Probably not as long as the internet has existed, but they've definitely had access to chatgpt-like AI for a little while. So has google of course. The problem with these learning models isn't the model itself, it's creating the sanitized, politically correct, idiots-can-use it version for the public. Private versions undoubtedly have been in use for a while.

I'm unsure how much this has been used as propaganda. It might have been harder than you think to properly train a model to generate propaganda, as you can see with the jailbreaks of chatgp3