The bias is built into the code. Otherwise, all would be almost the exact same.
AI is very good about anything not political. If it is political the answer will be biased.
So you've trained AI models with a couple hundred billion parameters?
It's been my experience that a model's output mimics its training data.
The Interwebs is crawling with Lefty comment. Train a model on Interweb data, and that's what you get.
Garbage-in, garbage-out still applies.