Jack Dorsey, co-founder of Twitter (now X) and Square (now Block), sparked a weekend’s worth of debate around intellectual property, patents, and copyright, with a characteristically terse post declaring, “delete all IP law.”
X’s current owner Elon Musk quickly replied, “I agree.”
I’m still not getting it. What does generative AI have to do with attribution? Like, at all.
I can train a model on a billion pictures from open, free sources that were specifically donated for that purpose and it’ll be able to generate realistic pictures of those things with infinite variation. Every time it generates an image it’s just using logic and RNG to come up with options.
Do we attribute the images to the RNG god or something? It doesn’t make sense that attribution come into play here.
I would like to take a crack at this. There is this recent trend going around with ghiblifying one’s picture. Its basically converting a picture into ghibli image. If you had trained it on free sources, this is not possible.
Internally an LLM works by having networks which activate based on certain signals. When you ask it a certain question. It creates a network of similar looking words and then gives it back to you. When u convert an image, you are doing something similar. You cannot form these networks and the threshold at which they activate without seeing copyrighted images from studio ghibli. There is no way in hell or heaven for that to happen.
OpenAI trained their models on pirated things just like meta did. So when an AI produces an image in style of something, it should attribute the person from which it actually took it. Thats not whats happening. Instead it just makes more money for the thief.
I think your understanding of generative AI is incorrect. It’s not just “logic and RNG” It is using training data (read as both copyrighted and uncopyrighted material) to come up with a model of “correctness” or “expectedness”. If you then give it a pattern, (read as question or prompt) it checks its “expectedness” model for whatever should come next. If you ask it “how many cups in a pint” it will check the most common thing it has seen after that exact string of words it in its training data: 2. If you ask for a picture of something “in the style of van gogh”, it will spit out something with thick paint and swirls, as those are the characteristics of the pictures in its training data that have been tagged with “Van Gogh”. These responses are not brand new, they are merely a representation of the training data that would most work as a response to your request. In this case, if any of the training data is copyrighted, then attribution must be given, or at the very least permission to use this data must be given by the current copyright holder.