There is a famous painting by Francisco Goya (roughly 1820s) of Saturn Eating His Son – haunting and terrifying in equal measure. Reading Hemenway, Falk, and Tsoukalas (2026) article on the role of AI, and automation more generally, in the destruction of demand reminded me of that image.
There is a version of the AI automation debate that treats displacement as an inevitability — something that happens to workers, after which we argue about how, or even if, to compensate them. Universal basic income, sovereign wealth funds, equity stakes: the policy conversation is almost entirely about what to do with the proceeds once the automation has occurred.
That framing misses a deeper problem – the displacement isn’t fated – it is the aggregate result of individually rational decisions that, taken together, produce an outcome that is bad for everyone — including the firms making those decisions.
This is the prisoner’s dilemma of AI automation, and it deserves some reflection.
The demand commons
Here is the basic mechanism. When a firm automates a worker’s job, it captures the full labor-cost saving. That saving is private — it goes entirely to the automating firm’s margins. But the revenue loss from that worker no longer spending their wages spreads across every competing firm equally, because displaced workers are also consumers, and their lost purchasing power flows through the entire market, not just the firm that replaced them.
Each firm, in other words, imposes a cost on the shared demand pool — what we might call the demand commons — that it cannot individually internalize. The rational move for any single firm is to automate. The consequence of every single firm automating simultaneously is to destroy the demand base that sustains all of them.
Hemenway Falk and Tsoukalas formalize this and show that the Nash equilibrium produces more automation than the cooperative optimum. Crucially, the outcome is Pareto-dominated: both workers and firm owners would be strictly better off if firms could cooperate. They cannot — because any firm that holds back while rivals automate simply loses market share. Automating is a dominant strategy even for a firm that can see the demand cliff ahead.
Two features of this mechanism deserve emphasis. First, more competitive markets make it worse, not better. As the number of competing firms rises, each bears a smaller fraction of the aggregate demand destruction from its own automation decision, so the wedge between the Nash equilibrium and the cooperative optimum widens. Competition, which economists normally treat as welfare-enhancing, amplifies the over-automation externality. The intuition is direct: each firm internalizes only its own fraction of the aggregate demand damage its automation imposes — in more competitive markets, that fraction is smaller, so the gap between what firms will do and what they should do widens. Second, as AI becomes more capable, the per-task cost saving from automation rises — which makes automating a more dominant strategy, not a less dominant one. Every advance in AI capability widens the wedge between what competing firms will rationally do and what they would collectively choose to do. The technology getting better makes the problem worse.
A standard objection at this point is worth taking seriously: automation historically creates new jobs. Mechanizing agriculture freed workers for manufacturing; automating manufacturing freed workers for services. The long-run record of technology is task substitution and task creation in roughly equal measure, and that record is real. But it has two limits in this context. AI automation is the first general-purpose technology to target cognitive and service tasks at scale — precisely the sectors where workers displaced by previous waves re-entered. And even if the long-run equilibrium is benign, the workers displaced in the interim are not abstractions: they are people in specific industries and communities, bearing concentrated costs while the aggregate rebalances over years or decades. The prisoner’s dilemma argument holds regardless of how the long-run resolves — over-automation relative to the social optimum occurs during the transition, and the demand destruction is real while it lasts.
This is the tragedy of the demand commons. It is not caused by corporate malice or regulatory failure. It is caused by competitive market structure operating exactly as designed.
What Sanders and Trump get right
Senator Sanders’s American AI Sovereign Wealth Fund Act would impose a mandatory 50% equity transfer on major AI companies, giving the federal government board seats and directing returns as direct payments to all Americans (Sanders, 2026). President Trump is in active talks with AI company executives about a voluntary version — companies donating stakes to a government fund, with dividends to households floated as one use of proceeds. That these conversations are happening simultaneously is itself telling: Sam Altman met separately with Sanders on Capitol Hill the same week Trump made his Air Force One statement — suggesting these are not opposing proposals but converging ones, forming a political center around public AI ownership rather than representing a partisan divide (Duncan & Beggin, Washington Post, 2026).
Both proposals rest on a correct diagnosis. The productivity gains from AI are accruing overwhelmingly to the owners of AI capital. The IMF’s analysis of 142 countries finds that in every plausible AI scenario — optimistic, pessimistic, and middle-ground — the gap between capital income and labor income widens (Cazzaniga et al., 2024). That is not a prediction about which scenario will obtain. It is a structural finding about all of them. Korinek and Stiglitz (2019) show formally that even in a model where AI raises GDP, labor’s income share can fall toward zero in the limit as automation expands — a result that undercuts the reassurance that aggregate growth makes distributional intervention unnecessary.
The instinct to capture a share of those gains for the public is correct or at least reasonable. The principle is sound. What neither proposal has stated clearly is which problem that share is meant to solve — and the answer determines whether the instrument fits the goal.
The gap neither proposal addresses
Here is what both proposals miss. A sovereign wealth fund — mandatory or voluntary, held or distributed immediately — addresses what happens to the proceeds of automation after the automation decision is made. It does not alter the per-task marginal calculus that makes automating a dominant strategy in the first place.
Firms will still over-automate relative to the social optimum regardless of who owns their equity. The prisoner’s dilemma operates at the margin where a firm decides whether to replace a specific task with a machine. A dividend paid to American households does not change that margin. Government board seats do not change that margin. Each of those instruments reaches the proceeds; none of them reaches the process.
The only instrument the academic literature identifies that operates at the right margin is a Pigouvian automation tax — a levy set equal to the uninternalized demand loss per automated task, applied at the point where the externality is created. Guerreiro, Rebelo and Teles (2022) reach a similar conclusion from a different angle: an optimal robot tax corrects the redistribution problem without requiring revenue replacement, because automation raises total output and tax revenue regardless. The US tax code currently moves in the opposite direction. Labor is taxed heavily through payroll and income taxes; capital investment in software and equipment is expensed immediately and subject to far lower effective rates — around 5 percent after the 2017 reforms, down from roughly 20 percent in 2000. That differential is an implicit automation subsidy: the tax code prices capital cheaper relative to labor, shifting the private automation calculus further from the social optimum (Acemoglu, Manera & Restrepo, 2020).
This is not an argument against public equity stakes or dividends. It is an argument that they solve a different problem — the income distribution problem — while leaving the automation incentive problem in place. Those are not the same problem. Treating them as interchangeable produces policies that look like answers but don’t address the underlying dynamic.
How large is the disruption — and can redistribution compensate for it?
The scale of displacement is uncertain. The IMF finds that roughly 40% of global employment contains tasks that AI could plausibly substitute or complement — rising to 60% in advanced economies — but whether that exposure produces displacement or augmentation depends on how AI is deployed in each occupation, and that is not yet settled (Cazzaniga et al., 2024). What is clearer is that AI’s aggregate productivity contribution is likely more modest than the most optimistic projections imply — Acemoglu (2024) estimates 0.53–0.66% TFP growth over ten years — which matters for how large a per-capita dividend funded from AI productivity gains would actually be.
Displacement, when it occurs, is concentrated. Acemoglu and Restrepo (2020) found that the robot-era displacement fell on specific industries, commuting zones, and demographic groups — the aggregate figure of roughly 400,000 jobs understates the intensity of the local shock. AI-era displacement is likely to follow a similar pattern, even if the aggregate trajectory is different.
This matters for evaluating whether a sovereign wealth fund dividend is sufficient to compensate those displaced. Nayebi (2025) models the AI productivity threshold required to fund an 11%-of-GDP UBI from AI rents alone, and finds it reachable but not guaranteed on any near-term timeline — somewhere between the early 2030s and the mid-2050s — a range driven almost entirely by assumptions about capability growth rather than the model’s structure. In the intervening period, a dividend funded from current AI valuations is likely to be modest. There is also a structural mismatch: displacement falls on specific workers in specific industries and communities, but a sovereign wealth fund dividend is distributed equally to every American regardless of whether they were displaced. A uniform small payment distributed across 330 million people provides very different support to a 55-year-old manufacturing worker in a single-industry town than the per-capita figure implies. The instrument that concentrates resources where the disruption is concentrated is a targeted support program — not a universal dividend.
That concentration in where the disruption falls is the design feature that determines which instrument we need. Before asking whether either proposal is adequate, there is a prior question the current debate has not clearly answered: what problem are we solving for?
A social dividend and a protection scheme are different instruments addressing different problems. A social dividend distributes a return on collectively-built productive capacity — the logic is ownership. AI was built on our data, our research, our accumulated knowledge, and everyone receives a share because everyone contributed. Universality is the point, not a design compromise. A protection scheme directs support to those whose livelihoods have actually been displaced — it should be concentrated where the damage is concentrated, calibrated to what people have lost, and substantial enough to replace meaningful income.
Senator Sanders and President Trump are both proposing something closer to the first. A mandatory or voluntary equity stake, distributed as a per-capita dividend, is a social dividend. The workers who most need a protection scheme — whose jobs have disappeared, whose industries have contracted, whose communities have been economically hollowed out — receive the same uniform payment as everyone else. That may be the right answer to the social dividend question. It is not an answer to the protection question. Neither proposal is clear about which one it is trying to be — and that confusion is not incidental. It is the reason the proposals can claim broad popular appeal while leaving the people most in need of targeted support with a per-capita share of an AI dividend that may, at current productivity thresholds, be modest at best.
The deeper problem remains structural. A sovereign wealth fund — however generously funded — does not address the demand commons problem: firms continue to over-automate at the margin where the externality operates.
What would actually address it
The defensible answer is an automation tax calibrated to the uninternalized demand loss — a Pigouvian correction at the margin where the externality is created. What that looks like in practice does not require observing individual automation decisions in real time. A workable first step is already visible in the literature: Acemoglu, Manera, and Restrepo (2020) calculate that restoring the pre-2000 differential between labor and capital taxation — eliminating the effective automation subsidy that accumulated through accelerated depreciation and the 2017 reforms — would reduce over-automation materially without requiring new legislative architecture. The harder design problem is calibrating the tax to the actual demand externality rather than a revenue target; the information regulators would need to set it correctly is not currently collected.
The practical obstacles are real and should not be understated. The political economy of taxing technology investment is hostile in both parties — the 2017 reform that deepened the effective automation subsidy passed with broad acquiescence. And any proposal that sounds like taxing innovation faces a labeling problem regardless of its technical design. None of that makes it the wrong answer. It makes it the hard answer.
The easy answers — equity stakes, dividends, sovereign wealth funds — are not wrong; they address a real problem. But they leave the primary distortion in place while distributing its proceeds. That is better than doing nothing. It is not the same as solving the problem.
The social disruption from AI-driven automation may be large (or not, we really do not know) and the period over which this disruption will occur is uncertain. The displacement will be uneven and concentrated in ways that aggregate statistics will obscure. What we can say is that we do not yet have a policy in serious political circulation that addresses the underlying prisoner’s dilemma rather than the distribution of its proceeds. The proposals on the table are better than the status quo. They are not a solution to the structural problem driving the disruption.
In the myth of Saturn, he devours his children not out of malice but out of fear — fear that what he created will eventually replace him. The prophecy comes true anyway, because the act of devouring is what sets it in motion. That is the prisoner’s dilemma exactly: every firm that automates to protect its competitive position accelerates the demand destruction that threatens the very market it is competing in. Saturn could not stop eating. Neither can any single firm. The only exit is structural — and the proposals currently in circulation do not reach it.