In the last six months the frontier AI labs have begun publishing economic policy papers about what happens if their technology sharply reduces the demand for human labor. Anthropic says that if AI delivers even a fraction of its potential, it could create unprecedented abundance while acting as a general substitute for labor. Its June framework sorts the future into three tiers of rising unemployment, the last beyond anything in the historical record. Alongside it Anthropic announced a $200 million Economic Futures Research Fund and a $150 million national fellowship program.1
The framework pairs each tier with a set of responses. At roughly 5 percent unemployment, most are meant to keep people working: wage insurance, training grants, job matching and licensing reform. At roughly 10 percent, it extends unemployment insurance and adds monthly payments for people whose benefits run out. Only at the third tier does it turn to new taxes, a possible basic income, sovereign wealth funds and equity sharing. Each step makes the question of who receives the gains more urgent, and the numbers become less concrete. At the last tier, Anthropic says it has moved beyond what economists know how to map, and that it is “not yet ready to advocate specific policies for this scenario.”2
The labs do their prognosticating very publicly, about a labor market their own products would weaken, and perhaps one day topple. For the millions who lose their wages, Anthropic proposes a public stake in the wealth its systems may create. A natural next question is how large that stake should be. It is not an exotic question, but Anthropic never answers it, and neither do most of the others.
I found seventeen AI economic-policy papers and statements that address distribution. Fourteen never say how large the public’s share should be; three do.3 I call this preparedness theater: institutions publicly building more and more elaborate preparations for a problem whose basic accounting is simple. The labs build this apparatus to advertise that they take the problem seriously.
What is simple about distribution
Roughly half the population does not work for pay at any given time. Most nonworkers are children, students, unpaid caregivers, the unemployed, disabled people and retirees.4 Most of these are stages nearly everyone passes through. Nonworkers still need income. Without some claim on what the economy produces they cannot buy food, pay for housing, or obtain the other things required to live.
This is one of the basic problems the welfare state exists to solve. We support children and caregivers with family transfers, the unemployed with unemployment insurance, disabled people with disability benefits, and retirees with Social Security. Matt Bruenig has spent much of his career explaining the welfare state in these terms: national income is paid to labor and capital, and some of that income is then routed to people who are not currently earning a wage.5
In 2023 Bruenig applied this directly to AI job displacement. His point was that technological displacement does not create a new economic category. The immediate problem is still lost labor income. His conclusion was aggressively unoriginal: “The policy solution to unemployment remains the same.”6 If AI leaves many more people without labor income for much longer, the same old problem gets larger.
The division between labor and capital is not fixed. For much of the postwar period, labor’s share of nonfarm business output sat between 60 and 66 percent. By the second quarter of 2026, after more than two decades of decline, it stood at 52.8 percent, the lowest reading in the BLS series going back to 1947.7 AI did not cause that decline. The point is that the division between labor and capital already changes over time.
The preparedness literature’s own severe scenarios move it much further. If AI substitutes for labor while producing the growth the labs forecast, less national income will be paid to workers and more will be claimed by owners of capital. There is also, on their account, much more income available. The resource problem gets easier, while the fight over who receives the gains gets bigger.
If labor’s share shrinks and capital’s grows, the transfers that once drew on wages have to draw on capital income, because that is where the income has moved. The policy literature often treats this shift as background rather than the central question, and leaves the public claim on capital income for later. Bruenig’s social-wealth-fund proposal answers this directly: public ownership of capital and a universal dividend from its returns, giving the public a durable claim on capital income.8
The question is simple: if less national income is paid to labor and more is claimed by capital, how much of that capital income should belong to the public?
What is complicated about preparedness theater
Anthropic has been specific about the institutional response it wants to build. Its $200 million research fund says it wants interventions that can “spread the gains broadly,” and adds that “the most promising solutions are ones nobody has tried yet.”9 The national fellowship sits beside the fund. So do the Anthropic Institute, the Economic Index, an expanding collection of advisers and outside researchers, and, since July, Ben Bernanke. The former Federal Reserve chair joined the company’s Long-Term Benefit Trust, whose members hold no equity in the company, and will contribute to its economic research. The potential of AI, he said on joining, is enormous, “and so is the range of outcomes.” How it plays out “will depend, in part, on the institutions we build around it.”10
There is plenty for this apparatus to learn: how to capitalize a public fund, how to keep capital taxes from being avoided, how the fund should vote its shares, and how much of its return to distribute each year. Bruenig’s own proposal spends pages on questions like these, all of which arise after deciding that the public should have a substantial claim on capital.8 Anthropic’s research agenda instead covers occupational-licensing reform, a retraining “fire drill,” AI-enabled job matching, relocation assistance, apprenticeships and new curricula. It also proposes trials of capital accounts and equity sharing, and studies of different tax bases.9 None of it says how large a public claim on capital these mechanisms should deliver.
Anthropic does want capital accounts funded partly with equity in AI companies, and says that “the earlier they are seeded, the more they are worth when they are needed.” How much equity is a natural next question. The framework does not say. Later it says benefits might converge toward a common floor without specifying the floor. When public budgets come under strain, it turns to reforms that need no new spending, including “private insurance mandates.” At the end Anthropic says companies like itself should be “ready and willing to pay our fair share.” It does not state the share.11
DeepMind manages to make the question considerably more complicated through unorthodox methods. Julian Jacobs and Alex Imas evaluate eleven household policies across several possible AGI futures. Alongside literature review and survey material, they build fifty-one AI agents from the past survey answers of economists on the Clark Center’s expert panel and ask the agents to rate the policies. The result is a series of color-coded decimal-precision tables scored from zero to one hundred. Universal basic capital receives 94.9 for Ownership of Gains. Negative income tax receives 57.0 for Democratic Voice. Universal basic income gets 46.2.12
I do not know what the 10.8-point difference between the last two numbers means. The paper does not translate it into anything outside its own scoring system. It did survey 2,019 Americans, but in its main results table their answers fill one column out of twelve. The simulated economists fill ten, including the one for Democratic Voice.12
The authors chose the policies, scenarios and criteria the agents were asked to rate. One feasibility score, “Ready,” is eyeballed: the public version of the paper says it includes an “author-coded” judgment of operational maturity, and their work forgoes any explanation of how that becomes decimal-precision figures like 95.3 and 48.4.12 Two of the economists whose simulated judgment helps score policy for a “full AGI economic transformation,” Alberto Alesina and Edward Lazear, died in 2020. It would be a mistake to hand any distributive decision to these scores.
Unorthodox methods aside, the paper’s severe scenario looks like the rest of the preparedness literature. Growth detaches from labor income while returns to capital rise, and the paper proposes universal basic capital, an ownership stake in that growth, as a backstop. It recommends designing the implementation details now so the program can launch quickly. Once again, the size of the stake is not among them.12
The paper also advises waiting. It warns that premature intervention could produce a “cure worse than the disease,” and it would trigger universal basic capital only once the data show the break has happened. Like many warnings in this literature, it does not say what the worse cure would be; the closest the paper comes is a passing reference to “unnecessary costs and market distortions.” Against those costs stands its own severe scenario, in which growth stops paying wages. Waiting for the data means acting only after people have lost their incomes. The rest of the literature shares the instinct: Anthropic’s first-tier supports are meant to be withdrawn as the labor market recovers, and OpenAI’s safety net switches on at thresholds it never defines and is built to phase out.13
This caution about protecting labor’s share of the national income, and with it the income of most of society, sits oddly next to the optimism DeepMind advertises about the technology itself. Demis Hassabis, who has said the end of disease may be within reach in the next decade or so, has described AI as potentially having ten times the impact of the Industrial Revolution and arriving perhaps ten times faster.14 What ten Industrial Revolutions in one-tenth the time would do to GDP is left as an exercise for the reader.
Dario Amodei speculates much further. In Machines of Loving Grace, he considers the point at which comparative advantage finally fails and humans may no longer contribute meaningfully to production. A large UBI might be only a small part of the answer. Perhaps AI systems operate a secondary economy and distribute resources according to their judgments of what humans should be rewarded for. Perhaps the economy runs on “Whuffie,” the spendable reputation currency from Cory Doctorow’s Down and Out in the Magic Kingdom. Amodei expects “lots of iteration and experimentation.”15
By The Adolescence of Technology, distribution is back in human hands. Amodei explicitly supports more robust progressive taxation in the high-growth, high-inequality world he anticipates, including the possibility of taxes aimed specifically at AI companies. But he also warns billionaires that if they refuse a “good” version, they will eventually get a “bad version designed by a mob.”16 He never says what makes the second version bad, or bad for whom. If the bad version is one that gives the public more say over concentrated wealth than Amodei has in mind, I suspect I am part of the mob.
How did fictional reputation currency remain inside the bounds of serious consideration while a strong democratic claim on concentrated wealth became mob politics? The contrast is sharper because Amodei is allegedly comfortable with enormous voluntary transfers. He reports that Anthropic’s co-founders have pledged 80 percent of their wealth to philanthropy.16 Giving away four-fifths of a fortune is numerically more radical than most tax proposals. Perhaps the important difference is control. Philanthropy leaves the owner in charge of the wealth and its allocation. Taxation makes the disposition of some of that wealth a political decision.
Across all three cases, preparedness theater grows around the same unanswered question: how large a public claim on capital are they actually prepared to support? Some of the preparedness literature admits that the obstacle is not a lack of knowledge. Windfall Trust came out of one scenario exercise saying that “the bottleneck is institutional, not intellectual.” Fathom describes the barriers as “largely political, not technical.”17
Geoffrey Hinton was less circumspect. Asked by British broadcast journalist Robert Peston how the gains from AI should be properly distributed, he answered, “Socialism.” Asked whether it was really that simple, Hinton said, “Yep.”18
There is no ground-truth value for the public’s proper share. Who gets what has always been settled through contests over power and resources, and none of these proposals will let us circumvent that political choice. An exotic technology does not require a correspondingly exotic politics. A simple theory of distribution will suffice.