Anthropic Models How AI Could Hollow Out Knowledge Worker Wages by 2030

Anthropic put a number on the question every knowledge worker has been quietly asking: what happens to my job if this stuff actually works? Its Economics team released an interactive scenario explorer that models how AI could reshape US GDP, wages, and employment by 2030. You can plug in your own assumptions about capability and adoption to see what economy falls out.

The framework treats every occupation as a bundle of tasks drawn from the US Department of Labor's O*NET taxonomy. For any given task, AI can help a person work faster, take over entirely, have no effect, or spawn new work. Aggregate across millions of daily task instances and a macro picture emerges: GDP, the labor share, and the unemployment rate.

Three futures, one growing pie

The explorer highlights three scenarios. All grow the economy, with wildly different distributional consequences.

• Modest: AI has roughly the impact the internet did, adding 1.6% to 2030 GDP for a $34.1T economy.
• Substantial: AI is capable of doing half of all knowledge work by 2030, most of it autonomously, though it isn't adopted for all of that work. GDP comes in 8.3% higher at $36.3T, with the economy growing at roughly twice its normal rate.
• Extreme: AI outperforms humans on the vast majority of knowledge-work tasks, does nearly all of them autonomously, and creates essentially no new knowledge tasks for people. This case likely requires recursively self-improving systems adopted quickly across knowledge work. GDP hits $44.4T (+32.4%), annual growth reaches 15%, and the economy doubles every 4.5 years.

Who gets the paycheck

The interesting result is how the pie gets sliced. Today about 60 cents of every dollar produced goes to workers and 40 cents to capital. Under the substantial scenario, labor's share drops to 56.1%. Under the extreme scenario it collapses to 45.2%, handing capital an additional 14.8 points.

Wage effects split sharply by occupation. In the substantial case, pay for knowledge workers is essentially flat. In the extreme case, it falls by more than 10% by 2030. Non-knowledge workers such as electricians, nurses, and construction crews see pay rise, because AI-accelerated design and permitting drives more demand for physical work that models can't do.

Job churn follows a similar pattern. Under the substantial and extreme scenarios, knowledge workers face heavy automation and displacement, with coders and call-center agents potentially pushed into the trades. In the extreme case, unemployment spikes above typical recessionary levels because switching occupations takes many workers a long time.

What the public thinks

Anthropic paired the model with a Morning Consult survey of more than 10,000 Americans. The typical respondent's answers imply outcomes close to the substantial-change scenario: GDP is 10% higher by 2030 than it would be without AI, and unemployment has risen to around 5%. About 10% of respondents hold views in line with the extreme scenario.

The methodology, and its limits

The model is documented in a technical report (Korinek et al., 2026) circulated to economists including Daron Acemoglu, David Autor, and Emi Nakamura. Two of their comments visibly shaped the current version: rising returns to capital and diverging wages between exposed and unexposed occupations.

Anthropic is unusually candid about what's missing. The explorer leaves out policy responses, business cycles, potential aggregate demand or financial market disruptions, and possible catastrophic risks. It also excludes hyper-capable robotics, doesn't track individual workers through displacement, and, as reviewers noted, ignores the aggregate demand effects of the current data center buildout.

Why a frontier lab is telling on itself

Why is a frontier lab publishing a model that suggests its own technology could hollow out knowledge-worker wages? The answer fits Anthropic's broader policy posture: the outputs will feed the grants it makes through Economic Futures and inform its policy recommendations. The company is essentially arguing that the live question is no longer whether AI will grow the economy, but who ends up holding the larger pie.

For anyone building or deploying these systems, the practical takeaway is narrower than the headlines suggest. Task-level augmentation dominates in the modest and substantial scenarios, autonomous execution dominates in the extreme one, and the gap between them turns mostly on how quickly agentic systems become reliable enough to run without a human in the loop.