Research

Algorithmic Priors: LLM-Mediated Belief Homogenization and Gender Segregation in Labor Markets

Statistical discrimination models have a built-in self-correction: employers with wrong gender-occupation beliefs incur costly mismatches, and competition erodes the error. This mechanism fails when employers share a large language model as their information source. An LLM trained on historically gendered text encodes a fixed representation of gender-job fit in embedding space — a structural artifact of gradient descent on biased corpora, not a correctable surface-level error. Because every user draws from the same parameters, cross-employer belief variance collapses as adoption spreads. Without that variance, no market signal punishes biased priors. Occupational segregation mirrors the training corpus rather than employer experience or taste, and persists even when every decision-maker is fully Bayesian.

We formalize this in a two-sided model of belief formation and search-match. Employers and workers update gender-occupation beliefs toward the LLM attractor via DeGroot dynamics; converged beliefs enter as endogenous bias parameters in a search-match framework. Cross-employer belief variance collapses at rate (1 - λ)^2t, where λ is the LLM adoption rate, destroying the heterogeneity that statistical self-correction requires. The equilibrium gender wage gap scales as λ²: a doubling of adoption quadruples the gap. When employers and workers also update beliefs against each other, they reach internal consensus before either converges to the true productivity allocation — a shared-misconception equilibrium that accelerates divergence further. In industries where technology intensity raises both LLM adoption and employer-worker interaction, the gap is monotone in technology intensity; where LLM use substitutes for direct contact, an interior maximum emerges at intermediate intensity — a counter-intuitive, testable prediction.

Two policy instruments follow. Restricting adoption speed slows convergence but leaves the biased attractor unchanged; retraining the model shifts the fixed point itself. The welfare cost of misallocation scales as λ²ε², making training-level governance strictly more effective in the long run.

Old-Age Pensions and Three-Generation Household Reorganization: Evidence from Rural China

China's 2016 pension reform raised the monthly benefit floor for rural residents by 27 percent. Using the reform's age-60 eligibility cutoff in a difference-in-differences framework, we show how this pension windfall reshapes decisions across three generations of the same rural household. Pension eligibility raises grandchild care provision by 8.2 percentage points while precautionary medical spending falls. When a grandparent becomes eligible, grandchildren spend 1.4 fewer months per year with the father and 1.6 fewer months with the mother. Complete paternal absence falls while moderate absence rises. The pension reorganizes parental work toward shorter-distance, more returnable trips rather than pushing parents further from home. Greater provincial pension generosity also reduces migrants' self-reported childcare burden and raises the likelihood of migrating with children in tow. The transfer channel confirms the mechanism: remittances from pension recipients fall by roughly 40 percent among those with adult children but are unchanged among those without. The pension substitutes for, rather than supplements, the informal family insurance that had tethered adult children to the village. The adjustment runs entirely through grandchild care, not elder care. The pension does not make grandparents more dependent. It frees their children to leave.

Global Value Chains, Specialized Inputs, and Sino-U.S. Trade Friction: Evidence from China's Customs Data

Standard trade-in-value-added (TiVA) methods assume each firm allocates imported intermediates in proportion to its overall input mix, regardless of destination or product. This paper relaxes that assumption by combining WIOD input-output tables with China's HS8-level customs records, which allow destination- and product-specific intermediates to be tracked directly. The resulting foreign value-added measures are substantially larger than conventional TiVA estimates for U.S. content embodied in Chinese exports, particularly in electronics and other high-technology sectors. The finding suggests that the trade balance attributable to U.S. inputs is routinely understated in standard accounting, with implications for how trade friction episodes are measured and interpreted.