Level 1 Techs goes over the history of copyright and the coming war with ai scrapers

with image copyright, the US courts have already decided: copyrights requires human conscious intent for protection. images created by animals(pets painting, monkey stealing a tourist camera, etc) is not protected. images generated by automation/software from random input is also not protected.

human intent/execution is required to be protected. so someone has to have the concept and has to do some action to create the product. the degree and amount of action is the part that remains the grey area in the generative space.

people submitting word prompts has been likened to a client telling the painter what they want from a commission. regardless of the idea being the commissioner's, they art copyrights remain with the person who put down the brushstrokes.


[in regards to models "remembering" art]
there is no protection for current image diffusion models because they used unauthorized images in the training. those billions of images in ~50+ stages of gradual gaussian blurring still remain inside the model so it is still cut and paste usage of the original copyrighted work. some image gen will still include distorted versions of the original artist's signature. so everything they generate is all fruit of the poison tree.

if the model trainers want to start from scratch using images they have licensed/paid the artists for, then fine knock yourself out. but that will reduce the range, variety, quality of the output images; which will reduce the value of the software/service they can offer to the end user and therefore reduce profit. (you could try to charge everytime a user used a specific studio/artist's name as a prompt, and then pay the artist after the license agreement rather than a large speculative upfront payment)
a clean non-scraped model will have access to public domain art(anything prior to 20th century) and a decent amount of real world photography, so those types of prompt will still work fine. but anything using a particular trendy artist's name will be irrelevant garbage output. there are artists licensing their catalogs for training and adobe forcing their user to let them scrape, which should improve the quality of a clean model but that model will be far less attractive/profitable as it wont be able to imitate the artists that user likes. probably useful for corporations and advertising types that just need a background image to slap together for websites and flyers.

i posted in the ai thread, poisoning tools for images and music already exist. whenever a model trains with scraped images the output is corrupted with discordant results to the prompt words. in 5 to 10 years when every artist is poisoning their work, the generative models will be worth far less as the range of what they can create will have a ton of holes.