Training AI models on protected works without permission can violate copyright law when they reproduce “memorized” data, the US Copyright Office said in a pre-publication version of its latest report on artificial intelligence.
Both collecting data and using it in the model-training process can implicate copyright holders’ reproduction right, the office wrote, and training can also touch on the right to control derivative works. But the question “turns on whether the model has retained or memorized substantial protectable expression from the work(s) at issue.”
The office rejected the argument that training AI models on copyrighted works is inherently “fair ...