Large Language Model Adoption and Firm Innovation: More Patents within Existing Knowledge Boundaries
Li Lai, Yikun Wang, Sipeng Zeng
Abstract
Large language models (LLMs) let employees search, explain, and recombine technical knowledge at low cost. We ask whether firms that adopt them innovate more and whether they innovate beyond the technologies they already know. Using a classifier validated against human coders, we date when Chinese listed firms first confirm, in replies to investor questions, that they have put LLMs or other AI applications into use, and we link these events to the patent applications of 1,836 firms over 2019--2024. Patent applications rise by 12.5% after adoption. In a second design based on the arrival of LLMs after ChatGPT's release, firms that had hired more heavily into AI-exposed occupations increased their patenting relative to other firms in 2024. The additional patents stay within firms' existing knowledge boundaries: utility-model and exploitative applications rise, new patents move closer to the firm's prior technology portfolio, and neither invention nor exploratory applications increase significantly. Adopters also raise R&D intensity, software-and-data investment, and the share of graduate employees. The patent response is concentrated in labor-intensive industries and in highly AI-exposed firms, it tilts further toward exploitation under myopic managers, and among business functions only market-facing uses are followed by significantly more patents. In the adoption year and the year after, LLM adoption intensifies local search: firms patent more in technologies they already hold and no more in new ones.