FWIW, I was using em-dash before it was actively the opposite of cool.
It's not just that, this construction:
> This isn't X, this is Y
is a huge ChatGPT signal.
But surely it is only a ChatGPT signal because it was a strong signal in the training data. You need more than one strong signal with that sort of potential for false positives to make a reasonably accurate identification.
If it is, it's one I've neither heard others mention before nor seen often enough myself to consider it a tell (but for the latter, I do use ChatGPT's customisation options).
I love how completely prosaic phrases are now "ChatGPT signals".