America’s debate about artificial intelligence often offers a false choice: either the technology will make everyone more productive, or it will make large numbers of people economically unnecessary. New research associated with Anthropic suggests a more uncomfortable possibility. The country could become substantially richer while some workers face weaker wages, fewer openings, and difficult transitions into different occupations.

The immediate backdrop is a September 2026 Anthropic Institute working paper, Economic Scenarios for Transformative AI, and two research releases dated October 16 about the company’s own workplace and the gap between public and employee expectations. Together, they raise important questions about America’s economic future. But they are three different kinds of evidence: a conditional model, company-level observations, and surveys. Treating them as one definitive forecast would misrepresent what the research actually establishes.

This analysis reviews those sources as available on October 18, 2026, alongside independent job-posting research. It does not report original interviews, a proprietary employment survey, or an independent audit of Anthropic’s internal performance measures. The distinction between a warning and a measured outcome runs through the entire story.

First, this is not a single prediction of America’s future

The September working paper is by Anton Korinek, Charles I. Jones, Szymon Sacher, Tess Cotter, and Peter McCrory. Its purpose is to translate assumptions about AI capabilities and adoption into possible paths for output, wages, employment, unemployment, and income distribution through 2030. The authors explicitly say their scenarios are not predictions and assign no probabilities to them. They also state that the paper’s views do not necessarily represent Anthropic or the Anthropic Institute. Those qualifications matter when discussing an ‘Anthropic forecast.’ 1

The accompanying scenario explorer allows users to change expectations and see the economic outcomes those assumptions imply. That makes it a structured way to examine disagreements, not a device that reveals what will happen. A dramatic result can emerge because someone supplies dramatic inputs. The relevant question is whether those inputs describe a plausible path, and which observations would support or contradict them. 2

The framework separates cognitive occupations—management, professional, sales, and office work—from other occupations. AI can automate or augment tasks in the cognitive group; workers displaced from that group may seek work elsewhere. The transition is not frictionless. A person who loses a software or administrative job cannot instantly become an electrician, regardless of how healthy demand for electricians might be.

What the modest, substantial, and extreme scenarios actually say

The modest scenario looks more like a conventional technological improvement than an economic rupture. At the start of 2030, GDP is 1.6% above the model’s no-AI path. Annual GDP growth is 2.4%, compared with a 2.0% no-AI baseline, and economy-wide unemployment is 3.9%, compared with 3.8% in that baseline. These are model outputs, not readings of today’s economy. 1

The substantial scenario produces a much larger change. GDP is 8.3% above the no-AI path at the start of 2030, with growth of 5.4% over the preceding twelve months. Cognitive employment is 3.9% below its mid-2026 level, and economy-wide unemployment reaches 4.6%. The average wage is 2.1% above its no-AI path, but cognitive wages are 0.3% below theirs. Aggregate progress and uneven outcomes coexist. 1

The extreme scenario is the source of the most alarming numbers. GDP is 32.4% above the no-AI path, annual growth reaches 15.4%, and cognitive employment is 21.5% below its mid-2026 level. Unemployment among workers who began in cognitive occupations reaches 17.9%; the economy-wide rate is 11.9%. These are different denominators. ‘Nearly one in five cognitive workers unemployed’ is not the same claim as one in five Americans unemployed. 1

Another distinction is easy to lose in headlines: an 8.3% higher GDP level by 2030 is not 8.3% annual growth. Likewise, a modeled unemployment rate is not a probability that a particular worker will lose a job. The paper’s tables specify the start of 2030, and define growth using changes over the twelve months leading to that date.

Why a richer country can still leave workers worse off

GDP measures output, not how comfortably a particular household lives. When AI makes production cheaper or faster, the resulting income can flow to employees, business owners, investors, customers, or some combination. The distribution depends on labor demand, competition, ownership, and the supply of the capital needed to deploy the technology.

In the substantial scenario, labor’s share of income falls from the model’s initial 60% to 56.1%. In the extreme scenario it falls to 45.2%. The average wage still rises in the extreme case, but cognitive wages are 11.5% below their no-AI path, while wages in other occupations are 33.6% above theirs. A national average therefore conceals sharply different experiences across occupations. 1

Labor’s income share is not a count of eliminated jobs, and the model’s cognitive category is not identical to every role conventionally called white-collar. But the underlying policy question is clear: if society creates much more value while the workers most directly exposed capture less of it, economic growth alone does not resolve the adjustment problem.

The assumptions behind the most dramatic outcome

The extreme scenario combines broad AI reach with extensive actual adoption, large productivity improvements on affected tasks, and a high proportion of automation rather than assistance. It also assumes no new cognitive tasks are reinstated through the model’s new-task channel. In the substantial scenario, by contrast, one new cognitive task is introduced for every four tasks automated. The difference is a major assumption about how much new human work appears, not an established fact about future employers. 1

The authors also examine how outcomes change when capital is easier or harder to supply, and when cognitive wages adjust at different speeds. A slower wage adjustment can shift more of the burden into unemployment; faster adjustment can shift it into lower wages. Neither is painless for affected households. Much of the divergence between scenarios occurs after 2027 because the model begins from a shared mid-2026 anchor.

These sensitivities are a reason to take the work seriously without treating its extreme case as destiny. They identify observable questions: how widely firms deploy AI, whether productivity improvements survive review and integration costs, how much new work emerges, and how quickly displaced people find suitable employment. The model simplifies a heterogeneous economy into two occupational groups; it cannot tell every profession, industry, or city its individual future.

October’s new finding: the public and AI workers see different futures

Anthropic’s October 16 report, The AI expectations gap, compares US adults with Anthropic staff. It reports that 39% of surveyed US adults expect AI to make a Nobel-level discovery by 2030, compared with 91% of staff. Among US workers who use AI at work, 26% say an AI-suited task takes a quarter of its former time or less; the corresponding figure among Anthropic staff is 59%. These figures measure expectations and self-reported experiences, not independently timed productivity gains. 3

More than half of Anthropic staff expect AI to be able to do at least 80% of their current work by 2030, compared with 5% of US workers. That is a belief about future capability, not evidence that those jobs have disappeared or that firms will choose to remove them. It nevertheless reveals a striking difference between people working at an AI frontier company and people across the broader labor market. 3

One interpretation is that frontier workers see changes earlier. Another is that their experience is unusually concentrated in AI-friendly tasks, tools, and organizational incentives. Their confidence may reflect useful information, but it cannot establish that a hospital, local government, manufacturer, or small retailer will experience the same trajectory. The survey gap itself does not decide which group will prove more accurate.

What the public survey can—and cannot—add to the model

The September paper describes a Morning Consult survey of 10,980 unique US adults fielded August 11–23, 2026, with population weights. Its model-implied outcomes use the 3,259 respondents who answered all five relevant items for the 2030 exercise. Across those responses, median modeled GDP is 8.6% above the no-AI path and median cognitive employment is 4.2% below its mid-2026 level—close to the paper’s substantial scenario. 1

Those are not respondents directly predicting an 8.6% GDP gain. Researchers convert answers about capabilities, adoption, speed, automation, and re-employment into model inputs. Some other parameters remain fixed. This is a useful translation of public beliefs into economic implications, but neither a consensus among economists nor a probability-weighted forecast.

Inside Anthropic, work is already changing—but one company is not America

A separate October 16 release says Anthropic has been hiring for a smaller share of junior positions and that a smaller share of employees work in technical roles. It reports that agents worked more hours than employees by March 2026 and that there were six times as many agents as employees by August. These are company-reported operational measures. Agent counts and agent hours do not translate directly into human full-time equivalents or jobs displaced. 4

The company reports more code per person, more sales closed per sales representative, and more contracts processed. It also acknowledges that code output is an imperfect productivity measure. Increased output among employees adopting tools is informative, but an observational relationship inside one fast-growing company is not an independent causal estimate of the effect on US employment.

One especially relevant observation is that human attention becomes a bottleneck as agents do more work. Employees devote more effort to monitoring their outputs. That suggests a change in job design: less direct production in some workflows, more specification, supervision, checking, and responsibility. It does not establish how many such roles will exist, what they will pay, or whether they will compensate for lost entry-level opportunities.

The Economic Index measures usage, not a national unemployment rate

Anthropic’s Economic Index is a different research program from the September scenario paper. The June 26 report, Cadences, describes a shift toward longer-running agentic tasks with Claude Code and Cowork. It also changes the measurement pipeline, including higher-rate sampling, a new output classifier, and separate reporting for chat/Cowork conversations and first-party API use. Comparisons across reports must account for those changes. 5

The March report illustrates why simple substitution narratives are inadequate. It found slightly increased augmentation—collaborative use where AI complements a person—in both Claude.ai and API traffic, alongside diversification of Claude.ai tasks. Its February sample found the top ten tasks accounted for 19% of traffic, down from 24% in the prior November sample. Those percentages describe sampled product conversations, not the share of American work automated. 6

A task appearing in Claude data can reflect experimentation, drafting, learning, repeated attempts, or successful production. It does not by itself show an entire job was replaced. Nor is Claude’s customer population a representative sample of all US employers. The Index is valuable as evidence about where one important AI system is being used; it must be joined to independent economic data before drawing economy-wide conclusions.

Independent hiring evidence complicates the displacement story

A May 14, 2026 analysis by New York Fed researchers Richard Audoly, Miles Guerin, and Giorgio Topa combined Lightcast US job-posting data with an occupational exposure measure based on Anthropic task research. It found little indication of a distinct AI-driven decline in labor demand in those postings, even though overall hiring had slowed. Relative declines in more-exposed occupations began before ChatGPT’s late-2022 release, complicating a causal interpretation. 7

The same study found no clear post-release divergence between junior and senior postings within highly exposed occupations. That does not prove AI has no effect on any company or worker. Job listings are not a complete census of actual hires, layoffs, hours, compensation, or informal hiring decisions. The finding is narrower: the patterns examined did not support a simple account in which generative AI was the main cause of the hiring slowdown.

This evidence and Anthropic’s workplace report need not be mutually exclusive. One AI-native company can reduce the junior share of its hiring while the broader market has not yet shown the same pattern. A forward-looking risk can be real without having become a dominant national outcome. Conversely, a serious forecast does not justify labeling every weak job market an AI consequence.

Entry-level work is a pipeline, not just a cost center

The US implications extend beyond the number of jobs available this quarter. Junior employees learn through routine assignments, feedback, and exposure to practical mistakes. If firms automate the assignments through which expertise develops, they may eventually face a thinner pipeline of experienced people able to supervise difficult decisions.

A smaller junior hiring share is also not automatically a decline in the absolute number of junior hires. A growing employer can increase junior hiring while increasing senior hiring faster. To assess the pipeline, observers need both counts and proportions, alongside training arrangements, promotion paths, retention, and the quality of the work beginners receive.

For employers, the question should therefore be broader than whether a task can be completed cheaply by a model. Which assignments build judgment? Who reviews automated work? How are beginners given enough responsibility to become competent supervisors? Eliminating visible training costs can create less visible future capability costs.

What professionals and businesses should do now

Professionals can start by separating tasks that involve routine transformation of information from tasks that require contextual judgment, trust, accountability, or coordination. This is not a guarantee of job security. It is a more useful planning exercise than treating a whole occupation as either safe or doomed. Learning to evaluate output and connect it to a real business decision is different from learning only to generate more text.

Businesses should measure completed, accepted work rather than prompts sent, documents drafted, or lines of code produced. Include checking time, corrections, data restrictions, integration costs, and the consequences of mistakes. A workflow that looks faster during a demonstration may be slower once review is included; another may create genuine additional capacity without requiring fewer staff.

Workforce plans should distinguish cutting existing roles from changing future hiring, and distinguish both from redeploying people. Each mechanism affects households differently. Transparent measurement also makes it easier to decide when a training investment is effective and when a supposedly automated workflow still depends on hidden human effort.

America’s policy question is how to manage the transition

The paper’s deepest challenge is not a particular unemployment number. It is whether the institutions governing work can adapt if change is faster than workers’ ability to move. Retraining can help, but it does not itself create openings, solve geographic constraints, cover living costs during a transition, or restore a lost professional network.

The authors discuss mechanisms for sharing gains, including income support and broader access to capital income, without resolving which policy package would work best. Policymakers must weigh costs, incentives, administration, and political feasibility. The existence of larger aggregate gains does not mean compensation happens automatically, and the extreme scenario should not be treated as a mandate for a single policy. 1

A credible US monitoring effort would track occupational hiring, earnings, hours, entry-level access, re-employment time, and firm-level adoption alongside output growth. It would examine differences across industries and regions, rather than waiting for a national headline statistic to summarize every local change.

The bottom line: preparation without inevitability

Anthropic’s recent research gives America a more concrete vocabulary for an uncertain future. It shows how rapid productivity growth could coexist with weaker outcomes for some knowledge workers, and how people at the AI frontier may form very different expectations from the public. Neither finding establishes that the extreme scenario will occur.

The responsible conclusion is neither complacency nor a declaration that mass unemployment is inevitable. Watch adoption, measure accepted productivity, preserve pathways into expertise, and distinguish a company’s experience from the country’s. The most important question is not simply how capable AI becomes. It is how American businesses and institutions convert that capability into prosperity people can actually share.

Sources and scope

Primary economic model: 1 — Economic Scenarios for Transformative AI, Anthropic Institute Working Paper No. 2026-02, September 2026; especially the introduction, Tables 3–4, and sensitivity analysis. Companion tool: 2 — Economic scenario explorer. All scenario figures in this article are conditional outputs, with GDP levels compared against the model’s no-AI path unless otherwise specified.

October company research: 3 — The AI expectations gap, October 16, 2026 and 4 — Working at the frontier shapes expectations for the future, October 16, 2026. Survey answers are beliefs and self-reports; internal company findings are not independently audited here.

Product-usage evidence: 5 — Anthropic Economic Index: Cadences, June 26, 2026 and 6 — Anthropic Economic Index: Learning curves, March 24, 2026. Independent labor-demand analysis: 7 — Do Job Postings Show Early Labor-Market Effects of AI?, New York Fed Liberty Street Economics, May 14, 2026. Its authors’ analysis concerns job postings and does not establish every employment effect of AI.