Sorry, that was expressed less precisely than I intended.
Human learning doesn't involve fixing anything, whether or not it would be a copy, into a fixed medium, and therefore even the question of whether or not it would be a copy if it were in a fixed medium, much less the question of whether or not such a copy would be exempted due to fair use, does not arise.
AI training does involve fixing something into a fixed medium, which makes it disanalogous to human learning. This raises the question of whether the thing fixes (parameter weights) is or is not legally a lossy mechanical copy of the training corpus, which further potentially raises the question of whether the incorporation of individual copyright protected works into the training corpus combined with its use in training would (before considering exceptions like Fair Use) violate the copyright on the works involved and whether, if it would, the use nevertheless satisfies the requirements of one or more of those exceptions (Fair Use being the one usually argued for.)
> And even if they did store something that's "effectively a copy", that's no more copyright infringement than when Google caches a site in order to handle search queries.
Among the reasons that Google removed outside access to its cached copies is that the evolution of the web meant that providing them increasingly potentially had negative revenue on the original content providers, which harms the Fair Use case for caching, since effect on the market of the original work is a fair use factor. (Google's cache was ruled as fair use based on the factual situation -- including implied consent, which AI model trainers have a much weaker arguent for -- in 2006, but changes in the facts relevant to fair use can change the outcome.)
But AI training is not so analogous to Google's cache (much less the situation of Google's cache in 2006) that one can simply leap with no analysis from one being fair use to the other in the first place. That's applying wishful thinking, not Fair Use analysis.