AI models threaten human expertise through 'value collapse', warns ICML 2026 paper

Wenjun Cao's ICML 2026 paper introduces a formal framework for understanding how current generative models undermine human expertise. The central concept is Human Temporal Learning (HTL)—knowledge and skills accumulated through path-dependent, sustained engagement with complex problems over time. As generative outputs mimic the surface features of HTL-intensive work, verifying whether a result stems from genuine human learning becomes increasingly costly relative to its benefit. Cao models this as a costly-inspection problem: once the cost of verification exceeds the expected reward, evaluators stop checking and reward outputs purely by their apparent quality. Producers who invested years in learning then compete on price with near-zero-cost AI outputs, leading to 'value collapse'—the economic devaluation of deep expertise.

The paper maps this dynamic across four domains: academic publishing, legal practice, content platforms, and software security, identifying four stages of verification erosion. Crucially, Cao demonstrates that better AI alignment (making outputs more truthful and helpful) actually intensifies the problem by narrowing observable gaps between human and AI outputs. Ironically, more aligned models make source verification harder, increasing competitive pressure on HTL-intensive work even as individual outputs improve. The paper was accepted at the ICML 2026 Position Paper Track and argues this is a structural risk that persists below AGI capabilities.