Getting rid of bias in LLM training is a major research problem and anecdotally e.g., to my surprise, Gemini infers gender of the user depending on the prompt/what the question is about; by extension it’ll have many other assumptions about race, nationality, political views, etc.