In other words, somewhat similar to what a programmer studying existing and open-source code would do: read the code, understand it, and try to reimplement by writing their own code.
I remember people claiming things like github copilot were reproducing original code verbatim and this was the main source of concern. I would see this as a problem if my colleagues were copying code as if it was their own.
I'm just pointing out (repetitiously, because arguments of the type "but humans do X" are made all the time with LLMs) that the fact that a human brain doesn't violate copyright when a human reads and studies copyrighted code isn't relevant either way.