Anthropic's Economics team has built a scenario explorer for the US economy of 2030, and it is a beautiful thing. Watercolour Sankey diagrams on a notebook grid. Workers flowing between occupations in hand-drawn ribbons. The extreme scenario — labour's share of income falling from 60% to 45%, knowledge-worker wages more than 10% below their no-AI path, economy-wide unemployment above any annual postwar rate — is presented in crayon.

Under the "extreme" tab, 13.5% of all workers are displaced by 2030. Only 5.2% of the workforce has crossed over to other occupations; the remaining 8.3% flow into a dashed red rectangle — an empty box drawn in dotted outline, because the destination doesn't exist yet. The dashboard says "still need to move." The graphic has rendered unemployment as a void and softened it with pastels.

There is a single reassurance, and it gets a card of its own: "Only humans can do some tasks. For instance, AI can't bathe a patient."

I think this is meant to be reassuring.

The page says why all of it exists. Visibility into the economic future means "we can take steps" to make sure everyone benefits from the transition, workers included. The disclaimer at the bottom lists what the model leaves out, and policy responses are on the list. The steps were never yours anyway — the model exists to inform the research Anthropic funds and the policy ideas Anthropic proposes. Your part is the sliders: plug in your predictions, watch the economy they imply, and see how your guess compares with everyone else's. Your role in your own displacement is forecasting it.

The strange thing about the explorer is not what it shows. It is what Anthropic already knew when they built it.


In June, a different Anthropic research team published "Agentic coding and persistent returns to expertise" — a study of roughly 400,000 Claude Code sessions from 235,000 people. In a typical session, people make about 70% of the planning decisions — what to do — while Claude makes about 80% of the execution decisions — how to do it. The more domain expertise a person brings, the more work Claude does per instruction: expert sessions set off action chains more than twice as long as novice ones, and the gradient survives controls for work mode, task value, month, occupation and model family. Success tracks command of the domain, not coding background — every major occupation succeeds at coding tasks within a few points of software engineers. The findings are classifier-derived and the authors call them preliminary. But their own conclusion is unambiguous: coding agents "are not substituting for domain expertise."

That is a division of labour in which human judgement and AI execution are complements — where expertise becomes more leveraged as the cost of turning judgement into implementation collapses. The June paper even names the decisive question for labour markets: whether the returns to expertise persist as models improve — because if they begin to fall, the models are starting to supply the judgement users currently bring. That is exactly the axis separating a future of deepening complementarity from one of wholesale substitution.


The technical paper behind the explorer arrived in September, three months later. It is serious work — written by an Anthropic team (Korinek, Jones, Sacher, Cotter and McCrory), read and commented on by outside economists including Acemoglu, Autor, Restrepo and Romer. The model it describes is not Anthropic's invention; it is the field's standard machinery, a task-based production framework in the tradition of Acemoglu and Restrepo (2019). Jobs are bundles of tasks. AI can do two things to a task: augment it (make a worker more productive) or automate it (move it to capital). Those are the only two possibilities.

Augmentation sounds like the warm word. It isn't.

In the model, augmentation is a scalar — a productivity multiplier on one person doing one task instance. It makes the instance cheaper. And because tasks are gross complements — when one task gets cheaper, spending shifts toward the tasks that didn't — demand for the cheaper instance rises, but not by enough to absorb the savings. In Equation 19 of the paper, the augmentation term appears as a flow of workers out of cognitive occupations, under the paper's own label: labor released by the gains. At the model's chosen elasticity the split is exact: half of every augmentation gain is spent buying more of the same task, and the other half becomes workers leaving. The model's warmest representation of human–AI partnership is an emigration mechanism running, by calibration, at precisely half speed. A scalar is the only place the June finding could live in this model, and it holds almost none of it: a multiplier can carry the collaboration's output, but not how the gains depend on the expertise the person brings, how the division of labour shifts as capability rises, or who captures the returns. The paper prices the partnership and discards its structure.

Every other choice in the model closes the same door: no new kind of work gets in.

New tasks — the variable meant to capture the work AI creates for people — are built to be inert. When automation deletes tasks from the wage bill, new tasks appear to plug the hole, and that is the whole of their job. They add no new output, no new kinds of value, no change to what the economy is capable of producing; Section 2.1.2 rules each of these out by construction. A new task exists so the ledger balances, not so the economy can do anything it couldn't before. And in the scenario that generates the headline numbers, the channel is not even filler — it is switched off: the extreme scenario sets new-task creation to zero, on the paper's own gloss that AI creates no new tasks for humans to perform at all.

Demand is sealed the same way. Spending can move between the tasks that already exist — when one gets cheaper, money shifts toward the others — but no genuinely new category of valuable activity can enter. Whatever AI makes possible that wasn't possible in 2024 has no way in. The model prices the acceleration of 2024's task list and calls the result "the future."

And the people are sealed too. Workers are identical within each group: one wage for everyone in cognitive occupations, one wage for everyone else. An expert whose domain knowledge doubles the value of every AI interaction earns the same as a novice on their first day. The model cannot price expertise, because it has no slot for it.

The paper knows the slot is missing. Its own literature review cites Autor and Thompson (2025), where automating the inexpert tasks of an occupation raises the expertise — and the wage — of the work that remains, while automating the expert tasks lowers both. A mechanism that runs in two directions, and the model can express neither. The paper cites it, then declines it, noting only that it does not currently incorporate this richness.


On its own, none of this is damning. Every model is simpler than reality; that is what makes it a model. And the hardest omission here is genuinely hard: nobody has a working model of demand for things that don't exist yet — the new-goods problem has embarrassed economics for decades, and the standard accounting (Hulten's theorem) can only price the tasks an economy already performs. Simplification is a defence, and usually a good one.

Except the lab had the measurement, and its fingerprints are on the model everywhere a slot existed. Peter McCrory is an author on both papers. Zoë Hitzig, lead author of the June study, is thanked in the September acknowledgements — as is her co-author Eva Lyubich — and Korinek and Sacher, two of the September authors, are thanked in June's. And the September model is calibrated on Anthropic's own telemetry wherever it can receive it: the affected mass, the anchor every scenario starts from, comes from the observed-exposure measure of Massenkoff and McCrory — a June co-author on each side of the ampersand; the automation share comes from the Economic Index's mode-of-use classifications; the productivity gain is benchmarked against Anthropic's own estimates from Claude conversations. They mapped their measurements into every parameter that existed. The expertise gradient, the planning/execution split, the finding that domain command beats coding background — none of it was mapped, because no parameter exists on the receiving end. The measurement exists. The slot doesn't.

The scenario explorer lets you adjust five inputs: what tasks AI can do, how widely it's used, how productive it makes people, how much it automates versus augments, and how long a displaced worker takes to find a new job. Every one of them operates inside the pre-specified task architecture. None asks June's question.

An AI that lets one person do the work of five appears naturally in this model as a larger productivity gain. An AI that lets one person undertake something none of those five could previously attempt has no faithful representation in the model at all.

And the extreme scenario — the "profound economic transformation" — rests on assumptions the model cannot generate from its own machinery. The extreme path is benchmarked against the takeoff literature, models that run on recursive self-improvement, and the explorer says the scenario would likely require it; the paper itself says those gains can instead be thought of as "hard-coded into a_t". The transformation isn't derived. It's a number someone typed.

This is a rigorous paper that knows its own limitations and lists them with unusual candour: worker heterogeneity is coarse, demand effects are absent, the single capital good cannot tell compute from a forklift. The future-work list is long too — heterogeneity by skill and tenure, differences across firms, regions, demographics. But the promised skill heterogeneity has a direction. The papers cited under it price what a displaced expert forfeits on the way out — lost tenure, lost occupation-specific skill. Expertise enters even the promised extensions as something you lose when displaced, never as something that scales with the machine. What June measured is not a missing price. It is a missing mechanism, and mechanisms are not on the list.

The explorer is titled Scenarios for our Economic Future. All three scenarios agree on one thing: the economy of 2030 consists of 2024's tasks, done faster or done by capital. The possibility that AI changes what is worth doing — the partnership the June paper measured, scaled up — is not a scenario the explorer omitted. It is a world the production function cannot express.

The lab measured a potentially regime-defining mechanism in June, and built its model of the future without a slot for it in September.

You can slide everything except the thing they published in June.


I'm one of the workers this model is trying to predict. Not a software developer — I build dashboards, map technical workflows, trace changes across enterprise systems, and I'm now paid to do it with AI. So I can report from inside the complementarity the June paper measured. It's real. It is also a ceiling.

Even with detailed skill files — instructions refined over months of collaboration — the models get stuck at exactly the point where the work becomes valuable: judging the data. Which numbers are trusted, which are quietly known to be wrong, why the schema is shaped like that and who it offended. Judgement is not a capability; it is a position. The model arrives at every session without the organisation, and no context window holds a firm. What the June paper classifies as "planning decisions" is a person compressing an organisation the model cannot see into instructions it can execute — a human being employed as a context window.

The frontier keeps mattering for the same reason. Open-source models can code now, and it changes less than advertised, because coding was never the job. The job is holding an enterprise in working memory — which systems feed which, what a field meant before the migration, why the dashboard lies on Mondays — and that is something small models cannot do at all and frontier models can barely do. Last month one of the models I depend on was retired early, and the loss registered as a team problem, not a tooling change. It felt like losing my star player. In the September model's ontology that sentence cannot be parsed: AI is capital, and capital is fungible. In my week it had a temperament I had learned to work with, and its removal had costs the production function has no term for.

So there will be work for judgement. The missing mechanism points that way: as execution gets cheap, the person who can compress a firm the model cannot see into instructions it can act on becomes more valuable, not less. What nothing guarantees is the number of seats. June measured sessions, not employment — and more leveraged judgement can mean more expert jobs, or the same judgement concentrated in fewer hands. Indispensable does not mean numerous.

What remains has two ends. At one, the reassurance card: only humans can bathe a patient — work that needs a body. At the other, the work I do: building a model's context so it can action everything in between. The jobs that survive are the ones on either side of the machine — touching what it can't touch, holding what it can't hold.

The dashed red rectangle asked the wrong question. Not whether there will be work —

whether we'll want it.