Do Large Language Models Complement Online Translation? Difference-in-Differences Evidence from ChatGPT Adoption in Korea

We examine how first use of a general purpose LLM, represented by ChatGPT, reallocates user time within the ecosystem of online translators. Using a matched difference-in-differences design with an event-time specification on a large Korean activity panel, we find that weekly translator use increases after adoption, peaks during the first six weeks, and then attenuates toward a near-zero steady state without turning negative on average. The average effect masks systematic heterogeneity: complementarity is larger for younger users, while a substitution pattern appears only among university and graduate students. Theoretically, we provide causal user-level evidence that a general-purpose LLM complements an online translation on average, link heterogeneity to capability, and offer a transparent template that combines event-time dynamics with a joint pre-tend test to separate initial exploration from steady state behavior. Practically, we translate the mechanisms into co-use design guidance, including one-click-pass through from LLM outputs to translator verification, and segment-specific onboarding that targets groups where complementarity is strongest.