Anthropic publishes three scenarios for 2030

The story that went around

On September 9, The Decoder ran a short piece under a headline that left nothing to interpretation: "Anthropic built an economic model that frames its CEO's bleakest job forecasts as an outlier scenario." It got picked up, summarized, aggregated. The story is a good one, and it has the shape of a small family betrayal. It also rests on a ranking the paper never makes.

Since May 2025, Dario Amodei has been repeating the same warning: by 2030, AI could wipe out up to half of all entry-level office jobs and push US unemployment to somewhere between 10 and 20 percent. Those numbers do line up with the harshest of the three scenarios in the new paper. The piece draws the conclusion out loud: "So Anthropic's own model is recasting its CEO's predictions as the most unlikely outcome or, dare I say, alarmist."

The document is free to read. It runs 57 pages, it's titled "Economic Scenarios for Transformative AI", and it carries five signatures, including Anton Korinek's and Charles I. Jones's. We opened it. It ranks none of its three scenarios, modest, substantial and extreme, and it takes the trouble to say so twice.

The sentence on page 3

In the introduction, before the model is even described, the authors set the rules of the game: "The scenarios are not predictions, and we attach no probabilities to them; their purpose is to make the consequences of different assumptions comparable."

This isn't boilerplate buried in an appendix. It's the second paragraph of the introduction, at the exact spot where a reader looks for what a document claims to be doing. A menu that lists three dishes doesn't tell you which one is best, and this one says so before it serves anything.

The conclusion, on page 39, shuts the same door in the same terms: "Today, none of the three can be ruled out." The ranking exists nowhere in the document. It was added on the way out.

The name that isn't in there

We searched the 57 pages for "Amodei". Zero hits. The CEO isn't quoted, referenced, discussed, or so much as mentioned in a footnote.

Better still: page one carries a line nobody picked up. "The views expressed in this paper are those of the authors and do not necessarily represent the views of Anthropic or The Anthropic Institute." That disclaimer is standard academic furniture, but here it has a direct consequence. Talking about "Anthropic's model" that "reframes its CEO" hands the company a position the document expressly declines to hand it.

What the three scenarios actually say

The figures the press ran with, on the other hand, are accurate. The model projects the US economy in 2030 against a path with no AI in it.

In the modest scenario, GDP ends up 1.6 percent above that path and unemployment among knowledge workers barely moves. In the substantial scenario, GDP lands 8.3 percent higher, cognitive employment falls 3.9 percent, and unemployment in those occupations goes from 2.9 to 4.5 percent. In the extreme one, GDP finishes 32.4 percent above the no-AI path, growth hits 15.4 percent a year, cognitive employment drops 21.5 percent, and unemployment among those workers reaches 17.9 percent.

The authors hand you the scale: at that pace, income per head would double every five years, against every 35 years in the US over the past century. Labor's share of income would slide from 60 to 45 percent, meaning 15 percent of GDP that stops being paid out in wages and starts being paid out as a return on capital.

Three markers, not a bet

If the authors won't rank them, it's because their scenarios aren't bets. They're reference points borrowed from the existing literature. They write it down: the three "broadly correspond to the wide range of published predictions."

Modest is calibrated on Daron Acemoglu's work and OECD figures. Substantial picks up 2023 forecasts from banks and consultancies, McKinsey included. Extreme sits well past that: the authors line it up against the literature on growth take-offs driven by general AI, and against narrative accounts like AI 2027.

Which is to say the extreme scenario isn't some outlandish idea the model pushes away. It's the top of a range that had already been published. The paper doesn't invent it in order to disqualify it, it borrows it to bound the interval. A thermometer marked up to 120 degrees isn't calling for a heatwave.

The one new number that went unnoticed

The paper does contain something genuinely new, and almost nobody picked it up. The authors surveyed 10,980 US adults through Morning Consult between August 11 and 23, 2026: when AI will be able to do each of eight named tasks, how much of what it can do will actually get used, how much time it saves, whether it works alone or alongside a person, and how many months a displaced worker needs to land somewhere else.

Feed the median answers into the model and the public comes out close to the substantial scenario. But the spread is the real finding. About 30 percent of respondents think AI saves no time at all on a task it's suited for, while 49 percent think it cuts that time in half or better. On a Nobel-level scientific discovery, 40 percent answer "never."

Those are two populations living in different worlds, and the average of the two describes neither.

The 17.9 percent is a dial setting

Which leaves the number that made every headline. It comes with a warning, and the warning is in the paper itself.

That 17.9 percent unemployment rate among knowledge workers rests on one parameter, wage rigidity, set to 0.5 in the baseline. The authors move it in the same document without touching anything else. With fully flexible wages, unemployment drops to 2.6 percent, but cognitive wages fall by more than 42 percent. With very sticky wages, unemployment climbs to 24 percent and pay holds up.

Their own phrasing is blunt: the cost gets paid in wages or in unemployment, and the baseline splits the bill across both. The dial doesn't decide whether there's pain, it decides where the pain comes out.

The authors are just as upfront about their limits. The framework ignores catastrophic risk, business cycles, financial crises and political economy. It models no robotics at all, which is a large part of why it stops at 2030.

What's left

So the paper doesn't say Dario Amodei is right, and it doesn't say he's wrong. It offers a way to turn disagreements about AI into disagreements about four or five measurable parameters, and it notes that the data will settle those on their own within a year or two.

That's less of a story than a CEO disowned by his own economists. It's also the only thing the document lets you claim, and its authors took the trouble to write it down twice.