When Assistance Becomes Infrastructure

Debates about the future of artificial intelligence often focus on a single decisive moment, when machines surpass humans. This essay examines a less visible path: AI may become infrastructure through a series of local, reasonable decisions that no one makes as a whole. We follow one question across three stages: can we still go back? First, services are redesigned around AI and non-AI channels close, a form of coercion without a coercer. Second, the human role shifts from producing to ratifying, and the skills needed to check AI output erode, as early clinical and labor-market evidence suggests. Third, the ability to compare work with and without AI disappears, so that a society may lose not only a capability but also the knowledge that it existed. A simple model shows how two reinforcing loops create hysteresis, in which the way back costs more than the way in. We then discuss brakes (laws, fallback options, reserves of skill), structural forcings (public debt, the decline of public research, geopolitical rivalry), and four possible end states: integration, bypass, managed coexistence, and partial return. We argue that the key question is whether human alternatives remain reproducible and measurable.