The issue is that LLMs generally give fairly "black and white" answers. They will generate answers, even "hallucinating" to do so. This presents things essentially as facts, when the reality is much more complex. They are programmed to offer some uncertainty and nuance (which doesn't mean "left wing"), but as shown in the various studies, there is a strong tendency towards "left wing" answers. Again, that doesn't mean nuanced answers, it means a tendency to generate responses that reflect particular political ideologies. Some of this is due to human intervention in fine-tuning. In the
summary PDF from HAI at Stanford:
The main issue seems to be the biases of people involved in fine tuning results. People are not generally good at understanding their biases. It turns out as covered in
this preprint, the models are also generally not good at knowing (not the correct word to use, but we'll go with it) their biases either (they aren't self-aware and how much "meta-cognition" is happening isn't clear). This means they generate usually biased output without understanding it is biased. That's fairly human-like, but what's generated tends to represent certain groups of people better than other groups.
The question about whether there should be a bias is a completely different one I won't get into.