I'd honestly be more worried about the NSA abusing OpenAI and not the inherent data collection which they already have access to anyway.
What if they target someone and give them a fake recipe for potato salad which turns out to be lethal.
Okay that's a dumb example, but do you get what I mean? They can arbitrarily send, insert, or delete data to a specific target if they want to. Would not be surprised if they already made their own in house tool for this like BeEF XSS to handle a bunch of targets at once.
I know the easy stupid solution to this is not to trust AI output in the first place, but people are lazy and dumb too.