Anthropic released a tool that projects three paths for the US economy through 2030. In the most aggressive one, GDP reaches 44.4 trillion dollars while knowledge worker unemployment hits 17.9% and labor’s share of output falls to roughly 45%. That branch also happens to be where its own CEO’s public forecasts land.
Key Takeaways
- The Economic Scenario Explorer maps three futures for the US economy out to 2030.
- The extreme branch pairs growth 32.4% above the no-AI path with 17.9% unemployment among knowledge workers.
- Dario Amodei’s public forecasts sit inside that branch, which the model presents as the least likely of the three.
Have an AI Sum Up This Article
ChatGPTThree Paths for the US Economy Through 2030
The lab put out an economic scenario explorer that projects American output through 2030. The tool hands the reader two dials: how far model capability climbs, and how fast companies actually put it to work.
The calmest path puts US GDP at 34.1 trillion dollars in 2030, 1.6% above where it would have landed without AI. Anthropic compares that shape to what the web delivered: real, diffuse improvement with no rupture in the labor market. Wages hold roughly steady, and the occupational mix shifts slowly enough that people move between roles over a career rather than inside a quarter.
The middle path lifts GDP to 36.3 trillion, an 8.3% gain. The scenery changes here. Growth doubles, but it arrives alongside heavy occupational switching, and wages for knowledge workers stop climbing while other categories pick up ground.
The third path is the one that makes headlines. GDP reaches 44.4 trillion dollars, 32.4% above the no-AI trajectory, with an economy doubling in size every 4.5 years. Anthropic is explicit that this branch would require recursively self-improving systems paired with very fast adoption across knowledge work.
None of the three is offered as a prediction. The point of the model is not to say what happens, but to make visible the conditions under which each outcome becomes plausible. That framing matters for reading the numbers that follow.
What separates the branches is not raw capability. It is the pace at which organisations genuinely hand work over. Knowledge worker unemployment barely moves when adoption stays slow, even with high capability on the other dial, and that is the sharpest methodological takeaway in the whole tool.
GDP Climbs While Labor’s Cut Shrinks
In the extreme branch, knowledge worker unemployment reaches 17.9%. Anthropic calls that a historic level, which it is: the natural comparison sits with major recessions rather than ordinary cycles. The distinction worth holding is that this rate does not describe the whole labor market, only a bounded population whose job is processing information. Wages for those who keep their seat fall 11.5% in the same branch, while manual workers see pay rise.
The most structural number sits elsewhere. Labor’s share of output drops to about 45%, down from roughly 60% today. The economy genuinely gets richer, but a much larger slice of that wealth flows to capital, meaning the infrastructure and the models producing the value, rather than to the people doing the work.
That deserves isolating, because it shifts the debate. Public conversation fixates on how many jobs disappear. The Anthropic model points at a different pivot, the split of the value created, which can happen in a country posting strong growth and no mass unemployment at all. A fifteen-point move in labor share is not a rounding error either. It is the sort of shift that rewrites tax bases, pension arithmetic and the political economy of an entire decade, and it would arrive without a single dramatic headline number to hang it on.
The two effects also stack, and the stacking is what makes the extreme branch politically hard to absorb. A country could post record growth, double-digit knowledge worker unemployment and falling pay in skilled roles inside the same year. Standard macro indicators would tell a success story while a slice of the workforce lived something else entirely.
The anxiety was already measured. We reported on a survey from the same lab where 64% of respondents feared for their job because of AI, a level of worry this model rationalises rather than contradicts.
One segment is already moving. Anthropic documented the entry-level role thinning out, the rung that historically served as the bridge between study and career. The extreme branch describes what a market looks like when that bridge stays closed.
More articles on Horizon
- Paul Christiano Joins the Board That Clears OpenAI Models
- Jacob Coxon Accuses Anthropic of Gambling With Our Lives
- Claude Code Automates Your Repetitive Office Work
A Model That Files Its Own CEO Under Unlikely
Dario Amodei publicly argued in May 2025 that AI could wipe out up to half of entry-level office roles and push unemployment to between 10% and 20% by 2030. Those figures do not float freely: they map onto the extreme branch of this model.
The consequence is awkward in-house. The tool Anthropic published files its own chief executive’s forecasts under the least likely of the three branches, without ever putting it that way. Nothing in the interface names Amodei or flags the overlap. The reader has to carry the May 2025 numbers in from outside and notice where they land, which is why the point took a day to surface after release.
Two readings hold up equally well, and the model settles neither. Either the lab is softening a message that had become commercially awkward, cushioning the alarm between two calmer paths. Or it is doing exactly what an honest modelling exercise demands, showing that the most-quoted forecast also needs the heaviest assumptions.
One detail leans toward the second reading: the shape of the artifact. A lab trying to mute an alarm does not ship a simulator with the dials exposed. It publishes a note. Here anyone can rebuild the extreme branch in three clicks and inspect what it assumes, which means 17.9% knowledge worker unemployment stays reachable inside the interface. It is conditioned, not buried.
The tension between Amodei’s public alarm and his reassuring register is not new. We flagged it when he pinned the growing backlash against AI on a trust problem rather than on effects workers were reporting directly.
For rival labs, the release creates an implicit obligation. Anthropic just planted a quantified, open, manipulable frame on ground nobody was occupying. Anyone disputing these projections now has to produce a model rather than a statement, and that raises the bar for the whole argument. For teams steering headcount the practical use is immediate and modest at once: the adoption dial is the only parameter a company actually controls, and a survey from the same lab already showed Claude handling half the workload on some roles. That dial, not the capability one, decides which branch an organisation lands in.
Follow the story on Horizon.
