On March 5, 2026, Anthropic published a research paper with an unassuming title: "Labor Market Impacts of AI: A New Measure and Early Evidence." Beneath the academic language lies something unprecedented — the first large-scale attempt to measure AI's real-world labor market footprint using actual production data from millions of conversations, rather than theoretical capability estimates or survey responses.
The paper is one output of Anthropic's Economic Futures program, a multidisciplinary initiative launched in 2025 that combines research grants, policy development, and longitudinal data infrastructure. Together with the Anthropic Economic Index — which tracks how Claude is used across hundreds of occupations — the program represents the most ambitious empirical effort to date to connect AI capabilities to labor market outcomes.
The central finding is deceptively simple: the gap between what AI can theoretically do and what it is actually doing in professional workflows is enormous. Computer programming has 94% theoretical exposure to AI, but only 33% observed exposure — meaning that despite the headlines about AI replacing programmers, actual automated usage in real professional settings covers roughly a third of the occupation's tasks. For the workforce at large, around 30% of workers have zero observed AI exposure.
This gap is the most important number in AI labor economics, and no one had measured it before.
The Measurement Problem: Why Every Previous Estimate Was Wrong
Before Anthropic's intervention, the field relied on two approaches, both with fundamental limitations.
The first approach uses theoretical capability assessments. Researchers take an occupational database like O*NET — which enumerates tasks for roughly 800 US occupations — and estimate which tasks an LLM could theoretically accelerate. The seminal work here is Eloundou et al. (2023), which found that LLMs could theoretically speed up 80% of US workers in at least 10% of their tasks. Goldman Sachs extrapolated this to estimate 300 million jobs globally exposed to AI automation.
The second approach uses labor market data — employment statistics, hiring rates, wage trends — to measure whether AI is actually affecting jobs. The Yale Budget Lab's October 2025 analysis found "no discernible disruption" in the broader labor market 33 months after ChatGPT's release. The Peterson Institute noted in March 2026 that "research on AI and the labor market is still in the first inning."
Both approaches have a critical blind spot. Theoretical estimates tell you what AI could do, but not what it is doing. Labor market data tells you what happened, but cannot attribute causation to AI specifically. No one was measuring the bridge between capability and outcome — the actual adoption patterns in real professional workflows.
Anthropic's insight was that they sit on top of the largest dataset of AI-assisted professional work ever assembled. Every conversation on Claude.ai and through their API contains evidence of how AI is being used, for what tasks, in which occupations, and with what degree of autonomy. The challenge was transforming this raw data into an economic measurement framework.
The Economic Futures Program: Three Pillars
The Anthropic Economic Futures program, announced in mid-2025 and expanded with a $10 million commitment later that year, operates through three interconnected pillars.
Pillar 1: Research Grants and Partnerships. The program offers grants between $10,000 and $50,000 for empirical research on AI's economic impacts, along with API credits and strategic partnerships with independent research institutions. The emphasis is on original empirical work using econometric methods, data analysis, or novel data collection — explicitly excluding theoretical modeling without empirical grounding.
Pillar 2: Evidence-Based Policy Development. Through Economic Futures Symposia — held in Washington, DC (September 2025) and London (November 2025) — the program brings together policymakers, researchers, and industry to evaluate concrete policy proposals. The approach is scenario-based: policy ideas are categorized by the severity of AI's labor market impact, ranging from "nearly all scenarios" (worker upskilling, infrastructure permitting reform) to "faster-moving scenarios" (sovereign wealth funds for AI revenue distribution, new taxation mechanisms).
Pillar 3: Economic Measurement and Data Infrastructure. This is where the Economic Index lives. Anthropic is building what it describes as "one of the first longitudinal datasets on AI's economic usage, diffusion, and impact." The Index uses Clio, a privacy-preserving analysis system, to classify conversations across occupational categories, task complexity levels, and usage patterns — all without exposing individual conversation content.
The three pillars are designed to be self-reinforcing. The data infrastructure produces empirical findings. The research grants fund independent verification and extension of those findings. The policy forums translate findings into actionable governance frameworks.
Observed Exposure: The Metric That Changes Everything
The program's most significant methodological contribution is a measure called observed exposure. It combines three data sources:
- O*NET task data: Enumerates the specific tasks associated with each of ~800 US occupations
- Claude usage data: From the Anthropic Economic Index, capturing what people actually use AI for
- Theoretical exposure estimates: From Eloundou et al. (2023), measuring which tasks LLMs could theoretically accelerate
The calculation works as follows. First, identify tasks that are theoretically feasible for LLMs. Then check which of those tasks actually appear in Claude usage data as work-related, automated (not merely augmentative) interactions. Weight automated usage fully and augmentative usage at half weight. Finally, aggregate to the occupation level, weighted by the time workers spend on each task.
The result is a single number per occupation: what fraction of its tasks are seeing real, automated AI usage in professional settings today.
This metric fills the gap between theoretical capability and labor market outcome. Theoretical exposure says "AI could do 94% of programming tasks." Observed exposure says "AI is currently doing about 33% of programming tasks in an automated way." The 61-point gap is where the entire labor market debate should focus — because that gap will close over time, and the rate at which it closes determines the pace of actual economic disruption.
What the data reveals: the exposure landscape
Anthropic's findings paint a more nuanced picture than either the utopian or dystopian narratives suggest.
| Occupation Category | Theoretical Exposure | Observed Exposure | Gap |
|---|---|---|---|
| Computer & Mathematical | ~94% | ~33% | 61 pts |
| Office & Administrative | ~85% | ~25% | 60 pts |
| Business & Financial Operations | ~78% | ~22% | 56 pts |
| Healthcare Practitioners | ~45% | ~8% | 37 pts |
| Construction & Extraction | ~15% | ~2% | 13 pts |
The gap is consistently 50 to 65 percentage points for knowledge-work occupations. This is not a small discrepancy. It means that every headline based on theoretical exposure — "AI threatens 80% of white-collar jobs" — overstates actual current impact by at least a factor of two.
Critically, workers with zero observed AI exposure represent approximately 30% of the workforce. These are predominantly physical occupations, service roles with high interpersonal requirements, and jobs where the work product cannot be digitized. The theoretical exposure literature often treats these workers as irrelevant to the AI story. Anthropic's data confirms they are — for now.
The demographic surprise
One finding that runs counter to popular assumptions: workers in the most exposed professions are more likely to be older, female, more educated, and higher-paid. This is because the occupations with the highest observed exposure — computer programming, financial analysis, customer service, technical writing — skew toward educated knowledge workers. The "AI threatens blue-collar men" narrative is precisely backwards. AI is currently most active in the highest-paid, most credentialed segments of the workforce.
This has policy implications that the Economic Futures program's symposia are designed to address. If AI disruption concentrates among educated, high-earning professionals — the demographic with the most political voice and institutional power — the policy response will look very different than if it primarily affects low-wage workers with less political leverage.
The Learning Curve Factor: Why Adoption Is Slower Than Capability
The Economic Index's March 2026 report, "Learning Curves," adds a crucial layer to the exposure data. It documents that high-tenure Claude users develop strategies and habits that make them significantly more effective than new users. Experienced users attempt higher-value tasks, achieve higher success rates, and use Claude for a wider range of work purposes.
This finding has a direct implication for the observed exposure gap. The gap between theoretical and observed exposure is partly a capability gap — AI cannot yet do everything theoretically possible. But it is also an adoption gap — even when AI can do something, professionals need time to learn how to integrate it into their workflows effectively.
The learning curve evidence suggests that observed exposure will increase not just because AI gets more capable, but because users get more skilled at deploying existing capabilities. This is the "cognitive amplifier" effect in action: AI's impact is mediated by the skill of the human operator, and that skill develops through sustained use over time.
Anthropic's data shows a specific channel through which skill-biased technological change may unfold. Early adopters with high-skill tasks have more successful interactions with AI than later, less technical adopters. These early-adopting users may be simultaneously the most exposed to AI-driven disruption and the most aided by AI in these initial, augmentative waves of adoption. The gap between the AI-native and the AI-illiterate may widen before it narrows.
The Hiring Signal: What the Data Does NOT Show
Perhaps the most carefully reported finding in Anthropic's labor market paper is what it does not find. The paper finds no systematic increase in unemployment for highly exposed workers since late 2022. This has been widely reported as "AI is not causing job losses" — a simplification that Anthropic's own researchers warn against.
What the data does show is suggestive evidence that hiring of younger workers (aged 22-25) has slowed in the most exposed occupations. This is a fundamentally different signal than mass unemployment. It suggests that AI's current impact is not on existing workers but on the pipeline — employers may be filling new positions with AI-augmented workflows rather than hiring junior staff, or they may be uncertain enough about AI's trajectory that they are slowing entry-level hiring while they wait for clarity.
Forbes' Hamilton Mann made a sharp critique of the "no unemployment" framing in March 2026: "AI might reduce hiring, slow promotions, reduce junior roles, or compress wages, without causing big layoffs. So concluding 'no unemployment effect' from current data is premature." The critique is valid, and Anthropic's paper explicitly acknowledges it. The observed exposure metric is designed to be a leading indicator — a tool for tracking the gap between capability and adoption as it narrows over time — not a definitive statement about current employment effects.
The Fortune interview with Anthropic's lead economist Peter McCrory in April 2026 crystallized the core insight: coding has something like 94% theoretical exposure, but based on actual adoption, "it was closer to 30% of the tasks across all the jobs in that pocket of the economy." The gap between those numbers — and the rate at which it closes — is what policymakers should be watching.
The Economic Index Survey: Extending Measurement Beyond Usage Data
In April 2026, Anthropic announced the Economic Index Survey, a monthly survey conducted through Anthropic Interviewer. The survey is designed to capture first-hand accounts of workplace change that may not yet appear in aggregate labor market data.
Combined with Claude usage data in a privacy-preserving way, the survey creates a two-channel measurement system. Usage data reveals what people are actually doing with AI. Survey data reveals how they perceive those changes affecting their work, their careers, and their economic prospects.
This addresses a fundamental limitation of pure usage data: it cannot capture displacement effects that happen outside the AI platform. If a company reduces its marketing team from 10 to 7 people because AI handles basic content generation, that displacement doesn't show up in Claude usage data. The three remaining marketers may be using Claude more effectively than ever, while the displaced workers' experiences are invisible to the platform. The survey is designed to capture these missing signals.
The Five Economic Primitives: A Measurement Framework
The January 2026 Economic Index report introduced five "economic primitives" — foundational measurements for tracking AI's economic impact over time:
| Primitive | What It Measures |
|---|---|
| Task Complexity | How challenging the tasks brought to Claude are, on a defined scale |
| Skill Level | The expertise required for the task, correlated with wage levels |
| Purpose | Whether the conversation is for work, education, or personal use |
| AI Autonomy | The degree to which the user delegates versus collaborates |
| Success | Claude's assessment of whether the conversation achieved its goal |
These primitives enable longitudinal tracking that goes beyond raw usage statistics. For example, if AI autonomy is increasing over time within a specific occupation, that signals a shift from augmentation toward automation — a leading indicator of potential displacement. If success rates are rising for complex tasks but not for simple ones, that suggests the model is improving on cognitively demanding work while basic tasks may already be saturated.
The primitives also enable cross-occupational comparisons that were previously impossible. Anthropic's data shows that Claude's task-level success rates vary significantly by occupation — and that these variations correlate with wage levels and education requirements. This creates a direct measurement channel between AI capability and labor market stratification.
Why This Matters: Data Over Narratives
Anthropic's Economic Futures program matters for three reasons that go beyond its specific findings.
First, it establishes a methodology that can be replicated. The observed exposure metric combines publicly available occupational data (O*NET), theoretical capability estimates, and proprietary usage data. While the proprietary component limits independent replication, the framework itself — measuring the gap between theoretical and actual AI adoption at the task level — is a methodological contribution that other researchers can adapt with different data sources.
Second, it provides longitudinal tracking infrastructure. The Economic Index is designed as a recurring measurement, not a one-time study. This means that the observed exposure metric will be updated over time, allowing researchers to track how quickly the gap between theoretical and observed exposure narrows. The rate of gap closure is a more policy-relevant signal than any single snapshot.
Third, it reframes the AI labor market debate from speculative to empirical. The current public discourse oscillates between "AI will replace all jobs" and "AI is creating more jobs than it destroys." Both claims are based on extrapolation and narrative. Anthropic's contribution is not to settle the debate — the data is too early for that — but to provide measurement tools that make the debate answerable over time.
As Peter McCrory told Fortune: "What we have in our data is how that theoretical ability of these models meets the real world, and by tracking it over time, we can have a sense of how the gap between theoretical exposure and actual adoption is taking place."
FAQ
What is the Anthropic Economic Futures program?
A multidisciplinary research initiative launched by Anthropic in 2025, funded with a $10 million commitment, that supports empirical research on AI's economic impacts, develops evidence-based policy proposals, and maintains longitudinal data infrastructure through the Anthropic Economic Index.
What is "observed exposure"?
Anthropic's novel metric that measures what fraction of an occupation's tasks are seeing real, automated AI usage in professional settings — as opposed to theoretical estimates of what AI could potentially do. It combines O*NET task data, theoretical capability estimates, and actual Claude usage data.
Has AI caused unemployment yet?
Anthropic's data finds no systematic increase in unemployment for highly exposed workers since late 2022. However, there is suggestive evidence that hiring of young workers (22-25) has slowed in the most exposed occupations. Researchers caution against interpreting "no unemployment yet" as "no impact ever."
What is the gap between theoretical and observed AI exposure?
For knowledge-work occupations, the gap is consistently 50 to 65 percentage points. Computer programming, for example, has ~94% theoretical exposure but only ~33% observed exposure. This gap is the critical metric for understanding the pace of actual AI adoption in the workforce.
Who is most affected by AI in the labor market?
Anthropic's data shows that the most exposed workers are older, female, more educated, and higher-paid — because the occupations with the highest observed exposure (programming, financial analysis, technical writing) skew toward educated knowledge workers. This contradicts the popular narrative that AI primarily threatens blue-collar workers.
What policy responses does the Economic Futures program recommend?
The program does not advocate specific policies. Instead, it organizes policy proposals by scenario severity: modest impact (upskilling, infrastructure reform), moderate acceleration (fiscal support for displaced workers, automation taxes), and rapid transformation (sovereign wealth funds, new revenue mechanisms). The goal is to prepare options for a range of possible futures.
How is the Economic Index data collected?
Through Clio, Anthropic's privacy-preserving analysis system, which classifies conversations on Claude.ai and the API into occupational categories and task types without exposing individual conversation content. The system analyzes millions of anonymized conversations to map AI usage patterns across the economy.
References
- Anthropic. "Labor Market Impacts of AI: A New Measure and Early Evidence." March 5, 2026. https://www.anthropic.com/research/labor-market-impacts
- Anthropic. "Introducing the Anthropic Economic Futures Program." 2025. https://www.anthropic.com/news/introducing-the-anthropic-economic-futures-program
- Anthropic. "Anthropic Economic Futures Program." https://www.anthropic.com/economic-futures/program
- Anthropic. "Anthropic Economic Index Report: Learning Curves." March 24, 2026. https://www.anthropic.com/research/economic-index-march-2026-report
- Anthropic. "Anthropic Economic Index: New Building Blocks for Understanding AI Use." January 15, 2026. https://www.anthropic.com/research/economic-index-primitives
- Anthropic. "Announcing the Anthropic Economic Index Survey." April 22, 2026. https://www.anthropic.com/research/economic-index-survey-announcement
- Anthropic. "Preparing for AI's Economic Impact: Exploring Policy Responses." https://www.anthropic.com/research/economic-policy-responses
- Fortune. "Anthropic's Research Shows That AI Can Already Do a Huge Portion of Many Jobs." April 7, 2026. https://fortune.com/2026/04/07/anthropic-peter-mccrory-ai-automation-white-collar-jobs-claude-recession/
- Forbes. "Anthropic's Study Does Not Measure AI's Labor-Market Impacts." March 8, 2026. https://www.forbes.com/sites/hamiltonmann/2026/03/08/anthropics-study-does-not-measure-ais-labor-market-impacts/
- Yale Budget Lab. "Evaluating the Impact of AI on the Labor Market." October 2025. https://budgetlab.yale.edu/research/evaluating-impact-ai-labor-market-current-state-affairs
- Peterson Institute for International Economics. "Research on AI and the Labor Market Is Still in the First Inning." March 2026. https://www.piie.com/blogs/realtime-economics/2026/research-ai-and-labor-market-still-first-inning
- Eloundou, T., et al. (2023). "GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models."