In terms of creating a less bias AI model, it is better to have a more diverse training model. Thus, if many data subjects opt-out to this training attempt, in the long run there will be a worse off propaganda machine.
This is one of those current dilemmas between quality of model and acquiring of data.