Generative Models Erode Human Temporal Learning Through Market Selection

Abstract: We argue that modern generative models create structural risks for knowledge and cultural production at current, sub-AGI capability levels. We define Human Temporal Learning (HTL) as path-dependent knowledge accumulation through sustained engagement with problems over time. Generative outputs increasingly resemble HTL-intensive work in surface features, so verifying whether a given output reflects genuine human learning grows costly relative to its expected benefit. Once verification loses economic justification, evaluators reward outputs regardless of production mode, and producers who invested years of learning compete on price against outputs that cost almost nothing to generate. We call this pathway value collapse and formalize it through a costly-inspection framework. Cross-domain evidence from academic publishing, legal practice, content platforms, and software security maps onto four stages of verification erosion. Alignment success is orthogonal. Better-aligned models narrow observable gaps between human and AI outputs, making source verification harder and intensifying competitive pressure against HTL-intensive work even when individual AI outputs improve.

• The paper demonstrates that generative models erode human temporal learning by triggering a value collapse via adverse market selection.
• It employs a reduced-form costly-inspection framework to show how high-HTL producers are outcompeted by low-cost AI outputs.
• The findings advocate for provenance-enhancing interventions to protect cumulative human expertise and sustain genuine skill acquisition.

Generative Models Erode Human Temporal Learning Through Market Selection

Introduction

The paper "Generative Models Erode Human Temporal Learning Through Market Selection" (2606.06572) develops an economic and sociotechnical framework for understanding how generative models, even at sub-AGI capability levels, erode the economic viability of Human Temporal Learning (HTL)—defined as path-dependent, cumulative expertise embedded in humans through sustained engagement with nontrivial problems. The central thesis is that improving output indistinguishability and declining verification ability, rather than overt intent alignment failures or AGI, already drive systemic displacement of HTL via adverse selection in market and institutional contexts.

Human Temporal Learning and Mechanisms of Erosion

HTL is described as a unique process in which continuous interaction with complex problems accumulates unobservable, implicitly encoded skills and judgment. Historically, institutions rewarded outputs that acted as reliable signals of temporal investment and expertise. Generative models, by optimizing for the surface features of HTL-intensive outputs—rhetorical structure, citation habits, stylistic coherence—fundamentally disrupt this equilibrium. Verification of genuine HTL entails costly deep inspection, while low marginal cost AI outputs continuously approach the superficial quality of genuine human work.

The paper formalizes this as a "value collapse" feedback loop: verification ability $g$ degrades as distinguishing features narrow, the quality gap $\Delta q$ becomes less economically actionable, and per-item verification cost $c_v$—often measured in reviewer or expert time—remains high or increases under volume pressure. Once $g \cdot \Delta q < c_v$, verification loses economic rationality and evaluators reward outputs blindly, independent of provenance or genuine HTL content, shifting competitive dynamics against sustained human learning.

Market Selection and the Costly-Inspection Model

A reduced-form costly-inspection framework, grounded in information economics and adverse selection theory [akerlof1970lemons], underpins the core mechanism. The model distinguishes between high-cost, high-quality HTL-intensive producers ($H$) and low-cost (AI-primary or low-HTL) producers ($L$). When verification is infeasible or economically unjustifiable, rewards pool across producers, rapidly outcompeting high-cost $H$ types and leading to their market exit. This dynamic is self-reinforcing: as $\lambda$ (the HTL share) falls, the average (pooled) reward $\bar{p}$ drops, the private value of verification recedes, and subsequent rounds select increasingly for low-HTL modes.

This formalism enables precise reasoning about channel-specific interventions: improving $g$ via watermarking or cryptographic provenance; reducing $\Delta q$0 via reviewer tooling and automation; and increasing $\Delta q$1 via institutional redesign or liability assignment.

Empirical Evidence and Domain Stratification

By analyzing cross-domain evidence, the paper presents a four-stage model of verification erosion:

• Stage 1 (Intact Verification): Clinical medicine, where high-stakes consequences and institutionalized review keep HTL-differentiating oversight viable.
• Stage 2 (Sanctions Maintain Verification): Legal practice, where formal penalties for AI-induced error make costly verification rational even as its difficulty grows.
• Stage 3 (Volume Overwhelms Verification): Academic publishing and open-source security, where the influx of formulaic or redundant generative outputs makes individual verification unscalable, triggering a 47-fold surge in automated papers and a steep rise in undetected fabrications.
• Stage 4 (Source-Blind Rewards): Digital content platforms and the creator economy, where engagement metrics are indifferent to content provenance and AI-generated works often secure rewards equal to HTL-intensive content. Current market structure, with platform power and scale economies, accelerates this convergence.

These stages are substantiated by extensive citation of recent measurement studies and field reports showing near parity in surface quality and increasing volume-driven review shortfalls [liang-etal-2025-quantifying, suchak-etal-2025-explosion, esau-etal-2025-gptzero, maupin-etal-2025-dramatic].

Theoretical and Practical Implications

A crucial conceptual contribution is the orthogonality of alignment to verification erosion. Stronger alignment—models that are more honest, helpful, and harmless—further narrow the visible gap between human and AI outputs. The paper thus claims that successful alignment intensifies competitive pressure against HTL, not the reverse, as higher local model alignment or output quality makes provenance-based value signals even less observable within surface-level evaluation regimes.

Long-term risks include "pipeline compression," where automation of entry-level tasks blocks skill acquisition pathways, gradually reducing the evaluator pool and making deep verification ever more expensive. This is empirically supported by labor market data indicating a disproportionate employment decline for early-career workers in AI-exposed fields [brynjolfsson2025canaries].

The paper connects value collapse to model collapse [shumailov2024curserecursiontraininggenerated, alemohammad2024selfconsuming]. As the active pool of human-generated (HTL) content shrinks, next-generation models trained on increasingly model-generated distributions lose diversity and utility, precipitating systemic degradation.

Governance, Provenance, and Future Directions

Governance recommendations focus on making provenance observable and auditable, reducing verification costs, rewarding HTL explicitly, and protecting the HTL pipeline. The paper addresses limitations and alternative views, noting the inadequacy of voluntary disclosure, difficulties in large-scale provenance enforcement, and the non-equivalence of terminal consumption efficiency with knowledge production ecosystems where recursive feedback cycles are present.

Future research and institutional responses must center on provenance-preserving mechanisms, auditing infrastructure, economic incentives that differentially reward temporally embedded skill, and regulatory structures to avert unchecked adverse selection. Advanced provenance solutions must be robust to adversarial tactics [sadasivan2025aigeneratedtextreliablydetected] and platform incentives.

Conclusion

This paper establishes that the displacement of HTL by generative models is an emergent property of ordinary market and institutional selection, not a distant AGI risk. Erosion of HTL and the supporting evaluator base precedes and exacerbates more advanced forms of systemic risk. The analysis demonstrates the necessity of coupling technical alignment work with economic and governance interventions aimed at sustaining temporally embedded human expertise for the integrity of knowledge-based institutions. The survival of effective collective judgment in AI-mediated environments depends on preserving incentives for temporal learning and reconciling surface-based evaluation with provenance-sensitive reward allocation.