With the current capabilities of LLMs I'd imagine removing the "it's not A, it is B" tell would be rather simple and involve a modification of the system prompt.

The fact that none of the frontier labs have done so is probably because it's another way to easily fingerprint their own slop. It's known that Anthropic already has a mechanism for ID'ing Claude output. As more and more of internet communication is mediated through agents, there is value in finding a way to prevent slop from being re-ingested during training and avoiding model collapse.