Anthropic Maps a 21.5% Employment Collapse by 2030

Anthropic's economics team published a technical paper laying out AI scenarios for 2030 ranging from a benign case that lifts U.S. growth 1.6% above baseline to a worst case where output grows 15.4%, employment falls 21.5% and white-collar unemployment reaches 17.9%.

The company building one of the world's most capable AI systems has put numbers on what could go wrong. Anthropic's economics team has published a technical paper sketching how the U.S. economy might look in 2030 under different assumptions about how far artificial intelligence gets at doing work people currently do. The range it describes is extraordinarily wide — and the ugly end of it is very ugly.

In the mild case, AI turns out to be a useful but ordinary technology. U.S. economic growth in 2030 comes in 1.6% above where it would otherwise have been, and unemployment barely moves. In the severe case, growth hits 15.4% in 2030 and the economy starts doubling roughly every four years, while employment collapses by 21.5%, overall joblessness reaches 11.9%, and unemployment among white-collar workers climbs to 17.9%.

The paper refuses to say which branch is likely

Anthropic is unusually explicit about the epistemic status of the work. "The scenarios are not predictions and we attach no probabilities to them," the paper states. That caveat is doing real work. Without probabilities, the exercise is a stress test rather than a forecast — closer to what a bank does when it models a 1970s-style inflation shock than to what a macro desk publishes each quarter.

That framing is also the reason the document is worth reading rather than dismissing. A forecast invites argument about the number. A scenario invites argument about the mechanism: what would have to be true about model capability, deployment cost and corporate adoption for employment to fall by a fifth inside four years. The paper's value to a policymaker or an investor lies in the conditional structure, not the headline figure. The details were reported by Baystreet.

The distribution numbers matter more than the growth numbers

Strip out the drama of the unemployment figures and the most consequential line in the severe scenario is about who gets paid. Labour's share of national income falls from 60% to 45%, while capital income rises more than 80%. Office wages fall 11.5%.

That is the crux. An economy expanding at 15.4% and doubling every four years is, in aggregate, vastly richer. But a fifteen-point shift in the split between wages and capital returns is a change in the basic architecture of household income in a country where most people fund their lives from a paycheque rather than a portfolio. Tax systems built on payroll withholding, pension systems built on contributions from employment, and consumer demand built on wage growth all assume the 60% figure holds roughly steady.

It is also the figure most directly relevant to equity investors, and the most uncomfortable one. A world where capital's income rises more than 80% is, mechanically, a world of very high corporate margins. The paper is describing, in the same breath, a labour crisis and an extraordinary profit environment. Those are not contradictory outcomes — they are the same outcome seen from two sides of the income statement.

What it means for AI-exposed employers

The scenario with white-collar unemployment at 17.9% is not evenly distributed across industries. The categories most exposed are the ones where output is text, code, analysis and process: business services, back-office finance, legal and administrative support, and parts of the IT services sector. Those are also sectors where headcount is the dominant cost line, which is precisely why the substitution economics could move fast if capability arrives.

For investors, the practical question is which side of the trade a given company sits on. Firms that sell labour by the hour face a revenue model under pressure. Firms that buy a lot of labour internally face a cost base that could fall. Firms that sell the compute and models sit at the point where the capital share accrues. A single company can sit on more than one side of that at once, which is what makes the exposure hard to price.

Markets are not pricing a labour shock

Nothing in the tape on Friday suggested investors read the paper as an imminent threat. At the close on Fri, 11 Sep 2026, the S&P 500 tracker (NYSEARCA: SPY) finished at $764.29, up 0.85% from the prior close of $757.83, within a day range of $763.60 to $766.38. The Nasdaq 100 fund (NASDAQ: QQQ) closed at $714.88, up 0.87%, and the Dow tracker (NYSEARCA: DIA) closed at $525.79, up 0.97%.

That is a broad, calm, modestly positive session — the tech-heavy index and the industrial index moving in near lockstep. Equity markets have consistently treated AI as a margin story rather than a macro-political one, and the late-2020s question the Anthropic paper poses is whether those can stay separated. A capital share rising more than 80% would be, for a while, enormously good for the index. The employment figures attached to it in the same scenario would not stay outside the political system for long.

What to watch from here

Several things would help distinguish which branch of the scenario tree the economy is on, and none of them require waiting until 2030:

• Job postings in exposed white-collar categories. Hiring freezes show up in postings before they show up in the unemployment rate.
• Wage growth in office occupations specifically. The severe scenario has office wages falling 11.5%; the benign one has them broadly unchanged. That divergence is observable well before the endpoint.
• Labour's share of national income. The move from 60% to 45% would appear as a trend in the national accounts, not as a single jump.
• Whether corporate disclosure starts tying headcount plans to AI deployment. When companies begin justifying reductions by capability rather than by demand, the mechanism the paper describes is live.

The paper's real contribution is that it comes from inside the industry. A firm whose commercial interest lies in selling model access has published a scenario in which the same technology produces 11.9% joblessness. That does not make the scenario right. It does make it harder to write off as the work of outsiders who do not understand what the systems can do.

Key facts

• Worst-case U.S. growth, 2030: 15.4%, with the economy doubling every four years
• Worst-case employment change: Employment falls 21.5%; white-collar unemployment 17.9%; overall joblessness 11.9%
• Income split: Labour's share drops from 60% to 45%; capital income up more than 80%
• Market close, Fri 11 Sep 2026: SPY $764.29 (+0.85%); QQQ $714.88 (+0.87%); DIA $525.79 (+0.97%)

Frequently asked questions

What did Anthropic's economics team actually publish?

A technical paper laying out a range of scenarios for how the U.S. economy might look in 2030 as AI becomes more capable of performing human tasks. It spans a benign case, in which AI adds 1.6% to growth and unemployment barely moves, and a severe case with 15.4% growth alongside a 21.5% collapse in employment.

Is the paper a forecast of mass unemployment?

No. The paper explicitly states that "the scenarios are not predictions and we attach no probabilities to them." It is a structured stress test of possible outcomes rather than a projection, which means the severe figures describe what could follow from certain capability assumptions, not what Anthropic expects to happen.

How bad is the worst-case scenario for workers?

In that branch, total employment falls 21.5%, overall joblessness reaches 11.9%, and unemployment among white-collar workers hits 17.9%. Office wages decline 11.5%. The paper pairs those figures with very rapid output growth, so the economy becomes richer while the gains shift away from wages.

What happens to the split between wages and profits?

In the severe scenario, labour's share of national income falls from 60% to 45%, while capital income rises by more than 80%. That fifteen-point shift is arguably the most consequential number in the paper, because payroll taxes, pension contributions and consumer demand are all built around wages remaining the dominant income source.

Did markets react to the paper?

There was no visible reaction. On Friday 11 September 2026, the S&P 500 tracker SPY closed at $764.29, up 0.85%; the Nasdaq 100 fund QQQ closed at $714.88, up 0.87%; and the Dow tracker DIA closed at $525.79, up 0.97%. All three finished broadly higher in a calm, correlated session.

Which sectors would be most exposed if the severe scenario played out?

Occupations whose output is text, code, analysis and administrative process are the most directly substitutable: business services, back-office finance, legal support and parts of IT services. These are also sectors where headcount is the largest cost line, which is why the substitution economics could move quickly if capability arrives.

Sources

• Anthropic Warns That A.I. Could Lead To Mass Unemployment — Baystreet

Photo: Mikhail Nilov · Pexels Licence — source