The trade effects of AI provisions in PTAs: Does non-binding matter?

This paper examines whether non-binding ("soft") AI-related provisions in preferential trade agreements (PTAs) move trade and how their effects compare with binding ("hard") provisions. The motivation is straightforward: digital rules are spreading quickly, but legal enforceability is uneven. We ask whether non-binding instruments (standards work, data-innovation clauses, interoperability programmes) can deliver trade gains in AI-intensive sectors even without hard obligations. Drawing from TAPED AI-related provisions in PTAs, we construct a measure of AI-related depth, distinguishing binding from non-binding provisions. To identify where these rules should matter most, we interact agreement depth with a sectoral proxy for AI intensity based on firm adoption. We estimate a structural gravity model using PPML with exporter-year, importer-year, pair, and pair-trend fixed effects, and run extensive robustness checks including pooled sector interactions and leads/lags.