Generative AI lets a child obtain the essay, the answers, the analysis — without performing the cognitive work the assignment was built to elicit. A dynamic theory of skill formation asks when that cognitive delegation is harmless, when it isn't, and what governs the difference.
When ChatGPT arrived, The Atlantic announced that “the college essay is dead.” New York City’s school system — the nation’s largest — banned the tool in January 2023, then reversed course four months later, conceding the ban had “overlooked the potential.” Adoption, meanwhile, raced ahead: the share of U.S. teens using ChatGPT for schoolwork doubled in a single year — from 13% in 2023 to 26% in 2024 — and by 2025 more than half of teens reported turning to chatbots for schoolwork.
The argument has fixated on cheating and access: who uses it, who can detect it, who should ban it. This paper asks a different question — not whether children use AI, but what its use does to the formation of the mind.
For two centuries, schooling worked because the technology of producing output enforced the underlying process: to finish the problem set, a child had to think. Generative AI breaks that link — the same output can now arise from active engagement or from passive reception.
The model makes that distinction a primitive. Each period a child chooses effort e, delegation a, and verification v. A governance function φ(e ; g) — weakly increasing in both arguments, with φ(0 ; g)=0 — sets the process-preservation threshold. A canonical “first-attempt rule” lets AI scale with the child’s own prior effort.
Raising governance g lifts the boundary, so more of a child’s AI use stays on the process-preserving side.
The child accumulates three capital stocks — knowledge h, discipline D, and judgment J — that enter output through a CES technology and are dynamically complementary. When AI replaces the process rather than supporting it, a wedge opens between the active-engagement counterfactual Wactive and realized welfare WR.
The Capacity-Wedge Decomposition (Theorem 1) shows the wedge is exactly additive. The first term is the discipline a child fails to build by skipping the struggle. The second is the cascade: foundational knowledge missed today is an input into all later learning, so the deficit propagates down the curricular sequence — and in typical parameter regions it is the larger term.
Begin two identical children at the same capacity. One uses AI to support her own engagement; the other delegates the cognitive work. The contemporaneous gap looks small — but it does not stay small.
This is self-productivity and dynamic complementarity (Cunha–Heckman) running in reverse: each period’s missed knowledge lowers the productivity of the next period’s investment. A present-biased, sophisticated child (quasi-hyperbolic β, δ) already under-invests; process-replacing use compounds that gap rather than closing it.
The marginal welfare cost of process-replacing use, M(t) = δt [ πt γa ΨD + λh Ψh(t) ], is strictly decreasing in age. Three forces drive it down: discounting δt, the exponential decline of plasticity πt as habits set, and the cascade-distance term Ψh(t), which shrinks as the remaining curricular runway (T−t) shortens.
The implication is sharp: policy attention is most consequential early. A unit of substitution at age six carries a welfare cost orders of magnitude larger than the same unit at sixteen.
Whether AI access raises or lowers welfare — the sign of ∂V0/∂θ — flips at a sharp threshold g*(θ). The cross-partial ∂2V0/∂θ∂g > 0: more governance makes a given capability more beneficial. And the threshold itself rises with AI capability (dg*/dθ > 0) — more powerful models demand more governance to stay net-positive.
Worse for equity, the cross-sectional variance of welfare equals 2 · Cov(V0, ∂V0/∂θ) > 0 and grows in θ. The equity question is therefore not who can afford the technology, but who can govern its use.
The public debate asks whether children cheat and whether schools can detect it. This framework reframes the question: generative AI does not merely change the price or availability of an educational input — it changes the production function for human capital itself. The same technology preserves the cognitive process for one child and replaces it for another, and which mode obtains is set by governance, not by who owns a subscription.
That matters because the capacities at stake — discipline, judgment, the habit of independent reasoning — are durable life-course inputs, the very non-cognitive skills the Heckman tradition links to earnings, health, and civic participation. A wedge opened in childhood is a wedge in life trajectories, not in transcripts. And because the cost of substitution is largest early and the welfare effect of access flips sign at a governance threshold that climbs with capability, universal access deployed on its own can convert unequal governance into unequal lives — widening the very gaps it is often meant to close.
The levers the model identifies are concrete and decidedly not access-based: effort-contingent designs such as the first-attempt rule; verification requirements that keep the child’s judgment in the loop; and governance support targeted where the marginal cost is highest — the early years. The cohort that met large language models in kindergarten reaches the labor market around 2040; the institutions that decide between the two modes are being set now.
The question is not whether to give children the machine — it is whether we build the scaffolding that decides what the machine does to them.