Anthropic’s AI Economy Scenarios for 2030: Growth Could Surge While Knowledge Workers Lose Ground

Anthropic’s Economic Scenario Explorer is designed to show what different assumptions about AI capability, adoption, automation, new tasks, and worker mobility would imply for the U.S. economy by 2030. It is not a forecast: Anthropic assigns no probabilities to its modest, substantial, or extreme scenarios. 12

Its central implication is straightforward: AI-driven output growth and broadly shared labor-market gains are not the same thing. In the faster-automation cases, the economy grows more quickly while knowledge workers face weaker wages, more difficult occupational transitions, and a declining share of national income. 12

The three modeled 2030 outcomes

Scenario	Core assumption	2030 GDP	Main labor-market result
Modest	AI has an effect broadly comparable to the internet, with gradual adoption.	$34.1 trillion in 2025 dollars, 1.6% above the no-AI path.	Unemployment stays near its historical range and labor’s share is about 59.4%, close to the model’s 60% starting point. 12
Substantial	AI can perform about half of knowledge work by 2030, much of it autonomously, but is not used for all of that work.	$36.3 trillion, 8.3% above the no-AI path.	Growth is roughly twice the normal rate, but knowledge-worker wages are broadly flat while workers in other occupations gain more. 12
Extreme	AI outperforms humans on most knowledge-work tasks, performs nearly all of them autonomously, creates essentially no new knowledge tasks, and is adopted rapidly.	$44.4 trillion, 32.4% above the no-AI path.	Annual growth reaches about 15%, but knowledge-worker wages fall by more than 10% and labor’s share drops to about 45.2%. 12

The extreme case is intentionally extraordinary. Anthropic describes it as an outcome that would likely require recursively self-improving AI rather than a simple continuation of present-day deployment trends. 12

Why higher GDP does not guarantee better outcomes for workers

The explorer separates cognitive or knowledge occupations—the work AI directly affects in the model—from other occupations that are more physical or in-person. AI can augment people, automate tasks, or create new tasks. The final employment result depends on the balance among those forces, not on a single count of jobs “replaced.” 1

When AI substitutes for cognitive tasks, the model can produce lower cognitive wages, less employment in cognitive occupations, and unemployment while displaced workers seek work elsewhere. The severity of that transition rests heavily on assumptions about wage rigidity, the pace at which firms open vacancies, and whether workers can move across occupations. 1

That makes retraining a real constraint in the exercise, rather than an automatic solution. A worker displaced from a knowledge occupation is not assumed to immediately find an equivalent job at equivalent pay. 1

What happens to physical occupations and capital owners

Physical and service occupations appear relatively protected in the substantial and extreme cases because advanced robotics is outside the model’s scope. Their relative insulation can raise their wages and employment prospects as AI boosts demand elsewhere in the economy. That is a modeling result, not a claim that those occupations are permanently safe from automation. 1

The counterpart to labor’s declining income share is a larger claim on output for capital owners. In the extreme case, labor’s share falls from 60% to about 45.2%, meaning the gains from rapid AI-enabled production accrue much more heavily to capital. 12

The model’s sharpest warning is therefore distributional: an economy can expand rapidly even as the benefits are concentrated and the workers most exposed to cognitive automation lose bargaining power.

Unemployment: distinguish the aggregate from knowledge work

The extreme case should not be summarized as 17.9% economy-wide unemployment. That figure refers to cognitive or knowledge-worker unemployment in the model. A contemporaneous summary of the scenario reports 11.9% overall unemployment, alongside 17.9% cognitive unemployment. 45

That distinction matters. The model’s disruption is concentrated in occupations where AI can directly perform a large share of tasks, while workers outside those occupations are relatively insulated by the model’s exclusion of advanced robotics. 1

What more than 10,000 Americans expected

Anthropic says the explorer lets participants compare their assumptions with those of more than 10,000 Americans. 23 According to Anthropic’s reported survey comparison, the typical respondent’s assumptions implied an outcome near—though somewhat stronger than—the substantial scenario: GDP around 10% above the no-AI path by 2030 and aggregate unemployment near 5%. About 10% of respondents gave inputs broadly consistent with the extreme case. 2

Those results measure expectations, not likelihood. Participants supplied beliefs about future AI capabilities, adoption, autonomy, task creation, and the ease of changing occupations; the model then translated those inputs into economic outcomes. 12

How to read the scenarios alongside Evan Hubinger’s warning

Evan Hubinger’s statement that he personally sees a greater-than-10% chance of AI killing all humans within a decade is not a probability attached to any Economic Scenario Explorer outcome. In the same statement, he said Anthropic does not yet have a plan to solve alignment for superintelligence and is not clearly on track to do so. 31

The two analyses operate at different levels:

• The explorer asks how output, jobs, wages, and income shares might change if a functioning U.S. economy continues through 2030 under specified AI assumptions. 1
• Hubinger’s statement concerns a catastrophic tail risk in which sufficiently capable, misaligned AI could prevent that ordinary economic future from occurring. 31

As a result, the $44.4 trillion extreme-case GDP figure is not an expected-value prediction for the United States. Anthropic’s framework explicitly treats scenarios as conditional illustrations, not forecasts, and does not assign them probabilities. 1

It also omits several forces that could substantially change the result: policy responses, business-cycle dynamics, advanced robotics, a possible AI-investment bubble, and existential risk. 1 Some omissions could soften disruption—for example, effective redistribution, labor-market institutions, or better transition support. Others could worsen it or make conventional GDP projections irrelevant.

The most useful takeaway is not that 15% annual growth is inevitable, nor that a personal extinction-risk estimate establishes a known probability for AI catastrophe. It is that rapid cognitive automation creates two separate questions: how gains are distributed in a functioning economy, and whether advanced systems can be made safe enough for that economic future to be meaningful at all. 131