Brief #14: The AI labs measure how AI impacts work and suggest solutions
Anthropic drafts the safety net for AI displacement, while OpenAI measures the shift from chatting to delegating
Welcome! This bi-weekly newsletter, published by the Windfall Trust, curates the most important developments in AI economics research and policy.
Need to Know
Anthropic proposes a three-tier framework for AI-driven displacement: capital accounts and wage insurance now, expanded unemployment insurance at recession-level joblessness, and UBI-style payments plus new tax bases only under unprecedented unemployment levels.
OpenAI analyzes token usage among individuals, businesses, and OpenAI employees. More senior and more intensive adopters mostly use the coding agent, while individual users still mostly stay on the chatbot.
AI 2040’s Plan A projects 50% growth in 2032 under a US-China compute cap-and-trade. Critics dispute its assumptions about R&D automation returns and diffusion speed, while warning that a slowdown could make any later breakdown more explosive
In the news
The Stanford Digital Economy Lab coordinated a group of nearly 200 economists and tech leaders, including 16 Nobel Prize Laureates, to warn about the economic risks and opportunities posed by AI, including job displacement and gains in living standards. The New York Times reports that even prominent AI skeptics like Daron Acemoglu have joined this letter.
The corporate expense manager, Ramp, and the workforce data provider Revelio Labs find that high-intensity AI adopters grow overall and early-career headcount by 10%-12% compared to similar firms that are slower adopters. However, this might be due to growing market share rather than overall employment growth. In a response to this piece, Guy Lichtinger and Seyed Hosseini (Harvard) suggest explanations for the correlation, arguing that there is not enough evidence to distinguish between “AI leads to hiring” and “growing firms start spending on AI alongside many other investments.”
The AI Futures Project published an optimistic scenario for AI by 2040, assuming the US and China follow their Plan A to verifiably slow down AI development. Their Plan A is a verified slowdown that consists of an international cap-and-trade system to restrict compute access. It makes all AI research transparent, limits algorithmic progress, and allows for mutually assured compute destruction. AI is deployed broadly outside of the riskiest domains such as AI R&D. In their accompanying piece on the economics of Plan A, the authors lay out how their cap-and-trade compute system captures a large share of AI surplus. A core goal of Plan A is to avoid recursive self-improvement. They assume that AI will soon be able to replace human workers in R&D and that inherently human or relational tasks only contribute negligibly to output.
These aggressive modeling choices, which generate 50% growth in 2032, assume away most economic frictions. A timely calibration by Tom Cunningham (METR) et al. investigates the possibility of R&D automation: by modeling the feedback loops behind recursive self-improvement, they find that current returns to AI capabilities in R&D fall short of the threshold for self-sustaining acceleration, though returns are rising. Their model also implies that automating R&D doesn’t necessarily have a big economic impact if it only speeds up narrow capabilities like AI R&D instead of broader capabilities that could automate the full economy.
Meanwhile, Tom Davidson (Forethought) worries that restraining algorithms while scaling compute builds a “dry tinder” overhang, making any intelligence explosion far faster should the deal collapse. Seb Krier (DeepMind) notes that a multi-year plan with centralized authority over compute access and research directions is too heavy-handed for experimentation. He is skeptical of the rapid diffusion of AI, the 50% growth in 2032, and the feasibility of increasing welfare-state spending solely through data-center permit revenues. Jaime Sevilla (Epoch), by contrast, approvingly calls it the most accelerationist plan yet spelled out in detail. Discussing and measuring the assumptions behind Plan A allows us to gauge early how transformative AI will be.
In detail
Anthropic suggests how the US Government should handle AI-enabled labor displacement
Anthropic laid out its Economic Policy Framework, which focuses on AI-enabled automation of the US labor market across three scenarios with different levels of unemployment. The measures rely on prior improvements in economic measurement, a dedicated government analytical unit, and modernized unemployment insurance systems that deliver more easily and quickly. If displacement outpaces the economy’s capacity to adjust, Anthropic conditionally supports a slowdown of AI development, provided it applies uniformly to all firms.
The first tier of AI automation looks similar to today, with unemployment around 5%, and requires policymakers to prepare for disruption and ensure broad participation in AI’s upside. Therefore, Anthropic’s main proposal is universal capital accounts. Children, young adults, and exposed workers can withdraw from their accounts during career transitions and participate in AI equity. Additional recommended measures include wage insurance, retention tax incentives, and workforce training grants.
If AI continues to displace workers and create recession-level unemployment of 10%, retraining and expanded temporary income support become necessary. Automatic, uniform, and longer-term unemployment insurance is Anthropic’s core policy proposal in this scenario, supported by sector-specific transition support.
If unemployment exceeds historical peaks, Anthropic suggests building new redistribution mechanisms to sustain the incomes of many unemployed people. Over time, payments should be less linked to previous wages and more towards a universal basic income. To sustain redistribution amid an AI productivity boom, taxes need to shift from labor to consumption, capital gains, and levies on AI use. Additionally, public investments in human-facing work keep jobs available for those who want them. Anthropic also points out that AI companies should pay their fair share if they automate the workforce while generating outsized returns.
What this means: This framework combines reactive unemployment benefits with proactive sharing of AI’s upside via universal capital accounts that are more targeted to AI than the similar Trump Accounts: eligibility extends to exposed workers, withdrawals are permitted during career transitions, and the asset mix includes AI equity. Temporary income support may also dampen the political backlash, while capital accounts allow the median voter to benefit from AI as a stakeholder.
In the end, it all comes down to the government’s checkbook to clean up the mess of AI automation, even though Anthropic states it wants to contribute its fair share and considers AI usage taxes at a later time. Sharing equity with the public at a potential IPO would be very informative regarding Anthropic’s intent of this document as credible commitment or cheap talk, especially since Anthropic hasn’t promised this.
Hiring numbers and the labor force participation rate might be better measures of AI disruption than unemployment. Reduced hiring is an early warning indicator and would shift focus away from defending incumbents via retention tax credits. In such scenarios, many of those affected would either drop out of the workforce, never enter it altogether, or become underemployed. This would make the unemployment rate a lagged lower bound for those affected.
Work shifts from chatbots to coding agents
Delegation appears to depend more on expertise than on age
OpenAI analyzes the shift to agentic AI among employees, businesses, and individuals. OpenAI employees delegate more tasks to the coding agent, Codex, across a range of tasks rather than consulting the chatbot. The authors classify each user’s persona, job title, task category, and estimated task complexity and primarily use tokens as the output measure.
Adoption is rapid, but uneven. While 98% of OpenAI employees used Codex in the last month, this drops to 17% and 1% for organizational and individual users, respectively. Among external organizations, engineers and analysts lead, while HR and legal lag; within OpenAI, this gap has closed.
Intensive users delegate work via concurrent agents and reusable skills. The tasks they delegate are growing in estimated complexity - e.g., the share of individual users submitting at least one request estimated to require at least 8 hours of human work rose from 2% to 26% in six months. The most intense OpenAI users accumulate 71 hours of agent runtime per day, summed across concurrent agents. According to the authors, this internal use previews what diffusion will look like once workflows are restructured. Overall, Codex is used for ever longer and more varied tasks.
What this means: OpenAI observes relevant facts about AI usage and changing workflows, but tokens are the wrong units of measurement. Coding agents use more tokens per request, and token growth doesn’t imply that more work is shipped, as discussed in Brief #12.
OpenAI employees not only receive unlimited tokens and best practices from developers on Slack, but also probably work there because of their enthusiasm for AI. Other firms might not be able to recreate these favorable conditions for diffusion, even if OpenAI has now made Codex the standard desktop app.
Early-career workers use Codex the least, consistent with Anthropic’s finding that effective delegation is complementary to domain expertise - a result the OpenAI paper itself cites. Meanwhile, contractors use AI the most. It’s unclear whether this is due to tasks already packaged into well-specified, modular bundles for external handoff or to fewer status concerns when using AI for temporary workers.
In Other News
Labor Market & Employment
Goldman Sachs expects AI adoption to temporarily displace over 9% of the US workforce over a 10-year transition.
The US think tank Fathom published a framework for action under uncertainty across different scenarios of AI’s labor-market impact.
Arul Murugan (Berkeley) et al. show that AI exposure is lower in developing countries and among men, but that remittances increase exposure in poor countries with emigration.
Joseph Emmens (MIT) finds evidence that common ownership exacerbates automation.
AI Capabilities & Impact
Amelia Michael (FAI) argues that industrial robots will be more useful than humanoid robots under transformative AI if factories reorganize their workflows.
Karen Dynan (Harvard) et al. build on the Congressional Budget Office’s Budget Outlook to analyze four AI scenarios. The authors recommend robust, scalable “no regret” policies such as a limited modernized Trade Adjustment Assistance program and public equity purchases.
Lukas Freund (Boston) and Vasco Carvalho (Cambridge) develop a model of innovation in which firms must decide between exploiting current research and exploring new fields of research. If AI makes exploiting existing fields easier, it restores growth, but narrows the direction of innovation further.
Tom Cunningham (METR/ Elastic) et al. model the economics of recursive self-improvement as discussed above. The authors call for more data from AI labs on training, inference, and experimental compute, as well as staff numbers to better calibrate their model.
AI Adoption
MIT Future Tech found that AI adoption among S&P 500 firms quadrupled from 5% to 21% between 2022 and 2025. The authors don’t find effects on capex or productivity.
Nicola Borri (Luiss) et al. estimate firm-level comovement with an AI factor. There’s a large and heterogeneous AI premium.
According to SemiAnalysis, Anthropic will become profitable in the third quarter. More than 80% of the revenue stems from API calls, not subscriptions. Revenues have increased sevenfold so far this year. This data might justify a high valuation at a potential IPO and assuage fears of a sector-wide AI bubble.
Research Opportunities
Checks and Balances requests proposals from researchers on, e.g., economic power and benefit-sharing.
Benjamin Moll (LSE) published his slides on using Reinforcement Learning for economic research.
Your comparative advantage is commenting. Ours is overthinking AI and labor markets.
Thanks to Deric Cheng and Joel Christoph for contributing to this week’s edition of the newsletter.