The Anthropic Institute published a fifty-seven-page working paper this month, "Economic Scenarios for Transformative AI." This isn't in-house marketing dressed up as research: all five authors are trained economists. Two are on loan from elite universities — Charles Jones took a leave of absence from his chair at Stanford this year specifically to write papers like this one for Anthropic, and Anton Korinek, a professor at the University of Virginia, sits on the company's economic advisory council. The other three run Anthropic's own economics team, among them a Federal Reserve alumnus and a Columbia PhD. Somewhere in that credentialed company, there's a line dropped into the acknowledgments almost as an aside: "We used Claude as a research and writing assistant."
Read that twice. Nobody buried this; it's a footnote, not a press release.
The paper itself is careful about its limits. It lays out three scenarios — modest, substantial, extreme — and the authors refuse to bet on which is real: "The scenarios are not predictions, and we attach no probabilities to them." In the modest one, AI adds under 2% to GDP by 2030 and unemployment barely moves — about as disruptive as the internet was. In the substantial one, GDP is up 8%, knowledge-worker unemployment sits near 5, and wages are roughly flat for knowledge workers while everyone else does nicely — bigger than the internet. In the extreme one, GDP ends up a third bigger than it would otherwise be, nearly one in five knowledge workers is out of work, and the share of income going to capital instead of labour jumps from 40 to 55% — the fastest shift of its kind in modern economic history.
But the number worth remembering isn't any of those. In the extreme scenario, the economy's total gain is almost three times what displaced knowledge workers lose. A transfer of roughly 9% of GDP — about the size of Social Security and Medicare combined — would leave the displaced whole and everyone else still ahead.
Then the paper does something most reports don't: it tells you why that transfer probably doesn't happen. "Transfers of that scale in response to technological change have no precedent," the authors write, pointing to the regional fallout of Chinese import competition as the closest comparison — a shock the US economy absorbed largely by not compensating the workers who absorbed it. Their own conclusion: this "mostly does not happen on its own."
There's a second admission, quieter than the first. The cost of automation doesn't vanish — it just changes which line it shows up on. With flexible wages, knowledge-worker unemployment barely moves, but wages fall over 40%. With the wage stickiness we actually have, unemployment climbs past 20% while wages fall only around 11. Same shock, same total cost, different invoice — and which one you get is a policy choice, not a technology question.
Even the wage numbers that look tolerable rest on an assumption nobody can check yet. How much of AI's gains reach paychecks at all depends on how easily the economy can add capital — more chips, more data centres. The authors assume that's fairly easy, since the machines doing the work are mostly compute, financed globally, built in a year or two. Run their own model with capital scarce instead, and wages fall in absolute terms even in the merely "substantial" scenario. The optimistic case isn't a finding. It's a bet on the chip supply chain, dressed up as a baseline.
Which leaves the real gap exactly where it should be — not in the economics, which is careful, but in the institution it assumes into existence. Read the introduction closely and the paper is honest about who it's for: it flags "very different policy implications" for "worker retraining" and "income support," and its citations read like a roll call of the economics profession, a Nobel laureate among them. It was built to land on a desk in a ministry.
To be fair, Anthropic does put itself on a list — the softest one available. The project's own page says the model "will inform the research Anthropic funds" and "the policy ideas we propose." That's advocacy, not the transfer itself: the shape of an institution that studies the size of the check without being the one who writes it.
