The central difference is scale.

If a person "trains" on other creatives' works, they can produce output at the rate of one person. This presents a natural ceiling for the potential impact on those creatives' works, both regarding the amount of competing works, and the number of creatives whose works are impacted (since one person can't "train" on the output of all creatives).

That's not the case with AI models. They can be infinitely replicated AND train on the output of all creatives. A comparable situation isn't one human learning from another human, it's millions of humans learning from every human. Only those humans don't even have to get paid, all their payment is funneled upwards.

It's not one artist vs. another artist, it's one artist against an army of infinitely replicable artists.