Well, what about Gebru's findings? Data in and of itself is not beneficial, we do have certain underlying beliefs about appropriateness and goodness. There is an element of innate belief at play.

Otherwise, chatbots are very good at learning abusive and racist language. It's due to our belief that it's not right and decent that we train them using a bias. And it's only due to beliefs that we rein in the rogue AI which misclassifies black people.

Edit: If you still want to think of belief as an input node whose link has weight, then set the weight to infinite