The basic element of copyright is some level of creative input. Mass, automated transformative processes without human input traditionally don't count as creative (e.g. compiling source code with a compiler usually doesn't mean that compiler authors get copyright over the binary as a derived work).
AI is recently being presented as kind of a gray area where if you remove the inputs far enough from the outputs with huge, incomprehensible machine learning models the outputs count as original works, even when provenance is sometimes clear in the end (e.g. see the whole debate about GitHub Copilot regurgitating well-known pieces of code). I suspect that's mostly the big tech companies lobbying and pushing for this re-interpretation of copyright.