The bias is really towards recency, because LLMs are mostly trained on text scraped from the internet.

If you incorporated older content, the bias would be much less liberal. It'd be interesting to train one on only pre-1950 content.