This study quantitatively examines how the level of intellectual property protection over artificial intelligence (AI) training data affects AI capital accumulation. Korean copyright law provides no explicit exception for text and data mining; therefore, using protected works for AI training depends on ex-post interpretation of the general fair use provision. Drawing on a review of domestic and foreign regulations and disputes, the study identifies two channels through which protection operates: rising costs at the data procurement stage and binding constraints at the training stage. This distinction is consistent with 2025 U.S. court decisions that assess AI training separately from data acquisition. A dynamic general equilibrium model incorporating a Lucas span-of-control structure represents these channels as policy parameters and simulates how changes in protection affect aggregate output and knowledge capital, the model’s representation of AI capital. Under the baseline calibration, tightening protection by 20 percent reduces the knowledge capital stock by 17.1 percent and output by 3.4 percent, with the training-stage channel accounting for the larger share of the overall effect. This effect arises primarily from reduced use of knowledge capital rather than substitution toward labor.