Copyright Throughout a Creative AI Pipeline

Consider the following fact pattern: Alex paints some original works on canvas and posts photos of them online. Becca downloads those images uses to train an AI (training configures AI’s model parameters useful values). resulting trained parameter values her website under a license that reserves right use commercially. Cory in program is designed produce artwork. clicks create produces work. This work new Cory, but it looks lot like one Alex’s images. sells Advise about their potential copyright liability Alex (for substantially similar produced subsequently sold) taking Becca’s using commercially, contrary license). Cory again. The another work, this time quite different from any paintings. shares Instagram. Danny copies image Cory’s Instagram feed bunch postcards feature image. Cory. These scenarios are not as contrived they might initially seem. Copyrighted frequently required (more precisely: parameters). being shared licenses assume applies. People do these programs can novel content. be surprising end-user there generally no checks place ensure don’t take too directly training data. However, many will content already world. And end-users creative often claim ownership over work. I first present based neural network, popular forms basis state-of-the-art AIs. Then, I examine each issues just raised: 1. Does person managing automatic network’s obtain parameters? 2. artistic output output? 3. network requires large amounts example data (a set). Can around internet copied for purpose network? 4. What if examples? Is infringement? who infringing?