You can't "give up bias". You can choose what you prefer, but an "unbiased" model doesn't exist. If you put in information, you're adding bias.

So biased by both the choice of the sources and the information contained in the sources.

But crucially not biased by RLHF being applied afterwards to stop the models from making basic biological observations, telling jokes about left wing politics, or saying anything nonnegative about European historical impact on the world, and so on.