What if both A and B were implemented using fully automated systems that relied on next token probablity in language?