When I first opened Anthropic’s economic scenario explorer, I expected another polished argument about productivity. Instead, I found a more uncomfortable proposition. The United States could become materially richer by 2030 while a large part of the people who built their careers around knowledge work receive little of that added prosperity, or lose their position in the labour market altogether.
Anthropic is careful about the status of its work. This is a model of possible futures, not a forecast, and its authors are unusually explicit about what remains outside the model: policy choices, business cycles, financial shocks, demand effects, regional differences and the uneven cost of asking an actual person to start again in a new profession. Those caveats matter. They do not make the exercise less useful. They make its central tension clearer: economic growth and social progress are related, but they are not the same thing. [1] [2]
That tension extends an argument I made recently in The AI bill. We are often invited to admire the scale of the AI dividend before we have decided who owns it, who receives it and who is expected to absorb the risk while it compounds elsewhere. Anthropic’s scenarios put numbers around that problem.
“The main challenge is not achieving economic growth, but making sure the benefits are broadly shared and the costs aren’t unequally dispersed.” [1]
A model built from tasks, not job titles
The useful premise of the research is that jobs are bundles of tasks, each with a different relationship to context, responsibility and human trust. A nurse may spend a shift documenting vitals, coordinating care, talking to patients, ordering supplies and making judgements that depend on all three. AI can accelerate some tasks, automate others and create a small set of new ones around supervision, quality control or coordination. The same logic applies to a designer, a lawyer, a financial analyst or a junior consultant.
That distinction is much more valuable than the tired binary between jobs that “survive” and jobs that “disappear”. The relevant question is how the task bundle changes, how quickly that change is adopted, and who has enough time, income and support to adapt before the old bundle stops paying for a life.
The International Labour Organization reaches a similar conclusion from a different direction. Its 2025 index finds that one worker in four is in an occupation with some exposure to generative AI, while also arguing that transformation is more likely than complete replacement for most occupations. Clerical work remains the most exposed category, although specialised professional and technical work is increasingly affected as the technology becomes more capable. [3]
The middle scenario is already uncomfortable
Public discussion tends to oscillate between two reassuring stories. One claims that AI will change very little because organisations move slowly. The other assumes that any sufficiently capable system will replace everyone almost immediately. Anthropic’s most useful case is the one in the middle.
In its substantial scenario, AI can perform half of all knowledge work by 2030, much of it autonomously, though firms do not deploy it across every eligible task. The result is an economy that grows at roughly twice its normal rate, with GDP 8.3% higher in 2030 than it would have been without AI. That is a serious gain in a short period. It is also a scenario in which knowledge-worker wages are essentially flat. [1] [2]
This matters because knowledge work has been sold for decades as the relatively safe side of economic change. Education, analytical skill and proximity to information work were supposed to create resilience. The model does not say that these skills become worthless. It says that their scarcity can be reduced faster than the institutions around them can adjust, especially when the work can be decomposed into digital tasks and delivered through existing software.
Anthropic uses examples such as a software engineer moving into electrical work, or a call-centre agent training as a nurse, to illustrate a much harder fact: occupational mobility is never as frictionless as a chart makes it look. Training takes time, professional identity is sticky, and care responsibilities, geography, licences, health and income all shape whether a person can move. A vacancy elsewhere in the economy does not automatically become a viable next step for the person whose role has changed.
Average wages can rise while a career still loses value
The wage chart is where the abstract language of “net gains” starts to fail people. In the substantial scenario, average wages increase because demand expands in work outside the most exposed knowledge occupations. In the extreme case, the difference becomes stark: wages outside knowledge work rise 33.6% relative to the no-AI economy, while knowledge-worker wages fall 11.5%. The average still rises 9.7%. [1]
An average is useful for a national accounts table. It is a poor summary of a household budget, a mortgage application, a professional identity or the decision to invest years in education. The social cost of transition is not distributed across a fictional representative worker. It lands on specific people at a specific time, often when their accumulated experience has become difficult to price in the new market.
The conversation about junior roles deserves more attention than it usually receives. Entry-level work has never been only a cheaper way to get routine tasks done. It has been an apprenticeship system, often badly designed, through which people learn the judgement that later lets them manage ambiguity, clients, risk and responsibility. If AI removes too much of the work through which that judgement is formed, the eventual shortage may be experienced people who had the opportunity to learn how to make decisions before the machine began proposing the first draft.
The distribution problem arrives before the apocalypse
The most revealing chart in the explorer is also the least cinematic. It shows the division of national income between labour and capital. In the modest case, labour receives 59.4% of GDP in 2030. In the substantial case, the share falls to 56.1%. In the extreme case, it reaches 45.2%, while capital receives 54.8% of a much larger economy. [1]
There is no contradiction in an economy becoming richer while workers receive a smaller part of its income. That is precisely the point. AI can make capital more productive because compute, models, data, energy infrastructure and proprietary distribution systems become useful across more tasks. The owners of those assets receive higher returns, even where aggregate wages rise.
The IMF has made the same structural warning in broader terms. AI reaches high-skilled work as well as routine work, and the returns can favour people and firms already equipped to deploy it. If the technology complements higher earners first and lifts returns to capital at the same time, inequality can widen within countries even while productivity improves. [4]
This is where the language of “everyone will benefit eventually” becomes insufficient. Eventual benefit does not settle the question of timing, ownership or bargaining power. A worker whose wage stagnates for four years while the firms around them become more valuable has not received a neutral share of a national success. They have financed part of the transition with forgone income and uncertainty.
Growth needs an operating system for distribution
I do not read this research as a case against AI adoption. Avoiding useful technology would not protect workers for long, and it would not solve the underlying weakness of labour markets that already struggle to reward care, craft, public service and junior learning properly. I read it as a case against treating adoption as a purely technical implementation decision.
A serious response begins inside companies. If a firm uses AI to remove a layer of routine knowledge work, it should be able to explain what it will do with the people whose jobs supplied that layer: redeploy them, train them, retain them during the transition or simply reduce headcount. Each choice has a cost. Pretending that the choice is made by an impersonal model is a convenient way to avoid taking responsibility for it.
It also begins outside companies, because no individual employer can solve a distribution problem created by a concentrated set of assets. Anthropic’s own policy paper considers wage insurance, job-matching infrastructure, sector-based training and forms of universal capital accounts that give more people a direct stake in AI-driven growth. These proposals are debatable and need local translation, especially outside the US. Their underlying premise is harder to dismiss: if income increasingly comes from ownership of the systems that do the work, broad participation in ownership becomes part of the social contract. [5]
The task for leaders is more demanding than a productivity dashboard. They need to measure which work is being augmented, which judgement is being hollowed out, which people are being asked to retrain and which assets capture the gain. Those questions belong in workforce strategy, procurement, compensation and education policy before they become a crisis-management exercise.
Anthropic’s model does not tell us which scenario will occur. It gives us a clearer map of the choices hidden inside each one. The technology may make the economy larger. The quality of the transition will be decided by whether we build institutions that let people share in the value they help create, rather than merely adapt to the cost of creating it.