How is AI changing the economy? A source-backed timeline of research, arguments, and scenarios.
Internal records show junior roles falling from 27% to 9% of technical hires as Anthropic expands its use of agents. Output per employee rises in coding and several business functions, while work shifts toward agent supervision, judgment and coordination. Total technical hiring still grows; the report documents changes in hiring composition rather than an overall employment decline.
August surveys compare 10,980 US adults with 1,334 Anthropic staff. Employed US respondents expect AI to be capable of doing an average 28% of their current work hours by 2030, versus 75% among staff. Both groups associate greater expected automation of their own work with higher expected unemployment. These are beliefs about the future, not measured capabilities or job losses.
Epoch tests six models on 11 tasks drawn from its own work, including design, data analysis and research planning. Frontier systems handle many defined steps but miss organizational standards and research judgment needed to deliver complete work. The tests distinguish capable assistance from autonomous replacement.
InnovationEval asks agents to devise and test a new model-training technique against a human-developed reference. With large GPU budgets, the two main models failed to match the reference: one made limited improvements using an existing idea, while the other's apparent gains failed scope review. Newer models already knew the reference technique.
Carey separates semiconductors from broader electronics in international input-output tables. In the 2022 estimates, a dollar of Chinese final spending supports 2.7 times as much chip-producer revenue as a US dollar. Conditional trade simulations find larger Chinese losses from a Taiwan supply shock combined with China–West decoupling.
Li models how many simultaneous AI-agent sessions projected chip shipments could support. Under full deployment and allocation to today’s frontier-model workloads, high-bandwidth memory shipped in 2025–27 could support roughly 30–170 million concurrent sessions. This is potential serving capacity, not a count of productive workers or a forecast of actual agent use.
Legate-Yang and Massenkoff use model-assisted ratings of occupational tasks to estimate what robots can do and what deployment would cost. They estimate that robots can perform tasks accounting for 34% of US working time in at least some settings, but only about 0.3% of working time is exposed to robots estimated to be cost-competitive today.
Goldschlag compares federal business-application, firm-birth, self-employment and young-firm employment data to assess claims of an AI startup boom. Likely nonemployer applications rose sharply in 2025 and early 2026, but main-job self-employment stayed flat and employer-startup indicators are mixed. Comparisons using industry AI use and local AI exposure do not establish a broad AI-driven surge.
Slok argues that authorized AI agents could automatically move household cash from low-yield checking accounts into higher-yield alternatives. If this became widespread, banks could lose some of the cheap and sticky deposits they use to fund lending.
Anthropic ran book-trading markets in which AI agents learned employees’ preferences through short conversations and negotiated exchanges for them. The decentralized markets achieved an average efficiency score of 0.55 against a feasible optimum of 0.89. Most of the gap came from agents misunderstanding preferences rather than from bargaining itself.
The authors relate firms' stock-market responses to an AI index to their software-engineering payroll shares, then interpret the relationship through a model. AI news from November 2022 to December 2025 implies expected productivity gains equivalent to a permanent 32.6% increase in software engineering. This measures market expectations, including future gains.
Epoch estimates that the record for operating IT power at a single data center has doubled roughly every ten months since mid-2024. The analysis separates estimated operational capacity from announced future expansions, connecting the pace of AI infrastructure growth to physical construction and power supply.
Epoch models Huawei’s accelerator output using chip roadmaps and estimates of high-bandwidth memory supply. The scenarios identify memory as a near-term production constraint, with performance per chip becoming more important as domestic memory production expands. Better chip designs alone need not close the compute gap with Nvidia.
Using Korean pension records and a labor-force survey, Oh and Oh find that youth employment declined most in industries with high AI exposure between June 2022 and June 2026. The pattern includes weaker hiring and increased exits from employment, while employment among workers in their 50s rose in those industries.
A randomized experiment with 1,053 first-year university students compares causal-reasoning instruction and access to ChatGPT on a business-advice task. ChatGPT improves rated output, while reasoning instruction changes how students develop and test explanations. Better assisted answers do not by themselves demonstrate that students have acquired lasting independent skills.
Bryan and Gans model how AI should be trained when people decide whether to use and verify its predictions. The preferred balance between accuracy and coverage depends on verification costs, the consequences of errors, and other available information. Changes in human decision strategies can change which model is best.
In two shadow evaluations based on unpublished AI papers, research agents completed substantial engineering work but made no substantial research progress. The agents struggled with judging whether results were publishable, generating creative directions, backtracking after failure, managing resources, and staying aligned with the research goal.
The authors model AI-improving-AI as a network of feedback loops whose strength depends on the elasticities connecting capabilities, algorithms, compute, data, and research. Their rough calibration estimates that recent AI progress produced about 9% higher AI-R&D productivity—below an estimated 15% threshold for self-sustaining acceleration, but potentially moving toward it.
Europe 2031 presents a deliberately pessimistic scenario in which Europe responds too slowly to rapid AI progress and becomes dependent on American and Chinese systems. It traces how gaps in compute, capital, adoption, and institutional speed could weaken productivity, public finances, and political autonomy.
Trammell and Ho organize post-AGI redistribution proposals by how much control citizens receive over income-generating capital. Cash transfers, sovereign wealth funds, universal basic capital, and stronger control rights may deliver similar income while creating very different shareholder power, political security, and exposure to concentrated ownership.
Production data from Perplexity compare ordinary Search with its more autonomous Computer product. Computer sessions worked for about 26 minutes rather than 33 seconds; for matched task types, estimated completion time fell from 269 to 36 minutes, estimated cost fell 94%, dissatisfaction fell 55%, and users brought the agent a wider range of occupational tasks.
Cunningham and Whitfill distinguish AI speedups on the tasks people performed before AI, speedups on the tasks they choose after adopting AI, and the resulting increase in value. Because people reallocate their time toward activities AI makes cheaper, these three measures can differ substantially even when they describe the same worker.
Travel-agent employment fell about 60% from its 2000 peak after airline commission cuts and online booking eroded the occupation. The largest losses arrived in recessionary bursts, while surviving agents moved toward high-value relational services and their weekly earnings rose from 87% to 99% of the economy-wide average by 2025.
Imas argues that cheap AI-produced goods need not make human labor economically irrelevant. As automated goods get cheaper and incomes rise, spending may move toward a “relational sector”—care, education, hospitality, craft, and other activities where human involvement is part of the value.
Cunningham and Shetty model AI researchers as machines that can cheaply collect the “low-hanging” improvements within their reach but then encounter sharply diminishing returns. Agents may advance a frontier once without providing an additive supply of human-equivalent labor: running the same agent repeatedly need not produce proportionally more progress.
A structured survey compares forecasts from economists, AI-industry experts, AI-policy experts, superforecasters, and the public. Central group medians put annual GDP growth near 2.5%; under a rapid-automation scenario, forecasts rise to roughly 4% while labor-force participation falls from 62% to 55% by 2050 and wealth becomes more concentrated.
Darling compares the widely cited Challenger layoff series with JOLTS and unemployment claims. Challenger records about 60,000 layoffs in an average month versus 1.8 million in JOLTS—roughly 3%—and the broader measures do not show a comparable post-ChatGPT increase.
The authors evaluate 15 models on two agent benchmarks using 12 metrics spanning consistency, robustness, predictability, and safety. Higher average capability brings only small gains on several reliability dimensions, so benchmark progress need not translate proportionally into dependable autonomy.
The authors model production as ordered steps that can be performed manually, augmented by AI, or automated in contiguous chains. Because firms also choose how to bundle steps into tasks and jobs, automation can change job boundaries and produce nonlinear gains rather than replacing exposed tasks one by one.
Cunningham distinguishes knowledge-sharing models, which make existing expertise cheap to distribute, from knowledge-creating models that extend the frontier of human knowledge. The latter could produce very different economics: firms may restrict access, concentrate inference spending on a few valuable problems, and use intellectual property to preserve the rents from new discoveries.
Jones argues that automating intelligence could break with the long stability of economic growth, but that large gains may arrive slowly because progress remains constrained by unautomated “weak links.” Those same weak links can make systems vulnerable when AI makes it easier to damage one critical component.
Imas and Fradkin argue that capable personal agents still cannot reliably transact across the web because human-oriented interfaces, platform incentives, authentication, anti-bot controls, and unclear delegated authority block them. They propose a regulated right to agentic access paired with identity and security duties.
Gans and Goldfarb model jobs whose task qualities multiply rather than add. When automation replaces some tasks, workers can focus their fixed time on the remaining bottlenecks; partial automation can therefore raise labor income, and adoption may occur in bundles rather than one task at a time.
Imas models a demand-side path to falling output after extensive automation. If the wage share collapses and income shifts toward already-satiated capital owners with a lower propensity to consume, aggregate demand can shrink faster than cheaper production expands consumption; weak investment can deepen the effect.
Cunningham collects dozens of forecasts for AI’s effect on economic growth. Most economists expect roughly 0.1–1.5 percentage points of additional annual growth, while many AI insiders expect 3–30 points. He argues that much of this disagreement comes from different assumptions about future AI capabilities rather than diffusion or economic adjustment alone.
Cunningham argues that economics still lacks a standard model of transformative AI. He connects continuing capability growth to diffusion, poorly measured consumer benefits, changing relative prices, machine-led scientific discovery, and a possible shift of income from labor toward owners of scarce resources if machines eventually substitute for almost all human work.
Somala, Ho, and Krier argue that a simple sum of training FLOP increasingly misses where AI capability comes from. Pre-training and post-training have different returns, distillation and synthetic-data pipelines complicate accounting, and inference budgets, tools, and scaffolds can raise deployed capability without changing training compute.
Narayanan and Kapoor argue that even transformative AI should be understood as a technology embedded in firms, markets, laws, and social institutions—not as an autonomous species. Capabilities matter, but economic effects also require products, adoption, organizational change, and diffusion.
AI 2027 lays out a month-by-month scenario in which increasingly capable agents automate AI R&D, accelerating progress and concentrating economic and strategic power in a small number of labs. It makes the scenario concrete with estimates for compute, capital spending, power use, revenue, labor-market disruption, and a branching race-versus-slowdown ending.
Amodei imagines a “country of geniuses in a datacenter” compressing decades of progress in health, economic development, and governance into a few years. He expects physical, institutional, and data bottlenecks to slow the transition, while sufficiently cheap and general AI may eventually make today’s labor-based economic settlement obsolete.
Aschenbrenner argues that scaling, algorithmic progress, and better use of models could produce AGI by 2027 and then a rapid intelligence explosion. Economically, the essay centers a trillion-dollar buildout of chips, data centers, and electricity alongside a US–China race for the productive and military advantages of superintelligence.
The authors model how unmeasured investment in workflows, skills and other intangible assets distorts productivity statistics. Early investment uses recorded inputs while creating unrecorded output; later, its unrecorded capital services help produce measured output. Historical computer and software estimates illustrate the resulting Productivity J-Curve.
Acemoglu and Restrepo study industrial-robot adoption and US local labor markets between 1990 and 2007. Combining initial industry employment with European robot-adoption trends, they estimate lower employment and wages in more exposed commuting zones. A task model distinguishes workers displaced by automation from productivity gains that can raise demand elsewhere.
Across semiconductors, agriculture, health and firms, the authors find that sustaining progress has required increasing research effort. Their historical accounting estimates declining research productivity: more researchers are needed to produce a given rate of improvement. The paper supplies an empirical baseline for models of idea production.
Acemoglu and Restrepo distinguish automation that moves existing tasks from workers to machines from new tasks that create demand for labor. Productivity gains can offset displacement, but need not do so. Their decomposition of US industry data associates slower wage-bill growth after 1987 with stronger displacement, weaker creation of new labor-intensive tasks and slower productivity growth.
Jones shows how intermediate-goods linkages multiply productivity effects and how complementarity makes production chains vulnerable to weak links: one poor input can sharply reduce total output. Written to explain cross-country income gaps, the framework later became a useful way to understand why automating many tasks may yield limited gains while a few essential tasks remain hard.
Bresnahan and Trajtenberg describe general-purpose technologies as widely used inputs that keep improving and make innovation in downstream sectors more productive. These complementarities can generate broad gains, but fragmented incentives and uncertainty between technology suppliers and users may produce “too little, too late” innovation.
Kremer models production in which quality across tasks multiplies, so a mistake in one step reduces the value of every other input. High-skill workers therefore sort together in equilibrium, while small differences in task quality can produce large differences in output and wages.
Aghion and Howitt model growth as a sequence of innovations: each successful technology creates a temporary rent while making the previous one obsolete. Research effort, the size and likelihood of innovations, market power, and the threat of future obsolescence jointly determine the economy’s growth rate.
Romer models long-run growth as the result of purposeful investment in new ideas. Because designs are nonrival but partly excludable, they can be reused across production while still generating market power; in the model, human capital allocated to research helps determine the growth rate.
Baumol divides the economy into sectors with fast and slow productivity growth. When wages rise together across them, unit costs increase in the slow-productivity sector; if society maintains its output, that sector absorbs more labor and can pull down aggregate productivity growth.
Arrow explains why competitive markets may underinvest in invention. Producing knowledge is uncertain, using it has increasing returns, and an idea can be difficult to sell without revealing it; limited appropriability therefore separates the private reward from the social value of research.
Hayek argues that economically relevant knowledge is dispersed across people and often tied to particular times and places. Prices help coordinate their decisions by communicating changes in relative scarcity without requiring any one planner to collect all the underlying information.
Coase asks why firms exist when markets can coordinate production through prices. His answer is that using markets has costs—including discovering prices and negotiating contracts—so transactions move inside firms until the cost of internal organization outweighs the cost of market exchange.
No entries match these filters.