Abstract: Generative AI lowers the cost of producing text, but much of that text is also an input another agent must process and act on. I study a cost shock that hits both sides of one production chain: civil litigation, where lawyers write petitions and judges process them to reach decisions. Using the full text and metadata for the universe of airline small-claims cases in São Paulo (2022–2025) and a validated AI-text classifier (near-zero false positives in a pre-GPT placebo and about 17% of lawyers and 8% of judges flagged by late 2024), I use early-versus-late adopter difference-in-differences to estimate the direct effects of adoption and its downstream spillover onto the judges who process adopters’ filings. Lawyer adoption has no clear net effect on whether a case is won or settled, a reduced form that nets offsetting channels: petitions are cleaner but contain more procedural mistakes, and adopters file more, lower-merit cases. Cases filed by adopters raise judges’ processing burden, with longer response times and more routine rulings per case. AI can make a lawyer more productive while imposing a downstream externality on the judges who process the result.

For Information contact angela.baldassarre@unibocconi.it or giulia.zenoni@unibocconi.it