That’s going to always happen when training data with the entire internet. The outliers will always skew thing more than the mainstream if the models are not designed to exclude them.

And there are a lot of contributing factors, for example with right leaning stuff being more available for LLMs to process as the platforms are generally less concerned about privacy and more concerned about policing and control (that’s just what right wing is), of course the models are going to see more of it than the left leaning stuff where people are more on the repressed side, more likely to use more private communication methods, and less likely to be able to safely, publicly share outlier kinds of views to skew things the other way. Even the people who pretend to be extreme left-wing, like the USSR or the Chinese Communist Party, are usually, in reality, right-wing.