Replicating copyrighted code from the training set only happens 1% of the time, it's the exception not the rule. And when it happens it's usually because the same text appears multiple times in the training set. So it will memorize boilerplate and popular code snippets, not unique stuff. Even a replicated piece of code 100 lines long is no big deal in my opinion, unless it contains some kind of unique thing never seen before, like an optimized matrix multiplication function. Certainly not FizzBuzz.

On the practical side, it is actually easy to filter out sequences of words that are too similar to the training set from the output of the model. You just generate another snippet until it is "original" enough.