There’s a new paper doing the rounds that a lot of thoughtful people are treating as good news. It’s called “Economic Scenarios for Transformative AI”, it comes out of the Anthropic Institute, and its lead authors are among the best macroeconomists working today - Chad Jones at Stanford, Anton Korinek, and colleagues. I want to say up front that this is clearly a serious, careful piece of work, and I am not here to try to tell you it is wrong.
I want to do something that’s harder than that. I want to take the model completely seriously (grant it every assumption, use only its own numbers) and show you that the calm feeling people are walking away with does not survive contact with what the paper actually says. If anything the model is built to be cautious: it deliberately leaves out the wilder possibilities, including the one where AI starts rapidly improving itself and races ahead. Even leaning conservative, here is where it lands.
Because here’s what really stuck out for me. The authors built three scenarios and labelled the middle one “substantial”. That word is chosen, and it just sounds measured. It is the central case, the one their public survey lands on, the one you are likely to file under “significant but manageable”. And in that supposedly moderate middle scenario, the share of the economy’s income that goes to workers rather than to the owners of the capital and machines (economists call this the labour share) falls from 60 per cent to 56 per cent. Four points, in four years.
To put that in perspective (and this is the authors’ own comparison, not mine) that four-point drop is roughly the entire decline in the US labour share across the four decades after 1980, compressed into a single presidential term. Over the same window, unemployment among knowledge workers rises from 2.9 per cent to 4.5 per cent - up by more than half. That is the reassuring scenario. Their word for it is “substantial”. I think a more realistic reading is: the reassuring case is an alarming one.
And it gets sharper. The paper’s own extreme scenario has the labour share falling to 45 per cent - which the authors describe, in plain words, as 15 per cent of the whole economy’s income moving from workers to the owners of capital. They are not hiding any of this. It is all right there in their tables. So the puzzle is not “what are they concealing”. The puzzle is:
If the numbers are this dramatic, why does the paper land as a calming influence?
I think there are three reasons, and clearly none of them are dishonest. But all three do make a dramatic result feel more moderate than it likely is.
Why it doesn’t feel dramatic
The first reason is how “substantial” got crowned the central case. The authors ran a survey of nearly 11,000 US adults, and the median response maps onto the substantial scenario. So “substantial” gets presented as the case the public expects - which quietly borrows the authority of a democratic majority, or the wisdom of the crowd.
But look at what that survey actually is. It is a very wide, very flat spread of opinion from people with no special knowledge of how fast AI will be adopted or how much work it will take over. Around 30 per cent of them expect AI to save no time on a task it is suited to. Around 49 per cent expect it to cut the time at least in half. Those two groups are describing different worlds. When opinion is scattered that widely, the median does not tell you what people collectively believe - it just tells you where the middle of a wide scatter happens to sit. And a wide scatter’s middle lands in the author-chosen middle bucket more or less by construction.
There is a second problem with leaning on that survey. Even where people do have a view, they do not seem to believe it about themselves. Asked about a list of professional tasks, respondents expected AI to reach 71 per cent of them. Asked about their own work, the figure dropped to 27 per cent. That is not a coherent forecast. That is a gut reaction that says:
AI will transform everyone’s job but mine
And the deepest issue is what the survey was actually measuring. People were asked about capability - how far and how fast AI will spread through tasks. That is the input to the model. They were never asked about the reassuring part - whether displaced workers get smoothly re-absorbed, whether wages rise outside knowledge work. Those are the model’s outputs, produced by its internal machinery, and no member of the public voted on them. So “the public expects substantial change” is, at most, a statement about the expected path AI takes. It says nothing about whether the gentle landing the model attaches to that path is the likely one.
To be scrupulously fair, the authors are candid about all of this in the body of the paper. They show the spread. They say plainly that these are not predictions and that none of the scenarios can be ruled out. The softening does not happen in the data - it happens in the emphasis, in the headline, and above all in what a busy reader or a press write-up carries away. But candour in the body is not the same thing as neutrality in the takeaway. And on the model’s own numbers, “substantial” is not moderate.
Even that understates the pain
Here is the part I find most important, and it stands even if you grant the model’s entire story about where new demand comes from.
The model’s comforting message is that displaced knowledge workers get re-absorbed, and that as AI makes office work cheap and abundant, wages rise in the work AI cannot do - the trades, construction, care work. On the model’s own logic, that wage increase is real.
The question the model cannot answer is: rise for whom?
Those higher wages flow to the people already doing that work - the incumbent electricians, the nurses, the carers who were there before any of this started. They do not flow to the laid-off analyst, because the analyst cannot do those jobs. A product manager made redundant in San Jose does not become a licensed electrician in rural Virginia. She needs an apprenticeship she has not done, a licence she does not hold, and she needs to move across the country and sell a coastal house to get to work that pays less than she used to earn. For a large share of displaced office workers, that is not a small delay you can wave away - it is a wall.
What actually happens to those who cannot cross it? They stay unemployed, or they crowd downward into the lower-skill work they can do (delivery, warehouse, retail) which pushes wages in those jobs down. So underneath the single cheerful “wages rise outside knowledge work” number, there are two very different stories: incumbents doing better, and the displaced doing worse. The model gets its reassuring average only because it treats every worker as interchangeable and treats the whole country as one frictionless labour market with no distance, no licensing, no lost experience.
And the one place where the model’s story clearly works (where the money from AI-driven growth genuinely does fund new manual jobs) is building the data centres. Those jobs are real. But they are being built in cheap-land, cheap-power regions far from where the displaced workers live, and wiring a data centre is something you do once. It is a construction phase sugar hit, not a permanent employer. So even the model’s best re-absorption story points to the wrong places, for the wrong people, on a temporary clock.
And it stops the clock early
The third softener is the calendar. The model runs to 2030 and no further, and the authors are honest about why:
Robotics and physical-task automation sit outside their framework,
and rather than guess, they stop
But look at what that does. The entire re-absorption story leans on manual and physical work being a safe refuge - the place cheapened office work flows toward. Robotics is precisely the thing that starts removing that refuge, and it sits just past the edge of the window. So does the deepening of adoption, where displacement tends to concentrate as a technology matures. The model stops at 2030 not because the story ends there, but because the harder chapters begin there. That is a good reason to be careful, and it is a very good reason not to read the comfort past the window. Beyond 2030 is now very very close!
Where that leaves us
Obviously, none of this makes Korinek and Jones wrong. Their model is a robust and legitimate benchmark, GDP really does rise in it, and they never promised a painless transition - they produce rising unemployment, and they concede that if capital is less abundant than they assume, average wages can go negative.
What it does mean is that the reassuring limb of the paper (displaced workers re-absorbed, everyone lifted by rising wages elsewhere) is far more fragile than the alarming limb. The alarming numbers come straight from the model. The comforting story rests on assumptions that bend in exactly the direction the paper’s own scenarios are heading.
So on the axis that actually matters to ordinary workers (who ends up with the income, and who actually gets re-absorbed), I would read “substantial” not as a central, most-likely estimate with room to move either side of it. I would read it as a floor for concern. The honest version of the headline is not “the experts expect a substantial but manageable shift”. It is the one the model’s own tables already tell you, if you let them.
The reassuring case is an alarming one.