A significant portion of the training set for most image generation tools is stuff made in the last 10-20 years harvested from the internet, if not the last 5 y

A significant portion of the training set for most image generation tools is stuff made in the last 10-20 years harvested from the internet, if not the last 5 years. We're not talking about 150 years of copyright protection here, we're talking about the time frames you suggest. Artists want to protect their own work and their livelihood, and AI is being trained on the work they're actively putting out right now. You would have to shorten copyright duration to something like 5 years to come remotely close to making modern image generation models possible without violating artist copyrights. Text is different and much less difficult since its history as a medium is much longer - if you excluded the last 10-20 years of prose from your LLM it would probably still be very good at writing. But excluding the last 20 years of digital illustration and photography would be limiting yourself to a much lower-fidelity training set.