Executive Summary

The late autumn of 2025 has marked a decisive shift in the economic literature surrounding Artificial Intelligence (AI). Moving beyond the initial hype cycle of 2023 and the speculative modeling of 2024, the research papers released in October and November 2025 by the world’s leading economic institutions—including the National Bureau of Economic Research (NBER), the Centre for Economic Policy Research (CEPR), the IZA Institute of Labor Economics, the International Monetary Fund (IMF), and the Organisation for Economic Co-operation and Development (OECD)—collectively depict a global economy entering a period of profound structural fracture.

This report synthesizes this emerging body of work to argue that we are witnessing a "Great Divergence" driven by the specific technological characteristics of Generative and Agentic AI. This divergence is manifesting across three critical axes: a labor market fracture where AI augments incumbents while displacing entrants; a firm-level fracture where "superstar" firms decouple from the rest of the economy; and a geopolitical fracture where the concentration of compute infrastructure threatens to siphon welfare from adopting nations to producing nations.

Unlike the Skill-Biased Technological Change (SBTC) of the late 20th century, which increased the premium on education, the evidence from late 2025 suggests that the current wave is "Experience-Biased." Empirical studies using massive administrative datasets now confirm that AI is substituting for junior human capital—the "canaries in the coal mine"—severing the traditional training pipeline that sustains the workforce. Simultaneously, while micro-level productivity gains are undeniable and often substantial, the aggregate macroeconomic impact remains constrained by the "Acemoglu Ceiling," creating a paradox of high local optimization and low global growth.

The following analysis is divided into four primary sections: Labor Market Dynamics, Industrial Organization and Market Structure, Macroeconomic Productivity and Growth, and Global Policy Implications. Each section rigorously deconstructs the methodologies and findings of the disparate working papers to present a unified, if unsettling, picture of the AI economy at the close of 2025.

1. Labor Market Dynamics: The Crisis of Entry and the Paradox of Experience

The most prolific and urgent area of research in late 2025 concerns the labor market. The theoretical debate between "substitution" (AI replacing workers) and "complementarity" (AI making workers better) has transitioned into a phase of rigorous empirical interrogation. The emerging consensus from high-frequency administrative data is that AI is not ending work, but it is radically restructuring the entry point into work, creating a crisis for early-career professionals that policymakers have largely unanticipated.

1.1 The "Canaries in the Coal Mine": Empirical Evidence of Junior Displacement

The most alarming empirical evidence released in November 2025 comes from the Stanford Digital Economy Lab and NBER. The working paper "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence" by Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen utilizes high-frequency administrative data from ADP to track labor market shifts following the widespread adoption of Generative AI ^[1]. This paper represents a landmark study because it moves beyond job posting data to actual payroll records, offering the first definitive look at who is actually being hired—or not hired.

The authors document a stark divergence in employment outcomes based on age and experience, which they distill into six stylized facts. The most critical of these is the finding that early-career workers (ages 22–25) in occupations highly exposed to AI experienced a 16% relative decline in employment compared to unexposed sectors ^[1]. This decline persists even after controlling for firm-level shocks, suggesting that the effect is driven by the technology itself rather than idiosyncratic business cycles.

In direct contrast, employment levels for experienced workers in the very same AI-exposed occupations remained stable or continued to grow. This finding overturns the conventional wisdom that automation targets "routine" tasks regardless of who performs them. Instead, it suggests that Generative AI is a substitute for junior human capital specifically. The tasks typically assigned to entry-level associates—drafting initial reports, summarizing meetings, basic coding, and data cleaning—are exactly the tasks where Large Language Models (LLMs) excel.

The mechanism of this adjustment is equally telling. The researchers found that the labor market is adjusting primarily through quantity (hiring freezes and non-replacement) rather than price (wage reductions). Firms are not cutting the wages of junior staff; they are simply ceasing to hire them. This creates a "silent" crisis: current employment statistics, which track the stock of employed workers, may mask the collapsing flow of new entrants. The "canaries" in this context are the recent graduates, whose labor is being devalued not because they lack potential, but because their initial utility to the firm can now be generated synthetically at near-zero marginal cost.

Furthermore, the study finds that employment declines are concentrated in occupations where AI is technically capable of automating tasks rather than merely augmenting them. This distinction is crucial. In roles where AI serves as a tool to enhance human capability (augmentation), employment remains robust. In roles where AI can fully execute the core deliverables of a junior employee (automation), demand is collapsing. The implications for the "training pipeline" are severe: if firms substitute AI for junior staff, they destroy the mechanism by which junior staff become senior experts, potentially creating a long-term shortage of high-level human capital.

1.2 Global Corroboration: The Collapse of Demand in Job Postings

Corroborating the payroll-based findings of the NBER, the World Bank released a significant report in November 2025 titled "Labor Demand in the Age of Generative AI: Early Evidence from the U.S. Job Posting Data" ^[4]. While the NBER paper looked at filled jobs, the World Bank study analyzes intent to hire, utilizing a massive dataset of 285 million job postings collected by Lightcast from 2018 through the second quarter of 2025.

This study utilizes the release of ChatGPT in November 2022 as an exogenous shock to identify causal impacts. The results paint a picture of accelerating displacement. The authors find that job postings for occupations with above-median AI substitution scores fell by an average of 12% relative to unexposed roles. Crucially, this effect was not a one-time adjustment but has intensified over time. The decline was 6% in the first year following the shock, growing to 18% by the third year (late 2025) ^[5].

The breakdown of these declines is instructive. Administrative support roles have seen a precipitous 40% drop, suggesting that "AI Agents"—software capable of planning and executing multi-step workflows—are rapidly replacing human administrative assistants. More worrying for the graduate labor market is the 30% decline in professional services and the 20% decline in positions requiring no experience. This data supports the hypothesis that Generative AI raises the threshold for employability. To be viable in the 2025 labor market, a worker must offer capabilities that exceed the "zero-shot" performance of frontier models. For fresh graduates without specialized domain knowledge, this bar is increasingly difficult to clear.

1.3 The Productivity Paradox: Augmentation for the Insider

While the outlook for the outsider (the applicant) is grim, the outlook for the insider (the employee) is characterized by significant productivity gains. This paradox is central to understanding the current economic moment: AI makes novices more productive, which paradoxically reduces the aggregate demand for novices.

An October 2025 IZA Discussion Paper, "How AI-Augmented Training Improves Worker Productivity" ^[6], provides detailed firm-level evidence from the financial services sector. The study analyzes the staggered rollout of an AI tool designed to provide real-time performance feedback to call center agents. This setting offers a clean "natural experiment" to measure the impact of AI as a performance-enhancing tool.

The study yields several key findings regarding the "Experience-Bias" of AI augmentation:

• Overall Productivity: The AI tool reduced the Average Handling Time (AHT) of calls by 10% ^[6]. This is a massive productivity boost in an industry where margins are measured in seconds.
• The Levelling Effect: The gains were not distributed equally. Short-tenured (novice) workers saw productivity gains of 17%, compared to just 7% for long-tenured veterans.
• Mechanism of Action: The study identifies exactly how the AI helped novices. It drastically reduced the time novices spent putting clients on hold to consult supervisors or look up information. Effectively, the AI acted as an "instant supervisor," closing the knowledge gap between a 6-month employee and a 2-year employee.

Combining the IZA findings ^[6] with the Stanford/World Bank findings ^[1] reveals the macroeconomic mechanism at play. Because one AI-augmented junior employee can now do the work of 1.2 or 1.5 non-augmented employees, the firm reduces its aggregate headcount for entry-level roles. The "augmentation" observed at the micro-level drives the "displacement" observed at the macro-level. The firm has optimized its internal operations by making its junior staff more efficient, but in doing so, it has reduced its capacity to absorb new entrants from the labor market.

1.4 Curricula and Long-Term Adaptation: The Race Against Obsolescence

How is the supply side of the labor market—the education and training systems—responding to this demand shock? Another significant IZA paper from November 2025, "Expertise at Work: New Technologies, New Skills, and Worker Impacts" ^[7], examines this question using a novel database of vocational training curricula in Germany over five decades.

The researchers use Natural Language Processing (NLP) to link curriculum updates to technological breakthroughs. They find that the German vocational system is responsive: technological change spurs curriculum updates, specifically shifting content toward digital and social skills while reducing routine task intensity.

The labor market consequences of these updates highlight the growing volatility of human capital value:

• The Premium for New Skills: "New-skilled" workers—those entering the labor market with updated training—earn wage premiums of up to 5.5% in technology-exposed occupations ^[7].
• The Penalty for Old Skills: In stark contrast, older incumbents (ages 55–65) experience wage declines of nearly 10% when new-skilled workers enter their field. This finding is consistent with rapid "skill obsolescence."

This implies that the "race between education and technology" (a concept popularized by Goldin and Katz) is intensifying. The shelf-life of a specific skill set is shrinking. While educational institutions are attempting to adapt, the speed of AI deployment—as evidenced by the rapid 3-year decline in job postings—may be outpacing the refresh rate of formal education. Workers are facing a reality where continuous re-skilling is not just a buzzword but a prerequisite for wage maintenance. The "career plateau," where a worker could rely on skills learned decades ago, has effectively collapsed.

2. Industrial Organization and Market Structure: The "Coasean Singularity" vs. The Compute Monopoly

Moving beyond the labor market, the research landscape in late 2025 has broadened to include the industrial organization of the AI economy itself. This literature grapples with a fundamental contradiction: downstream, AI "Agents" promise to fragment the firm and democratize the market; upstream, the infrastructure required to power those agents is driving unprecedented industrial concentration.

2.1 The "Coasean Singularity": AI Agents and the Boundaries of the Firm

A standout theoretical contribution from NBER in November 2025 is "The Coasean Singularity? Demand, Supply, and Market Design with AI Agents" by Peyman Shahidi, Gili Rusak, Benjamin S. Manning, Andrey Fradkin, and John J. Horton ^[8]. This paper revisits the Theory of the Firm, originally proposed by Ronald Coase in 1937, which posits that firms exist because transaction costs (search, negotiation, enforcement) make it too expensive to contract every task out to the open market.

Horton and his colleagues argue that "AI Agents"—autonomous systems that can perceive, reason, and act—act as "transaction cost destruction machines."

• Search: Agents can scan the entire market for goods, services, or labor in milliseconds, drastically reducing search frictions.
• Negotiation: Agents can engage in complex, multi-round negotiations instantly, balancing multiple variables (price, quality, delivery time) in ways that exceed human cognitive capacity.
• Contracting: Smart contracts and automated verification can reduce enforcement costs.

If transaction costs collapse toward zero—a theoretical point the authors term the "Coasean Singularity"—the economic necessity of large, vertically integrated firms diminishes. We might witness a shift toward a "hyper-modular" economy, where value is created not by monolithic corporations but by dynamic networks of autonomous agents contracting for specific micro-tasks. The boundary between the firm and the market dissolves. The authors suggest that agent adoption is a form of "derived demand"—users want the outcome (a booked flight, a coded module) and employ agents to minimize the effort of procurement. This shift could lead to more efficient markets but also introduces new risks, such as "congestion" (agents spamming markets with queries) and "price obfuscation" (algorithms colluding to hide true costs) ^[10].

2.2 The Reality of Concentration: The OECD on AI Infrastructure

However, the dream of a decentralized, agent-led economy faces a hard constraint in the physical reality of the AI supply chain. The OECD released a significant policy paper in November 2025 titled "Competition in Artificial Intelligence Infrastructure" which throws cold water on the idea of a decentralized utopia ^[11].

The report analyzes the AI "compute stack"—comprising chips, data centers, and cloud providers—identifying it as a "multi-layered" ecosystem characterized by extreme barriers to entry.

• Economies of Scale: The fixed costs of training frontier models and building semiconductor fabrication plants are astronomical, creating natural monopolies.
• Vertical Integration: The report notes a concerning trend where "Hyperscalers" (major cloud providers) are increasingly designing their own chips and taking equity stakes in model developers (e.g., the Microsoft/OpenAI or Amazon/Anthropic models). This vertical integration locks up the stack, making it difficult for independent players to compete at any single layer ^[12].
• Switching Barriers: The report highlights technical and commercial barriers that prevent customers from switching between cloud providers or model families, further entrenching the power of incumbents ^[14].

The OECD explicitly warns of "exclusionary practices," where dominant firms might bundle products or use data advantages to foreclose rivals. It calls for heightened merger control to prevent the "kill zone" phenomenon, where incumbents acquire nascent rivals before they can pose a competitive threat. The report suggests that without active regulatory intervention, the AI infrastructure market will calcify into an oligopoly, extracting rents from the rest of the economy.

2.3 Concentrating Intelligence: The Risk of Market Tipping

Supporting the OECD’s analysis, Anton Korinek and Jai Vipra’s paper "Concentrating Intelligence: Scaling and Market Structure in Artificial Intelligence" (published in Economic Policy Jan 2025 but circulated as an NBER Working Paper in late 2024/2025) argues that the "cost structure of foundation models" inherently leads to market tipping ^[15].

The authors identify "economies of scale and scope" in AI development that are so powerful they create a "winner-take-most" dynamic. The larger the model, the more capable it is; the more capable it is, the more data it generates to train the next iteration. This positive feedback loop suggests that "intelligence" may become a centralized utility, supplied by a handful of entities. Korinek and Vipra warn that this concentration is not just an economic issue but a political one: if a few firms control the "cognitive infrastructure" of the global economy, they wield unprecedented power over information flow and decision-making processes.

2.4 International Trade and Welfare: The "Double Harm"

Susan Athey and Fiona Scott Morton’s NBER Working Paper "Artificial Intelligence, Competition, and Welfare" (November 2025) extends this analysis to the global stage, modeling AI as a "priced, imported factor" ^[17].

They consider a general equilibrium model where a country imports AI technology from a foreign provider. If that upstream provider has market power (as suggested by the OECD/Korinek analysis), the importing country can suffer a "double harm":

• Labor Displacement: Unskilled workers in the importing country lose their jobs to the AI adoption.
• Rent Extraction: The efficiency gains generated by this displacement do not accrue to the domestic economy (as lower prices or higher profits). Instead, they are siphoned off by the foreign AI monopolist via high licensing fees (usage fees) or access fees ^[18].

This finding provides a robust economic rationale for "Technological Sovereignty." It suggests that nations cannot afford to be mere consumers of AI. To capture the welfare gains of the AI revolution, nations may need to ensure competitive access to AI technology, potentially through regulation of foreign providers or investment in domestic compute capacity. The paper concludes that broad gains for an adopting country depend on "pressure (or regulation) on both usage and access fees," as well as policies that support the productive absorption of displaced labor.

3. Macroeconomic Productivity and Growth: The Skeptics vs. The Optimists

While the microeconomic disruptions are clear, the debate over the aggregate macroeconomic impact of AI remains unresolved in late 2025. The tension lies between the "Techno-Optimists," who forecast a return to robust productivity growth, and the "Structural Realists," led prominently by Daron Acemoglu, who argue that the mathematical ceiling for these gains is lower than popularly imagined.

3.1 The "Simple Macroeconomics" of Limited Impact

Daron Acemoglu (MIT/NBER) continues to serve as the primary skeptic of the "AI Boom" narrative. His paper "The Simple Macroeconomics of AI," along with subsequent commentaries in late 2025, provides a rigorous counter-weight to industry hype ^[19].

Acemoglu’s skepticism is grounded in Hulten’s Theorem, which dictates that the aggregate productivity impact of a technology is the product of the fraction of tasks it affects and the average cost savings per task.

• Limited Exposure: Acemoglu estimates that AI currently impacts a fraction of tasks that contribute to roughly 4.6% of GDP.
• Finite Savings: Even if AI makes these tasks significantly cheaper, the aggregate lift is constrained by their small share of the total economy.
• The Forecast: Based on these parameters, Acemoglu projects a Total Factor Productivity (TFP) increase of roughly 0.66% over 10 years—a modest 0.06% per year ^[20].

In commentary published in the IMF's Finance & Development and other outlets in late 2025, Acemoglu warns of "So-So Automation" ^[22]. This occurs when technology is just good enough to replace a worker but not good enough to generate massive productivity surpluses. In this scenario, the economy suffers the negative demand shock of displaced wages without the positive supply shock of vastly cheaper goods. He argues that if AI is deployed primarily to cut labor costs (rent dissipation) rather than to enable new capabilities, the gains will be minimal and heavily skewed toward capital owners.

3.2 The Counter-Narrative: Deflation and Diffusion

The "Techno-Optimist" case is bolstered by data from the Stanford Institute for Human-Centered AI (HAI) and its 2025 AI Index Report ^[23]. This perspective focuses not on the current share of tasks, but on the deflationary shock that drives future adoption.

The report highlights a staggering decline in the cost of "inference"—the computational process of running an AI model.

• Metric: The cost to query a model with GPT-3.5 level performance dropped 280-fold between November 2022 and October 2025 ^[25].
• Hardware: Costs for AI hardware have declined by 30% annually, while energy efficiency has improved by 40% per year.

Optimists argue that Acemoglu’s static analysis misses the dynamic expansion of the "exposed task" set. When the cost of intelligence drops by two orders of magnitude, tasks that were previously uneconomical to automate (e.g., personalized tutoring, complex legal discovery, real-time code generation) suddenly become viable. This suggests a Jevons Paradox effect: as intelligence becomes cheaper, the economy will consume vastly more of it, expanding the "share of GDP" impacted by AI far beyond Acemoglu’s 4.6% estimate.

3.3 The IMF’s "Tenuous Resilience" and Global Divergence

The International Monetary Fund’s World Economic Outlook (WEO), released in October 2025, attempts to bridge these views, characterizing the global economy as showing "tenuous resilience" ^[26].

The IMF projects global growth at 3.0% for 2025, an upward revision reflecting the resilience of major economies. However, the WEO identifies a "Global Divergence" in AI readiness.

• AI Preparedness Index: The IMF uses this metric to show that advanced economies are well-positioned to harness AI to offset aging workforces. In contrast, emerging markets face a "double risk": they may lose their comparative advantage in low-cost labor (as AI automates business process outsourcing) while lacking the digital infrastructure to capture AI's productivity upside ^[29].
• Policy Warning: The IMF warns that "pressure on the independence of key economic institutions" could undermine the sound policymaking required to navigate this transition. The report suggests that without strong "guardrails" and social safety nets, the displacement effects of AI could lead to social instability that negates its economic benefits.

4. Firm-Level Adoption: The Rise of Agents and the "Superstar" Effect

Connecting the macro trends to the micro reality, reports from McKinsey and NBER examine how firms are actually adopting these technologies in late 2025.

4.1 From Chatbots to Agents: The 2025 Pivot

McKinsey’s "State of AI 2025" report identifies a critical pivot in corporate adoption: the move from Generative AI (content creation) to Agentic AI (action execution) ^[30].

While 88% of organizations report using AI in some capacity, only about 30% have successfully scaled it across the enterprise. However, 62% of survey respondents say their organizations are experimenting with AI agents. This aligns with Horton’s "Coasean Singularity" thesis—companies are preparing for a world where software doesn't just draft the email, but sends it, schedules the meeting, and negotiates the vendor contract.

Despite the hype, the "EBIT impact" remains elusive for the majority. Only 39% of companies report a measurable contribution to Earnings Before Interest and Taxes (EBIT) from AI adoption ^[30]. This confirms that we are still in the "Solow Paradox" phase: "We see AI agents everywhere but in the productivity statistics."

4.2 The "Superstar Firm" Effect

NBER Working Paper 33509, "Artificial Intelligence Adoption and Firm Dynamics," adds a critical dimension to the adoption story ^[31]. The authors find that AI adoption is heavily skewed toward larger, more productive firms.

• Concentration of Gains: AI adoption leads to higher firm-level sales, profits, and Total Factor Productivity (TFP).
• Divergence: Because adoption is concentrated among industry leaders (who have the data infrastructure and capital to invest), AI serves to widen the gap between "superstar firms" and the rest of the economy.

This finding reinforces the OECD’s concerns about market concentration. AI appears to be a technology that rewards scale. If the benefits of AI accrue primarily to the largest firms, the economy may become less dynamic over time, with reduced competitive pressure from smaller challengers who cannot afford the "table stakes" of AI integration.

5. Global Policy and Governance: Navigating the Fracture

The research literature in late 2025 culminates in a unified call for robust policy intervention. The laissez-faire approach of the early 2020s is widely viewed as inadequate for the structural challenges posed by Agentic AI.

5.1 Regulating Risk: The Sandbox Approach

NBER Working Paper 31921, "Regulating Artificial Intelligence," proposes a theoretical framework for optimal regulation under uncertainty ^[32]. The authors argue that Pigouvian taxes (taxing AI to pay for externalities) are insufficient because regulators cannot accurately observe the risks. Instead, they propose a "two-stage optimal policy":

• Stage 1: A "Sandbox" period of experimentation (beta testing, red teaming) to reduce uncertainty about societal costs.
• Stage 2: A binary decision to either publicly release or withdraw the algorithm based on the information gathered.

This "Regulatory Sandbox" model is gaining traction as a way to balance the need for innovation (highlighted by the Optimists) with the need for safety (highlighted by the Realists).

5.2 Political Economy: AI and Democracy

Finally, Daron Acemoglu and co-authors explore the intersection of AI and politics in "AI and Social Media: A Political Economy Perspective" (June 2025, cited in late 2025 literature) ^[33]. They identify mechanisms by which AI-driven social media platforms can distort democratic discourse through the "digital ads channel" and the "social media channel." This research suggests that the concentration of AI power is not just an economic threat but a threat to democratic stability, necessitating governance structures that enforce transparency and accountability.

5.3 Regional Responses

The global policy response is fragmenting along regional lines, as evidenced by specific summit declarations:

• Africa: The "Cotonou Declaration" (November 2025) from the Regional Summit on Digital Transformation in Western and Central Africa emphasizes AI as a driver of inclusive growth but demands investments in digital infrastructure to prevent the region from being left behind ^[34].
• Middle East: The IMF's selected issues paper on Qatar highlights the country's active preparation and "AI readiness," suggesting that resource-rich nations are attempting to buy their way into the AI tier ^[35].