A theoretical model finds that generative-AI automation can reinforce itself through local demand, creating a choice between two self-sustaining outcomes. In one, firms use usage-based automation and local demand is low. In the other, firms use human augmentation while carrying a team wage bill and demand is high. The finding is conditional on assumptions about wages, demand and how the two production modes complement each other.
The study describes a local service economy with a normalized set of differentiated sectors and one incumbent in each. Firms are treated as atomistic, meaning each takes total market size and the wage schedule as given when choosing a production mode. It is a model rather than an empirical sample, and its numerical exercise is illustrative rather than a structural estimate.
The tipping point is the story
The production choices are deliberately asymmetric. Automation is modeled as usage-based, with no material downstream capacity fixed cost. Human augmentation requires the firm to carry a team wage bill. That setup makes payroll more than a cost: it also supports household spending on local services, so a broad switch toward automation can weaken local demand.
Within what the authors call the coordination region, both endpoint modes are equilibria. A unique middle threshold separates them: under myopic adjustment, the endpoints are locally stable while the threshold is unstable. In plain terms, a starting share on one side tends to move toward the automated outcome, and a starting share on the other side toward augmentation.
The model also gives a conditional profit result. Even with flexible wages, it does not guarantee that augmentation pays more. But under an additional payoff condition, every downstream firm earns more in the human-augmented equilibrium than in the automated one.
Expectations can lock in a path
Static equilibria do not determine how the economy gets there. When firms revise at different times and look ahead, the same inherited employment share can support an all-automation path or a recovery led by human augmentation, depending on what firms expect later revisers to do. The model therefore treats expectations as part of the transition, not just a reaction to the final outcome.
A stochastic version adds public aggregate shocks and limited revision opportunities. Under the paper's stated regularity and boundary conditions, it has a unique state-contingent path. Yet finite frictions preserve history dependence, while the vanishing-friction limit selects the model's risk-dominant mode. The result is conditional: it does not say that arbitrary noise will choose the outcome.
Wages narrow, but do not erase, the choice
Allowing wages to adjust changes the map. Within the maintained strategic-complementarity domain, wage adjustment raises the amount of displacement needed to sustain the automated endpoint and narrows the interval in which both outcomes can coexist. The paper warns that its global classification may not apply outside that domain.
The welfare calculation points in one direction inside the same region. Human augmentation is the unique local first best, meaning it is the only allocation that maximizes the model's local welfare under its stated surplus, convexity and complementarity assumptions. That is a local result, not a national or global verdict.
The policy response follows that logic. The paper defines a state-contingent Pigouvian wedge, a policy adjustment matched to the spillover from adoption, to align private incentives with the social decision-maker's marginal incentive. If the low state is locally optimal, the model calls for a temporary coordination bridge. For forward-looking firms, support must last beyond the myopic threshold until pessimistic expectations no longer sustain an automation cascade.
The numbers are illustrative
The paper's numerical examples are meant to show how thresholds move, not to estimate what happens in actual economies. In a central high-autonomy scenario, the illustrative displacement share is d = 0.375 and the interior threshold is x* = 0.519. That sits just above the risk-dominance cutoff dRD = 0.374, so the case is reported as narrowly automation-risk-dominant. At d = 0.42, automation is dominant even at the high-demand endpoint.
A framework, not a forecast
Those figures should not be read as forecasts. The model does not observe firms or workers, estimate causal effects of generative AI on jobs, wages or demand, or attach uncertainty intervals to its illustrative inputs. Its conclusions depend on stylized assumptions about production, labor supply, household spending and local markets.
The dynamic results cover two monotone polar paths, not every possible nonmonotone route. Owner spending and displaced-worker re-employment are treated as extensions rather than part of the baseline. Another extension allows offensive automation to reduce augmentation profitability in defensive sectors through the shared labor-income market. Whether these mechanisms hold outside the modeled local economy remains an open empirical question.
Paper data and sources
Original title: The Reverse Big Push: Generative AI and Self-Fulfilling Automation
Authors: Soumen Banerjee, Jianguo Wang
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text