CtrlK
The Anthropic Economic Index is a research initiative run by Anthropic that tracks how its AI assistant Claude is used across occupations, tasks, and geographies in order to study AI's economic effects. The project was launched on February 10, 2025, and has since produced a series of reports based on privacy-preserving analysis of Claude usage data, an open dataset hosted on Hugging Face, and (from 2026) a recurring user survey.[1][2][3]
The Index is not a benchmark in the usual model-evaluation sense. It is an economic measurement program that documents where AI is being adopted, which tasks are being augmented or automated, how usage shifts as new models ship, and how adoption varies across U.S. states and countries. The flagship reports map each conversation to a task in the U.S. Department of Labor's O*NET occupational taxonomy, then aggregate the results to draw labor market inferences.[1][2][4]
Because Anthropic is also the seller of the product being measured, the Index sits in a peculiar middle position between marketing and labor economics research. The reports are produced by Anthropic's Economic Research team, often with named external collaborators, and are accompanied by open data releases under a CC-BY license.[5]
When large language model products like ChatGPT and Claude reached mass adoption in 2022 and 2023, most published estimates of AI economic impact on the labor market were exposure scores rather than usage measurements. The widely cited Eloundou et al. paper from OpenAI and the University of Pennsylvania, "GPTs are GPTs" (2023), estimated that roughly 80% of the U.S. workforce could have at least 10% of work tasks affected by LLMs and that about 19% of workers might see at least 50% of tasks affected, but it relied on human and model annotations of O*NET tasks rather than observed usage.[6]
Anthropicâs Economic Index was framed as the inverse approach. Instead of asking "which tasks could be exposed," it asks "which tasks are people actually bringing to Claude today?" Anthropic explicitly positioned the Index as a way to ground policy debate in observed adoption rather than speculation.[1][2]
The choice of ONET as the underlying taxonomy matters. ONET is the U.S. Department of Labor's database of about 17,000 detailed work activities mapped to roughly 900 occupations. Using it as a backbone lets Anthropic compare AI-assisted tasks to a standard labor classification that economists, the Bureau of Labor Statistics, and other AI researchers already use.[1][2]
Several contemporaneous developments made a measurement program like this practically useful. Policymakers debating AI regulation needed empirical grounding beyond theoretical exposure studies. The question of whether AI would primarily augment workers or automate them away was contested but largely unresolved by observed data. And the rapid pace of model releases meant that static exposure analyses became obsolete quickly, while a longitudinal tracking effort could at least document the trajectory of adoption over time.[1][4]
The core measurement engine is Clio, short for Claude insights and observations, an automated conversation analysis system Anthropic introduced in December 2024 (arXiv:2412.13678). Clio applies four steps to anonymized conversations: it extracts facets such as topic and language, clusters semantically similar conversations, generates cluster-level descriptions stripped of identifying details, and organizes those clusters into a hierarchy. Every step is performed by Claude itself rather than by human reviewers, which is the project's central privacy claim.[7]
Clio's privacy protections are described as defense in depth. Data is automatically anonymized and aggregated; the model is instructed to drop private details when summarizing; minimum-frequency thresholds prevent rare clusters from being surfaced (later reports specify a floor of 15 conversations from at least 5 unique accounts); and Claude verifies that final summaries contain no identifying information before they are exposed to analysts. The system was used internally for safety monitoring before being repurposed as the engine for the Economic Index.[7]
The Clio architecture raises a structural question: the system uses Claude to analyze conversations with Claude, so the model's own tendencies, blind spots, and categorization habits propagate into the measurement results. Anthropic acknowledges this in passing but does not publish detailed studies of how Clio's classifications compare to human raters across occupation types.
For each conversation, Clio's hierarchy is mapped to the closest matching ONET task. The first paper that accompanied the launch, "Which Economic Tasks Are Performed with AI? Evidence from Millions of Claude Conversations" (Handa, Tamkin, McCain et al., arXiv:2503.04761), described the approach in detail and used roughly four million Claude.ai conversations as its base. The March 2025 follow-up complemented this top-down ONET classification with a bottom-up taxonomy of about 630 granular usage clusters generated directly from the data.[2][8]
Later reports extended the methodology in several directions. The September 2025 report added first-party API data alongside Claude.ai consumer data. The January 2026 report introduced five additional measurement dimensions, called "economic primitives," covering task complexity, human and AI skill levels, use case (work, educational, personal), AI autonomy on a 1 to 5 scale, and Claude's self-assessment of task success.[9][10]
The augmentation versus automation classification deserves special attention because it appears in nearly every summary of the Index. Anthropic defines augmentation as conversations in which the user iterates with Claude as a thinking partner, refining and improving work interactively. Automation (also called directive use) describes conversations in which the user delegates a complete task with little back-and-forth. The classification is made by Claude itself based on conversational structure, not by the occupational category, which means the same occupation can appear on both sides depending on how an individual user approaches a task.
A March 2026 companion paper by Anthropic economists Maxim Massenkoff and Peter McCrory introduced a related but distinct concept called "observed exposure," which combines O*NET occupational task data with Anthropic's actual Claude usage patterns and Eloundou et al.'s theoretical LLM capability assessments. Their measure weights automated (directive) uses more heavily than augmentative ones and filters for work-related contexts, producing an occupation-level score that sits between pure theoretical capability and raw usage share. This framework was used to study whether higher AI exposure was associated with observable labor market shifts.[13]
The same paper found that, as of early 2026, computer programmers showed the highest observed exposure (75% coverage), followed by customer service representatives and data entry keyers. Importantly, it also found that theoretical coverage (about 94% of tasks in Eloundou et al.) substantially exceeded actual usage (about 33%), confirming a large gap between what AI could do and what users were actually asking it to do.[13]
The Economic Index is produced by Anthropic's Economic Research team. Across the six reports published through mid-2026, the recurring lead authors are Maxim Massenkoff (Anthropic economist), Peter McCrory, Ruth Appel, and Eva Lyubich. Earlier reports listed Alex Tamkin and Miles McCain as lead authors. Karan Handa, Alex Tamkin, and Miles McCain led the companion arXiv paper that accompanied the launch.
Anthropicâs research team acknowledges external contributors in each paper. The January 2026 report listed over 30 acknowledgments including Dario Amodei, Jared Kaplan, and Jack Clark alongside economists Anton Korinek and John Horton, both of whom subsequently joined the Economic Advisory Council.
| Date | Report | Main focus |
|---|---|---|
| February 10, 2025 | "Introducing the Anthropic Economic Index" | Launch report; first occupational mapping of about 1 million Claude.ai conversations |
| March 27, 2025 | "Insights from Claude 3.7 Sonnet" | Usage shifts after Claude 3.7 Sonnet, extended thinking adoption, 630-cluster bottom-up taxonomy |
| September 15, 2025 | "Uneven geographic and enterprise AI adoption" | Per-capita usage by country and state, first-party API data, automation overtakes augmentation |
| January 15, 2026 | "Economic primitives" | Five new measurement dimensions, productivity estimates, success rates |
| March 24, 2026 | "Learning curves" | Task diversification, model selection by task wage, evidence on learning by doing |
| April 22, 2026 | Anthropic Economic Index Survey launch and "What 81,000 people told us about the economics of AI" | Monthly user survey, qualitative perceptions of productivity and displacement |
The initial release analyzed roughly one million anonymized Claude.ai conversations. Computer and mathematical occupations dominated, accounting for 37.2% of queries, with arts, design, entertainment, sports, and media in second place at 10.3%. Physical occupations such as farming, fishing, and forestry registered only about 0.1% of conversations.[1][8]
At the task level, Anthropic reported that about 36% of occupations had AI use in at least 25% of their associated O*NET tasks, while only about 4% had AI use across 75% or more of their tasks. The implication, broadly compatible with the Eloundou exposure estimates, is that AI assistance is touching most occupations in some way, but reaching deep into very few.[1]
The launch report was also the first to publish Anthropic's augmentation versus automation breakdown, defined at the conversation level. About 57% of conversations were classified as augmentation (where the user iterated with Claude as a thinking partner) and about 43% as automation (where the user delegated a task with little back-and-forth). Within the augmentation category, roughly 31.3% of conversations were classified as task iteration, 23.3% as learning interactions, and 2.8% as validation. Adoption was concentrated in mid-to-high wage occupations such as software developers and copywriters, with both very low-wage manual jobs and the very highest-paid professional roles showing minimal AI integration.[1][8]
The report was clear that this wage distribution does not imply AI is leveling the playing field between income groups. Mid-range professional occupations are heavier users than either end of the wage spectrum, but that finding reflects who is using Claude, not necessarily who benefits most in productivity terms.
Notable occupations by usage included copywriters and editors (highest task iteration at about 58% iterative conversations), translators and interpreters (most directive approach), and librarians (heaviest learning interactions, about 56% of conversations). Computer and mathematical workers accounted for 37.2% of Claude conversations despite representing only 3.4% of the U.S. workforce.
The second report covered roughly one million Claude.ai Free and Pro conversations sampled after the launch of Claude 3.7 Sonnet. Headline findings included a small overall increase in coding usage (about 3 percentage points for computer and mathematical occupations) and a larger jump in learning interactions, which rose from roughly 23% to 28% of conversations.[2]
Augmentation remained at about 57% in aggregate, essentially unchanged from the launch report. The report broke this down by occupation. Community and Social Service tasks were the most augmentation-heavy at roughly 75%, while Production occupations and Computer/Mathematical occupations were closer to a 50-50 split between augmentation and automation.[2]
The new extended-thinking mode introduced in Claude 3.7 Sonnet was used disproportionately by technical and creative roles. Roughly 10% of conversations from Computer and Information Research Scientists, 8% from Software Developers, 7% from Multimedia Artists, and 6% from Video Game Designers used the feature. The bottom-up cluster taxonomy in the same report identified copywriters and editors as having the highest task iteration (about 58% iterative conversations), translators and interpreters as the most directive group, and librarians as the heaviest users of learning-style interactions (about 56%).[2]
This report was the first to publish the 630-cluster bottom-up taxonomy, which organizes Claude conversations into granular usage buckets generated directly from the data rather than mapped downward from O*NET. That approach complements the top-down classification by revealing usage patterns that do not fit neatly into traditional occupational categories.
The third report, authored by Ruth Appel, Peter McCrory, Alex Tamkin, Miles McCain, Tyler Neylon, and Michael Stern, introduced the Anthropic AI Usage Index (AUI), which measures Claude usage relative to a country or state's working-age population. By raw share, the United States accounted for 21.6% of all use, with India next at 7.2% and Brazil at 3.7%. By per-capita AUI, the leaders were Israel at 7.0x expected usage, Singapore at 4.57x, Australia at 4.10x, New Zealand at 4.05x, South Korea at 3.73x, and the United States at 3.62x. Indonesia, India, and Nigeria were among the lowest-AUI countries, at roughly 0.36x, 0.27x, and 0.2x respectively.[4][9]
The report documented a strong link between income and AI adoption: a 1% higher GDP per working-age capita was associated with a 0.7% higher AUI globally, with the relationship steepening to about 1.8% within the United States.[4]
The most-used U.S. jurisdictions on a per-capita basis were:
| Rank | Jurisdiction | AUI |
|---|---|---|
| 1 | District of Columbia | 3.82 |
| 2 | Utah | 3.78 |
| 3 | California | 2.13 |
| 4 | New York | 1.58 |
| 5 | Virginia | 1.57 |
District of Columbia and Utah surpassed California in per-capita usage, an unexpected ranking that the report attributed in part to document editing in DC and broad consumer adoption in Utah. Hawaii used Claude for tourism-related tasks at roughly twice the national average; Massachusetts was overrepresented in scientific research; Brazil used Claude for translation and language learning at about six times the global average.[4][9]
This report was also the first to include first-party API data alongside consumer Claude.ai data. Business API usage was much more automation-heavy than consumer use: 77% of API conversations were classified as directive automation, against roughly 50% on Claude.ai. Software development was the dominant API category, and tasks specifically aimed at "developing and evaluating AI systems" accounted for about 5% of API traffic. Computer and mathematical tasks made up 44% of API usage versus 36% on Claude.ai.[9]
From December 2024 through August 2025, directive (fully delegated) conversations on Claude.ai climbed from 27% to 39%, the first time the Index showed automation-style use exceeding augmentation in absolute terms during the period. Educational tasks also rose from 9.3% to 12.4% of usage, and scientific tasks from 6.3% to 7.2%, while Business and Financial Operations fell from about 6% to 3%.[4][9]
The report drew an explicit parallel to historical general-purpose technologies such as electrification, noting that uneven diffusion produced lasting differences in living standards. It suggested that similar dynamics could apply to AI if adoption gaps between high- and low-income countries are not addressed through infrastructure investment and skills development.
The January 2026 report covered approximately one million Claude.ai conversations and one million API records sampled from November 13 to 20, 2025. Its main contribution was the introduction of the five "economic primitives" listed earlier (task complexity, human and AI skill, use case, AI autonomy, task success). The lead authors were Ruth Appel, Maxim Massenkoff, Peter McCrory, with Miles McCain, Ryan Heller, Tyler Neylon, and Alex Tamkin as additional authors.[10]
Key statistics from this release:
| Metric | Claude.ai | API |
|---|---|---|
| Share of work-related conversations | 46% | 74% |
| Overall task success rate | 67% | 49% |
| Directive (automated) conversations | 32% | 64% |
| Augmentation conversations | 52% | 36% |
Directive use on Claude.ai dropped from the 39% peak in August 2025 back down to 32% in November, with augmentation rebounding to 52%. The report estimated that, on Claude.ai, conversations involving high-school-level tasks ran roughly 9 times faster than the unassisted human baseline, while college-level tasks ran roughly 12 times faster, although success rates declined with task complexity (about 70% for basic tasks, 66% for college-level tasks).[10]
The deskilling pattern in the data was notable: Claude covers tasks that require, on average, about 14.4 years of education, an associate's degree equivalent. That is above the economy-wide average of 13.2 years. If tasks currently handled by AI were removed from the labor mix, the average education requirement across jobs would fall, suggesting AI adoption is currently concentrated above the education midpoint of the workforce.
Geographic concentration in the U.S. continued to fall: the Gini coefficient across states declined from 0.37 to 0.32 over three months, with the report estimating two to five years to full per-capita parity. Globally, concentration changed less, with the U.S., India, Japan, the U.K., and South Korea remaining the top users.[10]
The report also revised earlier productivity estimates. Without adjustment for task success rates, Anthropic's model implied roughly 1.8 percentage points per year of additional labor productivity growth. Adjusting for observed success rates pulled the estimate down to about 1.0 to 1.2 percentage points, and tighter assumptions about task complementarity (elasticity of substitution below 1) cut it further to 0.6 to 0.9 percentage points. About 49% of jobs now had AI usage for at least a quarter of their tasks, up from the 36% reported at launch.[10]
The "Learning curves" report, led by Maxim Massenkoff, Eva Lyubich, and Peter McCrory, covered February 5 to 12, 2026, three months after Claude Opus 4.5 was released and on the eve of Opus 4.6. Task concentration on Claude.ai continued to fall: the share of conversations covered by the top 10 tasks dropped from 24% in November 2025 to 19% in February 2026. Coursework fell from 19% to 12% of Claude.ai usage, while personal use rose from 35% to 42%. Coding traffic increasingly migrated to the API.[3][11]
A central claim of this report was that more experienced users were getting more out of Claude. Users with at least six months of tenure showed about 10 percentage points higher task success rates, asked about 10% fewer personal questions, and brought inputs requiring on average 6% more years of education. They also tended to interact more collaboratively rather than directively. Anthropic argued that the pattern was consistent with learning by doing, while explicitly acknowledging cohort effects (early adopters being more technical) and survivorship bias (users who churn are not observed) as competing explanations. They noted that controlled regressions ruled out simple versions of these confounders, such as long-tenured users simply bringing different kinds of tasks.[3]
The report also documented strategic model selection: for every additional $10 of hourly wage associated with a task, the share of conversations using Opus increased by 1.5 percentage points on Claude.ai, rising to 2.8 percentage points among API users. Users were, in other words, self-selecting into more powerful models for higher-value work.
Geographic convergence within the U.S. continued but slowed: the top five states' share of usage fell from 30% to 24%, but the estimated time to full state-level parity was extended from the earlier estimate of two to five years to a new estimate of five to nine years. Globally, concentration actually increased, with the top 20 countries growing from 45% to 48% of per-capita usage.[3]
On April 22, 2026, Anthropic announced the Anthropic Economic Index Survey, a monthly survey conducted through its Anthropic Interviewer product. Random samples of Claude users with accounts at least two weeks old are asked qualitative questions about workplace changes, productivity, hiring, role transformations, and longer-term expectations. The first published results, "What 81,000 people told us about the economics of AI," reported a mean self-rated productivity score of 5.1 on a 1 to 7 scale, with only about 3% of respondents reporting negative or neutral impacts. About 10% said their employer had responded to AI adoption by demanding more output. Workers in highly AI-exposed occupations were about three times as likely to mention worry about job loss as those in the least exposed quartile, and only 60% of early-career workers said they personally benefited, against 80% of senior workers.[12]
The survey approach complements the conversation analysis data by capturing user perceptions that do not show up in transcript structure, including attitudes toward displacement, reported hiring changes at the employer level, and expectations about career trajectory. It also extends the reach of the project to users who self-report qualitatively rather than just generating analyzed conversations.
This split is one of the most cited Index outputs and has shifted noticeably across reports.
| Report period | Augmentation share | Automation (directive) share |
|---|---|---|
| Late 2024 / launch (Feb 2025) | 57% | 43% (directive 27%) |
| Post-3.7 Sonnet (Mar 2025) | 57% | 43% |
| Aug 2025 (Sept report) | below 50% on Claude.ai | 39% directive |
| Sept 2025 API | n/a | 77% directive |
| Nov 2025 (Jan 2026 report) | 52% | 32% directive |
| Feb 2026 (Mar 2026 report) | recovering | falling |
The pattern is not a clean monotonic trend toward automation. Directive use spiked in mid-2025 as agentic and tool-use workflows matured, then partly reversed as longer reasoning models were used more interactively. Anthropic interprets this as evidence that the augmentation/automation balance depends as much on user habits and product affordances as on model capability.[4][9][10]
The API and Claude.ai channels have consistently behaved differently. Enterprise API usage hovers around 77% directive throughout the period covered, while consumer use has oscillated between 43% and 39% directive. This divergence reflects fundamentally different use cases: API integrations are typically purpose-built for specific automated tasks, while Claude.ai conversation use is more exploratory.
On April 28, 2025, Anthropic announced the formation of the Anthropic Economic Advisory Council, a group of distinguished economists appointed to advise the company on the economic implications of AI development and deployment. Two additional members were added on May 9, 2025.
The Council's mandate is to advise Anthropic on AI's impact on labor markets, economic growth, and broader socioeconomic systems, with the stated goal of informing research for the Anthropic Economic Index and helping frame findings for policymakers, researchers, and business leaders.
Council members as of mid-2025:
| Member | Affiliation | Specialization |
|---|---|---|
| Dr. Tyler Cowen | George Mason University (Holbert L. Harris Chair) | Economic growth, technology, culture |
| Dr. Oeindrila Dube | University of Chicago (Philip K. Pearson Professor) | Global conflict, development economics |
| Dr. John Horton | MIT Sloan School of Management | Labor economics, AI workforce effects |
| Dr. Anton Korinek | University of Virginia (Darden School) | Complexity, economic systems, AI |
| Dr. John List | University of Chicago | Field experiments, applied microeconomics |
| Dr. Ioana Marinescu | University of Pennsylvania | Labor economics, job search |
| Dr. Tomas J. Philipson | University of Chicago (former Acting Chair, White House Council of Economic Advisers) | Health economics, regulatory policy |
| Dr. Silvana Tenreyro | London School of Economics (former Bank of England MPC member) | Monetary policy, macroeconomics |
| Dr. Chiara Farronato | Harvard Business School (Glenn and Mary Jane Creamer Associate Professor) | Digital platforms, platform economics |
| Dr. Pascual Restrepo | Yale University | Automation, labor markets, inequality |
The Council's composition leans toward economists who have previously worked on technology's effects on labor markets. Pascual Restrepo, for instance, co-authored influential work with Daron Acemoglu on automation and the labor share. Anton Korinek had previously written on AI and macroeconomic growth. Several members (List, Philipson, Horton) had prior experience in government or policy advisory roles.
Formal Council outputs have not been published separately from the Index reports. The Council appears to function primarily as an advisory board rather than a co-authoring group, with individual members cited in acknowledgments of the research papers.
On June 27, 2025, Anthropic launched the Economic Futures program as a broader institutional umbrella for its economic research activities. The program has three main components.
The first is a research grants initiative offering rapid grants up to $50,000 for empirical studies of AI's economic impacts, along with API credits for research institutions. The grants explicitly target independent researchers rather than work produced internally at Anthropic.
The second is a policy forums component, which has convened two symposia: one in Washington, DC (with a July 25, 2025, proposal deadline, held at the McCourt School of Public Policy at Georgetown University), and one in London (with a September 12, 2025, deadline). The DC symposium generated publicly posted policy proposals covering workforce training grants, revenue taxes on AI-driven services, and business wealth tax designs. The proposals came from academics and think tanks and do not represent Anthropic's own policy positions.
The third component is the Economic Index itself, described within the program as its longitudinal measurement infrastructure.
The Economic Futures program sits organizationally alongside the Economic Advisory Council as the institutional scaffolding around the Index research. Together they represent Anthropic's attempt to build credibility for the Index by surrounding it with independent economists and externally visible policy engagement.
Anthropicpublishes the underlying aggregated tables for each report on Hugging Face under the dataset Anthropic/EconomicIndex. Data is licensed CC-BY and code is licensed MIT. As of May 2026, five data releases had been posted, corresponding to the launch and each major report:
| Release date | Contents |
|---|---|
| February 10, 2025 | O*NET task mappings, occupation shares, augmentation/automation classifications |
| March 27, 2025 | Updated occupation shares, 630-cluster bottom-up taxonomy, extended-thinking usage |
| September 15, 2025 | Geographic breakdowns (country and state AUI), first-party API data, trend tables |
| January 15, 2026 | Economic primitives measurements, productivity estimates, geographic Gini tables |
| March 24, 2026 | Learning curves data, tenure-based success rates, model selection by wage |
A separate folder in the repository hosts the "labor market impacts" data accompanying the March 2026 Massenkoff and McCrory paper, including their observed exposure scores by occupation.[5]
Data is intentionally aggregated rather than conversation-level. Cells with fewer than 15 conversations or fewer than 5 unique accounts are suppressed to limit re-identification risk. The dataset has been downloaded roughly 20,000 times per month as of mid-2026, making it one of the more actively used publicly available AI behavioral datasets.[5]
Anthropicâs contact for research questions about the dataset is econ-research@anthropic.com.
The most cited external use of Anthropic Economic Index data is a paper by Erik Brynjolfsson, Bharat Chandar, and Chenzi Xu (Stanford Digital Economy Lab), published as a working paper in late 2025 under the title "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence." The paper linked millions of worker records from ADP (the U.S. payroll processor) to occupational AI exposure measures that included data from the Anthropic Economic Index, specifically the estimated share of O*NET tasks touched by Claude conversations.
The study's headline finding was a 13% relative decline in employment for early-career workers (ages 22-25) in the most AI-exposed occupations from late 2022 to July 2025. It also found that these employment declines were concentrated in occupations where the Index classified AI use as directive (automated) rather than augmentative. This made it one of the first empirical studies to use the Anthropic augmentation/automation classification as a variable rather than just reporting it descriptively.
The Brynjolfsson, Chandar, and Xu paper was notable for several reasons beyond its findings. It validated the usefulness of the open dataset for labor economics research; it linked the Index's occupation-level measures to payroll outcomes, something the Index itself cannot do (because it has no employment data); and it was cited approvingly in subsequent Anthropic Economic Index reports as evidence of the dataset's external utility.[10][13]
A separate paper by Massenkoff and McCrory (Anthropic, March 2026) explicitly built on the Eloundou et al. framework by constructing an "observed exposure" measure that combines theoretical capability assessments with Anthropic's actual usage data. That paper found tentative evidence of slower hiring of workers aged 22-25 in AI-exposed occupations (roughly a 14% decline), broadly consistent with the Stanford findings, though the Anthropic authors were careful to describe their employment data findings as preliminary.[13]
Additional external researchers who have cited or worked with Index data include David Autor and Thompson (2025), who studied the educational composition of AI-exposed task bundles, and Gans and Goldfarb (2025), who engaged with the augmentation-versus-automation framing in their own work on AI's effect on specialization.
| Country | Share of total Claude use (Sept 2025) | AUI (per-capita relative) |
|---|---|---|
| United States | 21.6% | 3.62 |
| India | 7.2% | 0.27 |
| Brazil | 3.7% | not top in AUI |
| Israel | small | 7.00 |
| Singapore | small | 4.57 |
| Australia | small | 4.10 |
| New Zealand | small | 4.05 |
| South Korea | similar to Brazil | 3.73 |
| Indonesia | small | 0.36 |
| Nigeria | small | 0.20 |
The pattern Anthropic flagged is that small, technologically advanced economies are leading on per-capita adoption, while large-population countries with lower GDP per capita are catching up much more slowly. The report draws an explicit parallel to historical general-purpose technologies such as electrification, where uneven diffusion produced lasting differences in living standards.[4]
Within the U.S., the Index has tracked geographic convergence over time. The Gini coefficient across states fell from 0.37 in August 2025 to 0.32 in November 2025. However, by the March 2026 report, the estimated convergence timeline had lengthened from two to five years to five to nine years, as growth in leading states continued to outpace lagging ones.
Heartland Forward, a regional economic research organization, published a September 2025 analysis of Index data focusing on heartland states. It found that editing and improving written content ranked as the top Claude use case in 18 of the 20 heartland states it examined, and that AI adoption remained heavily concentrated in computer and mathematics occupations in every state it analyzed. Texas led the heartland in absolute usage; Kansas, Illinois, and Louisiana ranked in the national top ten for small business AI adoption; and Arkansas, Texas, and Mississippi led in student AI use.[14]
The launch report received coverage in technology and business press, with Axios first reporting it the day of release. Subsequent reports were picked up by Reuters, Bloomberg, the Financial Times, Built In, and Constellation Research.
Academic uptake has been steady. The companion arXiv papers (arXiv:2503.04761 for the launch and arXiv:2511.15080 for the September 2025 report) have been cited in labor economics and AI policy literature. The Stanford Digital Economy Lab's use of the data in an employment study lent the Index visibility in academic labor economics circles that it might not otherwise have reached. The Benton Institute for Broadband and Society has cited Index findings in discussions of regional AI readiness and broadband policy implications.
As of mid-2026, the Index is widely treated as the most detailed publicly available record of consumer LLM usage by occupation. Most academic users of the data explicitly caveat its findings against the single-platform and self-selection limitations described below.
The Economic Futures symposia generated formal policy proposals from academics and think tanks. The DC symposium at Georgetown's McCourt School produced recommendations on workforce training grants, AI revenue taxes, and business wealth taxes. These proposals do not reflect Anthropic positions but were generated through a process Anthropic organized and funded.
Some economists have used the Index to argue that AI adoption is not yet broadly visible in payroll data, which undercuts more alarming near-term displacement narratives. Others, particularly those who focus on early-career workers following the Brynjolfsson et al. findings, argue the Index data supports concern about concentrated effects on younger workers in specific occupations, even if aggregate employment statistics remain stable.
The data has been cited in U.S. congressional staff briefings on AI and labor, and in European policy discussions about AI Act implementation, though it has not been formally adopted as an official measurement framework by any government agency as of mid-2026.
Some critics have raised the naming issue: calling the project the "Anthropic Economic Index" implies broader economic coverage than a single-product usage study can provide, and alternatives like "Anthropic Usage Index" would be more modest. This critique does not challenge the data itself but questions whether the framing inflates its scope.
A methodological critique published on a data analysis blog noted identification problems: the paper cannot establish that it is measuring what it claims to measure because Claude's capabilities in different domains may lead people to use Claude differently, and Anthropic's own marketing positioning of Claude may influence which users show up. These endogeneity concerns apply to all product usage studies but are amplified here by the fact that the analyst and the product seller are the same entity.
Anthropicâs response has been to publish the data openly and acknowledge limitations prominently in each report, rather than to resolve the fundamental conflict of interest. The council of external economists serves partly as a credibility mechanism, but since those economists advise rather than independently audit, the tension remains.
The Index is one of several streams of empirical work attempting to measure AI's labor effects. It is distinguished mainly by its use of observed product traffic.
| Study | Method | Key finding | Limitation |
|---|---|---|---|
| Eloundou et al. "GPTs are GPTs" (2023) | Human + model annotation of O*NET tasks | ~80% of U.S. workers could have 10% or more of tasks affected | Measures capability, not usage |
| Brynjolfsson, Chandar & Chen "Canaries in the Coal Mine" (2025) | ADP payroll data + Anthropic Index exposure scores | 13% relative employment decline for ages 22-25 in most-exposed occupations | Cannot isolate AI causation from other factors |
| McKinsey Global Institute generative AI reports | Employer surveys + modeling | 30% of work hours potentially automated by 2030 | Survey-based, not behavioral |
| World Economic Forum Future of Jobs surveys | Employer surveys | Net job creation expected but large gross disruption | Self-reported employer expectations |
| Stanford AI Index (annual) | Synthesizes external sources | AI investment and adoption rising sharply | Aggregates secondary data |
| Massenkoff & McCrory "Labor market impacts" (2026) | Observed exposure combining Anthropic + O*NET + Eloundou | 14% hiring slowdown for young workers in highly exposed occupations; no aggregate unemployment signal | Preliminary; short time series |
The Eloundou et al. "GPTs are GPTs" paper estimated that about 15% of all worker tasks could be done significantly faster by LLMs alone, and up to 47-56% of all tasks with software built on top. The Anthropic Index's bottom-line task coverage estimate (about 36% of occupations seeing AI use in at least a quarter of tasks at launch, rising to 49% by January 2026) is broadly consistent with that exposure picture, but measures actual deployment rather than theoretical capability.[6][1][10]
A distinctive feature of the Anthropic data is that it captures consumer (Claude.ai) usage independent of formal employment relationships. That makes it an unusual complement to enterprise-survey work, but also limits direct comparison with payroll-based studies. The data covers users who paid for or used a free tier of Claude, which skews it toward affluent, English-speaking, technically oriented populations.
Anthropicpublishes a limitations section with each report, and external commentators have flagged additional concerns. The most important caveats are:
Single-platform sample. Until September 2025, the data covered only Claude.ai, not Anthropic's API or other LLM providers. The September 2025 report added first-party API data, but the Index still does not see usage of ChatGPT, Gemini, or other competitors.[1][9]
Self-selected user population. Claude users tend to be technical, English-speaking, and over-represent computer and mathematical occupations. Findings about occupation mix may say as much about who chose to sign up as about what AI can do.[3][12]
Survivorship bias on tenure analyses. The "learning curves" claims are based on currently active users, leaving out people who tried Claude and stopped. Anthropic explicitly acknowledges this and runs controlled regressions, but cannot fully rule it out.[3]
No counterfactual. Because the data only covers AI-assisted work, the Index cannot observe what the same users would have produced without AI. Productivity figures from the January 2026 report are model-derived estimates, not measured outcomes against a control group.[10][12]
Conflict of interest. Anthropic is reporting on usage of its own product. The Index's findings consistently support a narrative of broad, beneficial AI adoption with concentrated productivity gains, which is also what the company has commercial incentives to show. The transparency of the open data and the academic style of the papers partly mitigate this, but do not eliminate it.[5]
Privacy concerns. Even with Clio's defense-in-depth design, conversation analysis at this scale raises questions about informed consent. Anthropic's terms of service permit this aggregated analysis for Claude.ai accounts, but enterprise customers and API users typically opt out.[7]
Claude-analyzes-Claude circularity. The measurement system (Clio) runs on Claude, which means the model's own biases and categorization tendencies are embedded in the output. If Claude tends to over-classify certain types of content as, say, "computer science tasks," that will appear as elevated computer and mathematical occupational share in every report.
Theory-practice gap. The March 2026 companion paper found that AI theoretically covers about 94% of tasks using Eloundou et al.'s capability measure, but Claude actual usage only reaches about 33% of those tasks. The Index measures the 33% but says little about why the other 61% of theoretically feasible tasks are not being brought to AI.[13]