What We Know — and Don’t — About AI, Automation, and Basic Income

Dystopian or utopian — which AI future is coming our way? As an avid consumer of science fiction in my youth, and, truth be told, still an avid reader, the genre is rife with both sides: the Skynet apocalypse or the more benign futures of Isaac Asimov’s robot series.

Yet today we are deluged with functionally similar visions of an AI dystopia in which the precious few live a life of absolute luxury while the masses are consigned to destitution (think Elysium), or a life in which tedious tasks are outsourced to our “agents” while we live a life of comfort and meaning (The Jetsons). Even the great corporate titans of artificial intelligence seem caught on Occam’s razor between their technophilia and their techno-dystopian visions of the future – their own personal gain is, perhaps, the deciding factor. These futures are representative of our fears of devolving into highly unequal societies while yearning for greater equality. Given these futures, we must ask ourselves: what, if anything, do we owe each other in the face of great change?

The AI and jobs debate has become something of a proxy war between two camps who rarely engage each other’s evidence. Basic income has been pulled into this debate as an answer to AI disruption — sometimes by people who support it on entirely different grounds. Before asking whether basic income is the right policy response, it is worth asking what response is actually warranted. The answer requires separating several questions that tend to get run together.

What the robots actually showed us

The best causal evidence we have — not about AI, but about the most recent large wave of automation — is sobering. Acemoglu and Restrepo (2020) studied the introduction of industrial robots into US commuting zones between 1990 and 2007, using European robot adoption rates as an instrument to identify causal effects. One additional robot per thousand workers reduced the local employment-to-population ratio by 0.39 percentage points and wages by 0.77 percent. There was no offsetting increase in employment elsewhere. Roughly 400,000 jobs were lost in the aggregate over that period. The displacement effect — machines replacing workers in specific tasks — was larger than the productivity effect that traditionally offsets it.

This does not mean automation always destroys jobs in the aggregate. Autor (2015) synthesizes a longer historical record and reaches a more measured conclusion: automation substitutes for routine, codifiable tasks but consistently creates demand for new kinds of work. The reason employment did not collapse after the mechanization of agriculture or after the personal computer is that productivity gains raise real incomes and demand for goods and services, and human comparative advantage shifts to tasks machines do badly. Automation’s problem is not scarcity but distribution — who captures the gains and who bears the adjustment costs.

Both of these readings are supported by credible evidence. They are not obviously in conflict. What they leave open is the more urgent question: is AI different?

Three futures — and what we actually know

There are broadly three ways the AI and jobs story might unfold. I will call them dystopian, K-shaped, and utopian, though those labels carry more rhetorical baggage than the underlying analysis deserves.

In the dystopian scenario, AI displaces workers faster than new tasks emerge, labor’s share of income falls structurally, and the result is mass technological unemployment. The academic evidence does not rule it out. Korinek and Stiglitz (2019) construct a formal model in which, as automation expands without bound, labor’s income share asymptotically approaches zero even as GDP grows. That is a theoretical limiting case, not a near-term forecast — it has not materialized on any visible scale — but it illustrates why the most extreme versions of the dystopian scenario are difficult to dismiss on first principles alone.

In the utopian scenario, AI generates large productivity gains broadly shared across workers, reduces drudgery, and ushers in something closer to the leisure economy Keynes predicted a century ago. Goldman Sachs estimated in 2023 that AI could raise global GDP by 7 percent and lift US labor productivity by 1.5 percentage points a year over a 10-year adoption period. These are large numbers — and contested ones. The gap between the Goldman Sachs estimate and the best academic work comes down almost entirely to one assumption: what fraction of tasks can AI feasibly and profitably automate within a realistic time window.

Daron Acemoglu (2024) applied a framework that takes that question seriously, and arrived at a substantially smaller number: a TFP gain of at most 0.53–0.66 percent over 10 years, translating to GDP growth of roughly 0.93–1.56 percent. He is explicit that even this is an upper bound — slow corporate AI adoption, organizational adjustment costs, and the prevalence of tasks that are genuinely hard for AI to learn would all push the estimate lower still.

The K-shaped scenario — the middle ground — holds that AI raises aggregate output modestly, but the gains and losses are unevenly distributed. The IMF’s January 2024 analysis covers 142 countries and finds that about 40 percent of global employment is exposed to AI, rising to 60 percent in advanced economies (Cazzaniga et al., 2024). Crucially, it distinguishes between occupations where AI is likely to complement workers (raising their productivity and wages) and occupations where it is more likely to substitute for them. In all three of the IMF’s model scenarios, one result holds consistently: the gap between capital income and labor income widens. Even the most optimistic scenario is not good news for labor’s share of national income.

This is probably the most important empirical finding for the policy debate, and it is underappreciated in public discussion. The question is not only whether AI creates unemployment. It is also whether AI concentrates the gains with capital owners in ways that require a distributional policy response, regardless of what happens to the employment rate.

What the evidence cannot yet tell us

Whether the historical pattern of technology-driven job creation will repeat is genuinely unresolved. Prior waves of automation displaced workers from specific tasks but created entirely new categories of work — the administrative economy, the information economy, the personal services economy. Whether AI will generate comparable new demand for human labor is uncertain, and Acemoglu (2024) makes a pointed observation: current tech industry priorities — automation and data monetization — are not oriented toward creating new, human-complementary tasks. That could change, but it would require a deliberate reorientation of how AI is developed and deployed.

Whether an AI dividend is fiscally viable is equally open. Nayebi (2025) shows theoretically that AI would need to reach roughly 5–7 times its pre-AI automation productivity to fund a UBI equivalent to 11 percent of GDP from rents alone. His estimated timeline for crossing that threshold runs from the early 2030s to the mid-2050s — a range that spans the difference between a near-term policy challenge and a theoretical long-run possibility. And the current US tax code actively subsidizes automation rather than taxing it — effective capital taxes on software and equipment fell from around 20 percent in 2000 to roughly 5 percent after the 2017 tax reform (Acemoglu, Manera & Restrepo, 2020). The institutional machinery to capture AI rents does not yet exist.

The distributional case — and the problem redistribution can’t solve

Strip away the most speculative claims, and what remains is a limited but defensible case for intervention. The robot-era evidence is unambiguous that large-scale automation has, in specific places and industries, cost workers jobs and wages with no compensating local adjustment. The AI-era evidence is thinner but consistent with the finding that capital captures a disproportionate share of productivity gains. The IMF’s finding — that capital-labor inequality widens in all plausible AI scenarios — suggests the distributional challenge is real regardless of whether the employment apocalypse materializes.

There is also a structural problem that income redistribution, by itself, cannot solve. Think of it this way: when a firm automates a worker’s job, it captures the full labor-cost saving. But the revenue loss from that worker no longer spending their wages falls across all competing firms equally — because displaced workers are also consumers. Each firm imposes a cost on the entire market that it cannot individually internalize. The result is a prisoner’s dilemma: every firm would be better off in the cooperative outcome with less automation, but automating is a dominant strategy for each one acting alone. More competitive markets, paradoxically, generate worse outcomes. And more capable AI widens the gap rather than closing it. Hemenway Falk and Tsoukalas (2026) formalize this mechanism and show the outcome is Pareto-dominated — both workers and firm owners are strictly worse off than they would be under cooperation.

The only instrument that operates on the right margin is a tax on automation itself. Guerreiro, Rebelo and Teles (2022) reach a similar conclusion from a different angle: an optimal robot tax starting at around 5 percent and declining over time delivers meaningful welfare gains precisely because it corrects the redistribution problem without requiring revenue replacement.

Senator Bernie Sanders recently proposed one of the most ambitious policy responses in this debate — the American AI Sovereign Wealth Fund Act. The proposal would transfer 50 percent of the equity in major AI companies to a federal fund, directing returns as direct payments to all Americans. Sanders grounds it in a clear argument: AI was built on publicly generated knowledge — our books, data, journalism, and scientific research — and the wealth it produces should therefore benefit the public. That diagnosis is largely correct, and it connects to the same “collective resource” logic underlying data dividend proposals and Alaska’s Permanent Fund.

But the proposal conflates two different income-support logics that are worth separating. A basic income targeted at those in need — displaced workers, low-income households, people who cannot find employment in an automated economy — is designed around poverty reduction and income security. The payment level, the eligibility rules, and the fiscal structure all follow from that purpose: you need certainty of payment, you need it concentrated where hardship is greatest, and the universality question is really about how much targeting efficiency you trade away for simplicity and take-up. A social dividend — the model Sanders is actually proposing — is not targeted at need at all. It is a return on a collectively owned asset, paid to everyone because everyone contributed to building it. Alaska’s Permanent Fund does not ask whether you need the dividend; it pays it because the oil belongs to all Alaskans. These are different instruments addressing different problems — and the Sanders proposal is coherent as the second, but not as the first.

And regardless of which framing you adopt, neither addresses the prisoner’s dilemma. Firms will still automate beyond the socially optimal rate regardless of who owns the stock, because the externality operates at the per-task margin that income transfers — targeted or universal — leave untouched.

This is the prior question the AI debate keeps skipping: what is the instrument actually for? Getting that question right matters more than any of the specific proposals currently on the table. The design will embed the answer — whether we’ve thought it through or not.