritize job quality alongside productivity; direct investment in workforce training and workflow redesign; and the creation of industry standards for human-in-the-loop systems that ensure human oversight and judgment remain central.
Priority 2: Worker Voice and Influence: To ensure that AI systems are designed to support rather than undermine workers, those with direct knowledge of their own tasks must be involved in the technology’s design and implementation. The successful negotiation by Hollywood writers to place guardrails on AI’s use in their profession serves as a powerful model.21 Proposed actions include supporting sector-based bargaining and co-determination models, promoting the use of AI for worker empowerment (e.g., enhancing workplace safety or providing access to information) rather than solely for managerial control, and creating clear channels for worker feedback throughout the AI lifecycle.
Priority 3: Public Policy Levers: Government has a critical role in creating an economic and regulatory environment that steers AI development toward societal benefit and provides a robust safety net for those who are negatively affected. Current tax codes, for instance, often favor capital investment in automation over equivalent investments in labor and training; reforming these could level the playing field. Other proposed actions include directing public R&D funding toward human-complementary AI, modernizing social Insurance programs like unemployment to support longer and more frequent career transitions, and establishing lifelong learning accounts to fund continuous upskilling. These policies are essential for managing the “transition dynamics” of a rapidly changing economy.22

3.2 Defining the Research Frontier: The Nine Grand Challenges

Our understanding of AI’s economic and societal impact is still in its infancy. To guide effective policy and private action, a systematic and ambitious research agenda is urgently needed. A framework proposed by Erik Brynjolfsson, Anton Korinek, and Ajay Agrawal identifies nine “Grand Challenges” for the economics of Transformative AI (TAI), providing an authoritative roadmap for this work.22

Economic Growth: How will TAI affect the fundamental drivers of productivity and long-run economic growth?
Innovation: How will TAI change the process of scientific discovery and invention itself, potentially accelerating the rate of new ideas?
Income Distribution: Who will be the winners and losers? How will TAI affect the distribution of income between labor and capital, and across different skill and wage groups? This directly addresses the tension between the Bifurcation and Augmentation scenarios.
Decision-making Power: How will AI shift economic and political power within firms, between organizations, and among nations?
Geoeconomics: How will TAI impact international trade, competition, and the stability of the global economic order?
Information Flows: How can societies manage the challenges of AI-generated misinformation and ensure a healthy, trustworthy information ecosystem?
Safety Risks: How can we analyze and mitigate the catastrophic risks posed by TAI, from economic collapse to existential threats?
Human Well-being: How can we ensure that TAI enhances human flourishing beyond narrow economic metrics like GDP, encompassing factors like health, community, and purpose?
Transition Dynamics: How can we best manage the potentially rapid and disruptive transition period to a TAI-enabled economy to minimize harm and maximize benefit? This challenge encompasses the policy questions raised by Kinder et al..10

3.3 Synthesis and Strategic Recommendations for Conference Deliberation

The future of the AI economy is not a technical question to be answered by engineers but a societal choice to be made by all stakeholders. The analysis presented in this report distills this choice into a set of core tensions that should frame the conference’s deliberations:

Automation vs. Augmentation: Is AI primarily a substitute for human labor, or is it a tool that complements human skills?
Inequality vs. Opportunity: Will AI’s primary effect be the concentration of wealth and the hollowing out of the middle class, or can it be a force for creating broad-based opportunity?
Determinism vs. Agency: Is the future of work something that will happen to us, or is it something we can collectively shape through our actions and policies?

Based on these tensions, the following recommendations are offered to guide the conference’s brainstorming sessions:

Develop Metrics Beyond Productivity: Acknowledge the limitations of traditional economic measures like GDP in the AI era. Brainstorm new metrics and methods for assessing AI’s impact on job quality, worker autonomy, economic security, and overall human well-being.
Launch Pilot Programs for Augmentation: Propose the creation of public-private partnerships to pilot and scale augmentation-focused AI deployments in key middle-skill sectors. Potential areas include community healthcare clinics, public school systems, and support services for small and medium-sized businesses. The goal is to create proven models that can help close the “adoption gap.”
Establish a Framework for Worker Voice: Move from abstract principles to practical implementation guides. Brainstorm concrete models for ensuring meaningful worker participation in the procurement, design, and deployment of AI systems in their workplaces.
Prioritize the Research Roadmap: Urge conference participants—including academics, philanthropic funders, and government agencies—to align their efforts and resources around the “Nine Grand Challenges.” A coordinated effort to accelerate our collective understanding is essential for developing the evidence-based policies needed to navigate the profound transition ahead.