If models are trained on internet content (Ex: ChatGPT), then surely there is some bias in their output, as the internet is infested with content who's sole purpose is to make you buy something, influence you in some way or straight up plug a product.