> Here it is, since you were not aware: "the amount and substantiality of the portion taken".
Yes. If you use only 1% of a work, then you are not using a substantial or large amount of the work and it is considered fair use.
But training doesn't use 1% of the work, it uses the entire work. No one is using 1/100th of an individual image to train, nor are they using 1/100th of a codebase to train, etc.
They're using entire individual works, and all those factors that are applicable are evaluated collectively, not in isolation.
Besides, all those factors become irrelevant if "...On the other hand, it is as clear, that if he thus cites the most important parts of the work, with a view, not to criticize, but to supersede the use of the original work, and substitute the review for it, such a use will be deemed in law a piracy ." (https://en.wikipedia.org/wiki/Fair_use)
It's hard to claim that the owners of ChatGPT and similar are not trying to supercede the works it is fed as input. They state as much everywhere.
> Instead, what I am saying, is that if there is a model, trained on millions and millions of images, the output of the model is fair use, because it is not taking significantly from your individual work.
Whether the output from the model is fair use or not is irrelevant to whether the input falls under fair use.
I must say your take is certainly novel, and no, I haven't seen anyone try to make that claim before; each time I have asked I have gotten a different answer.
I think a better case to be made is that ChatGPT is transformative, which would make it fair use.
If you read through the entire wikipedia article I linked above you'll see that:
1. All the factors are evaluated collectively, in relation to each other, not individually.
2. The burden of proof lies with the defendant, not the claimant. IOW, the court starts off with "prove that the use is fair", and not "prove that the use is not fair". From wikipedia "This means that in litigation on copyright infringement, the defendant bears the burden of raising and proving that the use was fair and not an infringement. "
In short, when the license says "not to be used as training data or learning data for any machine model", and it is ignored, the defendant is already in violation. If sent a cease and desist with a request for royalties, the defendant is already presumed to be in violation", and will have to prove fair use, which will (in order of factors) mean that they have to answer "No" to all of the following questions in court:
1. Is the output product being used for commercial purposes and/or profit?
2. Is the input work a freely available fact, or is of a nature that it's in the public interest to reproduce.
3. Is the proportion that is used of the input work insignificant (typically less than 1/100th) of the input work?
4. Does the output work harm the market for the input work?
The owners of ChatGPT are unable to answer YES to any of the above.