People don't understand what a generative image model is nor how it works, and so there's an entire (understandably) really upset movement whose entire argument is a strawman.
Just like any young, trained human artist, LLMs CAN copy existing works. However, like most young, trained artists -they DON'T most of the time.
An LLM or generative image machine learning system works like (emulates) a biological neural network at a very low level.
Sure, one can get R2D2 in one's sci fi image in response to a general sci fi prompt, but usually one doesn't. The model is functioning like what it is: an emulation of a human artist's cognitive-creative processes. If you TELL it not to plagiarise, it wont. Just like with a human artist.
LLMs are not JUST a statistical inference and pattern matching 'machine'. That description leaves out semantic vectors and low-level emulation of Hebbian neural (re)-wiring. It's the latter two features which are critical for classification ad understanding.
For example:
This just betrays an ignorance about what LLMs are and how they work. When you train a young artist - painter, filmmaker, director - to do their art, you train them on the masters and their genres and styles. Then you let the young artist loose on the world. You don't accuse them of theft unless they blatantly copy an existing work. Intertext is okay. Bricolage and pastiche are okay. Homage is okay.
Young artists become mature artists by emulating into their own inspiration. LLMs and generative image models do the same thing: they're a low-level emulation of Hebbian neural networks.
