That would likely depend on the meaning of "based on" (not a lawyer here!), and the actual lawyers would feel very happy charging exorbitant sums for arguing it out in court. As any sensible corporation (that is not deriving its income from lawyer's fees) would like to avoid that so the easiest approach would be "don't look at GPL code".

That raises an interesting question though. If I read the code, remember it, and then re-type it as is, that'd obviously fall under license's terms. But if I write an ML process, that knows to read code, understand it, and reproduce it, and it ends up reproducing the said code byte-for-byte, is that considered copying too? Suppose I didn't intend for ML process to work that way, I just programmed it in a way to find optimal solutions, and trained it on a dataset of GPL code, and it happened that this particular piece of GPL code, byte-for-byte, was chosen by the ML algorithm as optimal. Did I copy the code? Did I violate the license even earlier, when training the ML model on GPL code?