Maybe, but I think there is also a reasonably straightforward interpretation of self-modifications as being mere processing output and not a derivative work (since computers cannot currently hold copyright anywhere that I know of). Now, if a person ran the self-modifying code (for example, by emulating a Turing machine with a pen and paper) and distributed the output (for example, over a network, as a result of running the program), the person might have the obligations of the GPL (or be in breach of them, if he or she distributed a derivative work in violation of them as a result of interpretation), if they licensed the code under the GPL, in the same way that processing/modifying and then distributing any GPL code would, whether the tools used were GPL or not. If that person were the original copyright holder, however, they could still do whatever they want (for example, releasing the modifications without the GPL) since they have the full rights, not merely the GPL-granted ones.

The big question this raises is: on a network of devices, what constitutes the operator's computer and what is the threshold past which execution would constitute distribution. It goes without saying that IANAL.