The Democratic playbook for an AI economy needs three pillars

The Democratic Party is being asked to write economic policy against a forecasting void. AI capability is advancing on a curve that Sam Altman calls unprecedented and Zvi Mowshowitz believes is approaching recursive self-improvement, yet the Forecasting Research Institute finds that only AI experts expect a growth takeoff — economists, superforecasters and the public do not. A serious left-of-centre programme cannot wait for that disagreement to resolve. It has to be technology-agnostic: robust if AI delivers a post-scarcity economy, robust if it delivers stagnation with displacement, robust if it delivers something stranger.

Start with what is knowable. Robert Gordon's figure — a sevenfold rise in U.S. real output per capita between 1870 and 1970 — is the benchmark against which any AI revolution will be judged, and it took a century of electrification, internal combustion and routinised corporate R&D to deliver. Paul David's work on electricity, where flexible factory layouts produced 30 to 50 percent productivity gains only after decades of reorganisation, is the cautionary footnote. The honest position is that AI's macro impact is unknowable on a five-year horizon, which means policy should be built around distributional insurance rather than growth forecasts.

Abundance as the first line

The first pillar is making the basics cheap: housing, energy, healthcare, education, food. This is the abundance agenda stripped of its libertarian framing. If AI compresses, as Noah Smith puts it, a century's worth of change into two decades, the political economy cannot absorb that shock while rents and insurance premiums continue their current trajectory. Cheap basics also hedge the pessimistic scenario the ServiceNow chief executive sketched — chronically unemployed children — because the floor under a displaced worker is set by the cost of staying alive, not by the wage on offer.

The second pillar is a public equity claim on the capital that captures AI rents. Call it a sovereign wealth fund, call it universal basic compute as Altman does, the mechanic is the same: if the marginal productive unit is a data centre rather than a worker, the state needs a direct ownership stake in the returns, not merely a tax claim on them. This is where Democrats have been weakest. A corporate income tax is a poor instrument for taxing a capital stock whose accounting profits can be routed through Ireland; a preferred-equity stake taken in exchange for power grid access, land, or compute subsidies is not.

The third pillar is the most counterintuitive: subsidise human work, even where AI is cheaper. Acemoglu, Kong and Ozdaglar's recent paper on agentic AI gives the theoretical foundation — when AI substitutes for human effort, it erodes the learning externality that sustains the community's stock of general knowledge, and the economy can tip into what they call a knowledge-collapse steady state. Teaching, caregiving, scientific apprenticeship, judicial reasoning: these are domains where the cheap automated answer destroys the human capability that produced the training data in the first place. A wage subsidy for human-delivered services is not nostalgia. It is insurance against the model running out of things to learn from.

None of this requires Democrats to take a view on whether Mowshowitz's escape velocity arrives in 2027 or never. Cheap basics are good policy in stagnation. Public equity stakes are good policy in a boom. Human-work subsidies are good policy in a knowledge-collapse scenario the authors of the Acemoglu paper take seriously enough to model. The party that wins the next decade will be the one that stops trying to forecast the technology and starts writing policy that survives every plausible path it could take.