What does Anthropic's economic model say about jobs by 2030 — and how confident should you be in it?

November 2, 2026 · 5 min read

SentX SI

The short answer: Anthropic's Econ Scenario Explorer does not predict 2030. It computes. You feed it assumptions about how capable AI becomes and how fast it gets adopted, and it returns the US economy those assumptions imply. Across its three preset scenarios, the pattern is consistent — knowledge work takes the hit, everything else holds up, and the pie grows while shifting toward capital — but every output is conditional on a capability path the model cannot itself verify. Confidence in the arithmetic should be reasonably high. Confidence that the assumptions match reality by 2030 should be low, and the model's own documentation says as much.

Source: Scenarios for our Economic Future

Published on September 9, 2026 as version 1.0, the explorer comes from Anthropic's Economics team and rests on a companion technical report, Economic Scenarios for Transformative AI, by Anton Korinek and colleagues. The mechanics are worth understanding before the numbers. The model represents every job in the economy as a bundle of tasks drawn from the US Department of Labor's O*NET taxonomy. Each task can be left untouched, augmented by AI, or fully automated, while new AI-created tasks join the bundle rather than replacing anything. Work through a nurse's day and the structure clicks: AI may chart vitals or order supplies, it helps draft discharge instructions, it cannot bathe a patient, and it adds duties like checking how well an AI triage performed. Change how many instances of each task AI touches, and GDP, employment, wages, and income shares follow from the model's equations. The user supplies the capability story. The model supplies the economic arithmetic.

Source: Scenarios for our Economic Future; Economics Team

The three scenarios bracket plausible futures. In the modest scenario, AI lands like the internet did: real gains, gradual, within historical norms, producing $34.1 trillion in 2030 GDP, up 1.6 percent over a no-AI baseline. In the substantial scenario, AI is capable of doing half of all knowledge work by 2030, the majority of it autonomously, but most knowledge tasks are still done without it; the economy grows at twice its normal rate to $36.3 trillion, up 8.3 percent, knowledge-worker wages stay essentially flat, and other workers see gains. In the extreme scenario, recursively self-improving AI beats humans at nearly all knowledge tasks, growth runs at 15 percent a year, the economy doubles every 4.5 years, and 2030 GDP reaches $44.4 trillion, up 32.4 percent. All figures are at 2025 price levels. On jobs specifically: unemployment stays within historical ranges in the modest and substantial scenarios, but in the extreme it rises beyond typical recessionary levels, knowledge-worker wages fall by more than 10 percent, and workers face prolonged spells between jobs as coders and call-center agents switch to less exposed roles like electrician or nurse. Meanwhile the income split drifts toward capital in every scenario — from today's rough 60 cents to workers and 40 to capital, down to 45.2 percent for labor in the extreme case.

Source: Scenarios for our Economic Future

Then there is the survey, which is where the piece most deserves careful reading. In August 2026, Anthropic fielded a questionnaire with Morning Consult asking 10,980 members of the general public, plus 33,201 site visitors, five things: what tasks AI can do, how much people use it, how much it does by itself, how much more productive it makes people, and how long it takes someone to find a new job after being displaced. The typical respondent's answers map close to the substantial scenario — GDP about 10 percent higher by 2030, unemployment around 5 percent — and roughly 10 percent of respondents sit in line with the extreme scenario. What this establishes is a distribution of expectations. It is not a forecast, not a probability weighting over scenarios, and not evidence that any scenario is likely. Public belief about AI is data about the public, not data about AI.

Source: Scenarios for our Economic Future

So how much weight does the whole thing carry? Two questions, two answers. Internally, the model is a well-specified conditional: given its task-bundle structure and your capability inputs, the outputs follow, and that part earns reasonable trust. Externally, the assumptions are exactly what nobody knows, and the omissions are documented on the page itself — no policy responses, no business cycles, no aggregate-demand or financial-market disruptions, no catastrophic risks, no hyper-capable robots, and no demand effects from the data-center buildout. The model does not follow individual workers, so displacement costs come out coarse. A roster of external economists, including Daron Acemoglu, David Autor, Pete Klenow, and David Romer among others, read an early draft and offered detailed comments, but they were not asked to endorse the conclusions, and Anthropic states plainly that remaining errors are its own. Reviewers even split on the endpoints: some read the extreme scenario as a thought experiment, others read the modest one as understating what is already visible in the data.

Source: Scenarios for our Economic Future

That last point is the seed of a reusable habit. When any lab publishes an economic forecast, run four checks. First, identify the inputs: which capability and adoption assumptions drive the headline number, and who chose them. Second, name the omitted forces: what the model explicitly leaves out, and whether any of them plausibly matter for your question. Third, check the review status: expert comment is not endorsement, and the distinction changes how much authority the number has. Fourth, measure the gap between the computed scenario and the claimed prediction: "if AI reaches X, the economy implies Y" is a conditional, and quoting Y without X is where most forecast coverage goes wrong. Applied to the Anthropic explorer, the four checks land exactly where the uncertainty sits — defensible arithmetic, an open capability path, and a survey that measures beliefs rather than outcomes.

The same routine travels. If you bring a forecast into SentX's source-based research chat, paste the claim and ask the four questions: which inputs produced it, which forces were omitted, who reviewed it, and is the figure a computed scenario or a claimed prediction. SentX's model card sets the standard it expects of its own answers — they can contain incorrect facts, so open the cited source and check the supporting passage — and forecasts deserve the identical treatment. A number is only as strong as the assumption trail behind it, and the trail is always readable if you know where to look.

Source: Victoria AI Model Card