Anthropic's own economists model three scenarios for the 2030 economy, from "meh" to labor's share collapsing

Full analysis

The Researcher: Anton Korinek, Chad Jones, Szymon Sacher, Tess Cotter and Peter McCrory built a task-based model, meaning it slices the economy into tasks and asks what fraction AI does, then how that flows into output and wages. The scenario spread is driven by three dials: how fast AI capability improves, how widely it's adopted, and how much productivity rises. The three cases map cleanly onto real forecasts. The substantial change scenario, with GDP 8.3 percent higher by 2030, corresponds to the forecasts that circulated from financial and consulting institutions in 2023. The middle case is essentially the Goldman/consulting consensus dressed in cleaner math. The extreme case is the genuinely new claim: AI takes over close to a third of all economic tasks, output surges 32% beyond the no-AI trajectory, unemployment climbs toward 12%, and white-collar employment drops by more than 20%. Follow the dials. The headline number is downstream of them.

The Skeptic: A lab that sells the automating technology and then publishes a model of what automation does to the economy has a conflict worth naming out loud. And the framing does real work. Anthropic built a model that frames its CEO's bleakest job forecasts as an outlier scenario. Dario Amodei's earlier "half of entry-level white-collar jobs could vanish" warning now sits in the extreme corner, firmly outside the base case. Convenient. The reviewers pushed back hard in both directions. Some thought the extreme case reads better as a thought experiment than a forecast; others thought the modest case understates what is already visible. And the model quietly excludes the ugly parts. No recessions, no financial contagion, no feedback where job losses cut spending and trigger more layoffs.

The Compute Pragmatist: The whole spread hinges on one variable a business reader can actually watch: the share of economic tasks AI performs. In the modest scenario AI's impact resembles the internet; the extreme scenario has output doubling every 4.5 years, knowledge worker unemployment at 17.9 percent, and labor's share falling from 60 to 45 percent. The difference between "internet" and "nothing in history looks like this" is essentially adoption speed times task coverage. A major determinant of AI's impact will be how quickly it spreads through the economy. Skip the benchmark scores. Track what fraction of real work is actually being handed to models, and how fast that fraction climbs. Right now it's small. The extreme case requires it to reach roughly a third of tasks in under four years, which is a diffusion rate no prior technology has matched.

The Safety Lens: The distributional claim is the one that should keep buyers and policymakers up at night, and it's stranger than "jobs go away." In the biggest-impact scenario, labor's share of income falls from 60% to 45% by 2030; even though the economy is roughly a third larger, workers collectively earn as much as they would if AI did not exist at all. The entire gain accrues to capital. That's the political fault line. The model's own authors flag its limits honestly. The explorer is Version 1.0 and cautions that it isolates a few key forces while omitting many others: policy responses, business cycles, aggregate-demand or financial-market disruptions, and possible catastrophic risks. Leaving policy out is not neutral, though. It lets the model present a distributional crisis as a math result rather than a choice, then hand the fix back to government. Anthropic economist Peter McCrory says the scenarios are not predetermined. It's not an inexorable march, in his telling.

The Enterprise Buyer: For anyone signing AI contracts, the useful signal is not which scenario is "right." It's that a frontier vendor's own economists put the base case near the consulting consensus and the transformative case as an explicit outlier. That tempers the sales pitch. Anthropic published three scenarios; in the modest scenario AI's impact resembles the internet's. The public agrees, roughly. A separate Anthropic survey of nearly 11,000 US adults found median expectations align closest to the substantial scenario, with GDP roughly 10% higher by 2030. Reviewers also flagged the model can't see individuals. It does not follow individual workers, so it can only paint a very coarse picture of the costs of job displacement, and some questioned whether AI-exposed occupations will shrink at all rather than grow. Translation for procurement: plan around the middle, but don't let a vendor's dramatic top-end scenario justify a rushed workforce bet.

Where the real disagreements are. First: is the extreme case a forecast or a rhetorical device? The Researcher treats it as a legitimate corner of the parameter space; the Skeptic reads it as a way to relocate the CEO's own scariest claim into "outlier" territory. Both can be true. Second: does leaving policy and recessions out make the model cleaner or dishonest? The Safety Lens says omitting policy lets a distributional choice masquerade as arithmetic; the authors say it isolates the forces they can actually model. Third: the Enterprise Buyer and the Compute Pragmatist split on what actually matters: contract terms and the consensus middle, versus the raw task-adoption rate that decides which scenario you're actually living in.

What it hinges on. One number decides everything: the share of real economic tasks AI performs, and how fast it rises. Every scenario difference in GDP, unemployment, and labor share is downstream of that. It is also the most measurable and the most likely to disappoint versus the extreme case, because adoption inside real companies runs into workflows, trust, and integration cost, not model quality. For a business reader, the concrete move is to instrument your own task-automation rate: what fraction of your team's work is genuinely being done by AI this quarter versus last. That local number is your personal scenario dial.

Prediction: The share of US work actually performed by AI will stay far below the "extreme" scenario's near-one-third pace through the June 2027 update cycle of this model, tracking closer to the modest-to-substantial range, and no credible independent measure (US Bureau of Labor Statistics productivity data, Anthropic's own Economic Index, or Census business-adoption surveys) will show economy-wide task automation approaching a third by then.

Confidence: Medium — adoption diffusion is slow and measurable; the extreme case requires unprecedented speed.

Why: The entire spread in this model is driven by how fast AI reaches a large share of economic tasks, and the extreme +32.4% GDP case requires AI to cover close to a third of all tasks by 2030, a diffusion rate faster than the internet, electricity, or any prior general-purpose technology. Real adoption is gated not by model capability but by integration into messy company workflows, which historically takes years even after a technology works, so the near-term data almost certainly tracks the modest-to-substantial band that both the authors' own base case and their 11,000-person public survey already point to. Reaching a third of the economy in under a year would require enterprise deployment to break every historical diffusion pattern at once, which is why even Anthropic's own economists label that case an outlier rather than a forecast. The likely error is that reality comes in below the substantial case, not above it.

Revisit by 2027-06-30: We're right if BLS productivity data, the Anthropic Economic Index, or Census business-adoption surveys show AI performing well under a third of economic tasks and macro effects consistent with the modest-to-substantial range. We're wrong if any credible economy-wide measure shows task automation approaching one-third or GDP effects tracking the extreme scenario by then.