Copyrights and AI

A second requirement concerns the copyrighted work: it must contain a certain degree of originality and expression. In the early days of photography it was debated  whether or not to freely print the work of photographers, who after all only depicted reality as it was, and who could claim  copyright on reality? By now it has long been clear that choosing the subject, the camera angle, the lighting and everything else involved in a good photograph, together form one's own artistic input. The photo is therefore protected by copyright. Here too: the camera does not receive copyrights. The patent on the technique of taking photos has long since expired, so the inventors of photography cannot claim anything either.

Of the three best-known text-to-image generators, DALL-E 2, Midjourney and Stability AI, only with the latter it is known exactly on which images it has been trained. These come from Open Source Data collector LAION, a German non-profit organization that wants to democratize the training of AI, so that Big Tech cannot gain too much power. LAION has indeed scoured the entire internet and created a database with text-image pairs. They did not store the images in their own databases, but only recorded the link to each image. Stability AI then normalized these images into usable training data by scaling everything to more or less the same size, 512x512 pixels by default. I won't go into the details of training here, but what is stored in a neural network is not the picture itself: it is the similarities and differences between the pictures. Just as humans "train" their brains by seeing many cats, both in real life and depicted in pictures, an AI also learns what a cat is by seeing many pictures of cats in combination with the word "cat". It learns to recognize all the characteristics of a cat, and can then reproduce it. Because only one Zendaya is really known on the internet, a neural network will store and reproduce all characteristics of only this Zendaya.

Should training with celebrities be banned? Or should one mainly adhere to existing rules when publishing (generated) pictures of celebrities?

A form of collage that does not stand out, but is often used in digitally created art, is 'matte painting'. If an artist wants to render rocks, forests, or crowds as realistically as possible (and not draw all the details by hand), they can take a piece of an existing photo with the texture they want, rotate it, magnify it, double it, and just as long edit it until it fits seamlessly into the artwork. It is not possible to determine from the end result where that original piece of photo came from. Those who watch movies or play games will see matte painting in action without realizing it.