Scientists have reported that current generative AI systems exhibit significant limitations in their creative capabilities. Publishing their article in the journal Patterns , the researchers assert that these systems produce generic, repetitive, and "populist" images when operating without human supervision, falling far short of what humans consider artistic creation.
Arend Hintz , a professor at Sweden's Dalarna University, said : "Our experiments have shown that the creative capabilities of generative AI systems are very limited. The images they produce are generic, repetitive, and 'populist,' and AI will not be able to become a Picasso to create a painting like 'Guernica'."
The researchers reached this conclusion through an experiment in which they created a creative workshop using two generative artificial intelligence systems: one being the common Stable Diffusion model for generating images based on user queries, and the other being the Large Language Model (LLaVA) capable of creating textual descriptions of images.
The researchers prepared 100 descriptions for very different images , collected online using another algorithm, and then fed them through an image generator. Next, they used the LLaVA model to generate textual descriptions for the resulting images and fed them back into the Stable Diffusion algorithm to generate new images, thus starting a continuous cycle of image generation.
Professor Hintz pointed to what he called the "cultural homogenization of artificial intelligence ," saying: "We expected the images to stabilize after several iterations and align with the original queries, but in reality, the images began to shift from one subject to another periodically over thousands of iterations, for reasons that are not clear."
