To adress the mems side of the question: Memes aren’t a large portion of the original work. Often times they’re screenshots of video material, so the “portion taken from the original” is minute. Some meme formats, however, are digital art pieces in and of themselves. (Note the word format - the “background” of the meme, for example the “If I did one pushup” comic)

But even with that consideration, a meme doesn’t bring harm to the original - it’s basically free advertising. And as the memes are usually low quality abd not monetized, it can be passed off as fair use or free speech in some jurisdictions, while others merely turn a blind eye. And why shouldn’t they?

As I said, memes have a multitude of points going against them being copyright infringement. They’re low-effort, short-form media, usually with a short “lifetime” (most memes don’t get reposted for years). Most often they’re a screengrab of a video (so a ‘negligible portion of the original’) and almost never bring harm to the original, but only serve as free advertising. Again, usually. This means most meme formats’ involuntary creators have no reason to go after memes. You could probably get a court to strike a meme, but probably on defamation grounds - and even then, the meme will most likely die (not the format!) beforehand, so such suits are usually dismissed as moot.

Compare this to an AI model (not an AI “artpiece”): It’s usually trained on the entire work, and they’re proven to be able to recreate the work in large part - you just need to be lucky enough with the seeds and prompts. This means the original is “in there somewhere”, and parts of it can be yanked out. Remeber, even non-identical copying (so takig too much inspiration or in academic speak, “plagiarism”) is copyright infringement.

And to top it all off, all the big AI models have a paid tier, meaning they profit off the work.

If you were to compare memes to individual AI “artworks”, then it is the same thing as memes. Except if the generation is a near-verbatim reproduction, but even then, the guilt lies with the one who knowingly commited infringement by choosing what to put into the model’s training data, and not on some unlucky soul who happened to step on a landmine and generated the work.

“You”, the user of the AI model isn’t engaging in copyright infingement directly.

However, whoever made the model that you used did. Most using copyright protected works.

Some people are paying for these models. This is what’s the problem: financially benefitting off others’ work without permission (or royalties).

It’s like the age-old piracy dilemma: the person using direct downloads or streaming can’t be fined in most jurisdictions - it’s the duplication and sharing that’s forbidden.

This exact analogue exists with AI models: training a model and giving it to others to use is distributing access to copyrighted material. Using an AI model is not.