Maybe you could train it on a central database that keeps track of which conversations get the most consecutive replies from scammers.
You might not even need GPT. Maybe just the act of responding with a reasonable amount of text would create a large filtering burden for scammers.
I have to believe someone has tried this.