The tech is already there to do so: https://research.google/blog/mechanism-design-for-large-lang...

It was always there. Just don't train the model with data that you don't like.

Then you can’t make it targeted.