Large Language Models (LLMs) do not store entire images, such as complete frames from a Miyazaki film. Instead, they decompose these images into their component parts and are trained on those elements. This process creates a new product: a “style,” which is essentially a mathematical model of what could be termed “Ghibli-ness.” This model can then be applied to new imagery.
Currently, there is no legal precedent for protecting a “style”; copyright law only covers specific expressions of that style. In other words, you cannot copyright a music genre like “ska” or an art movement such as “impressionism.” While people may create and describe “styles” and “movements,” no one owns them—it is only specific artworks that can be owned.
This idea was reinforced by a recent conclusion from a U.S. District Judge in a case involving record companies. They claimed that when Anthropic’s Claude (another popular chatbot) remixes their song lyrics, it constitutes copyright infringement. However, as Enrique Dans explains, “if the system does not literally copy… there can be no infringement.”
