If your training set has 90% nazi opinions, don’t expect to see pro-Jewish arguments.
Deliberate being the key word here.
If I suck at sampling and make a bias model, I suck at sampling. If I shape a model to neuter it and ensure that it only puts out morally approved outputs, that slants significantly heavier towards censorship
"There are no un-representative models." and "Models by design represent opinions in the ratios they have seen."
That rephrasing matters when thinking about, for instance, what Pixar films should or should not be shown to fifth graders and why.
That which is regulated to be unrepresentative is likely suppressive.
If your training set has 90% nazi opinions, don’t expect to see pro-Jewish arguments.