Would this really work? Hardcoded models? When you want to influence a certain topic in a new way, on recent events - this means a new model needs to be created and adjusted all the time. I assume they rather use a complex traditional algorithm that can be tweaked - and if it is in use in US - it can be verified. If it is run through china, then things are clear as well. Do we have technical insights how it works?
You can simply use specific training examples that teach the model what you please. Eg. a set of examples which lead ranking/retreival/filtering models. The models are already online training and weights likely updated every ~1 hour ( or even less).
It’d be easy to go from a set of “moderators” who find examples and use it to query related content and use it as negative training samples. Just a guess.
