It is a natural consequence of these systems that they output the same biases that go into the input; e.g., the classic translation of "they are a doctor/nurse" in a language with gender-neutral pro/nouns to "he/she is a doctor/nurse" in English. Therefore, if one were to train any such system on the Founding Fathers or Nazi soldiers, it is only natural to expect that the system would output white/German/Eastern European men in each case. Yet Gemini doesn't; instead, Gemini introduces its own bias to erase the one in the input and instead always produce an "equal" cocktail of white/black/asian men/women (which, ironically, is still biased and racist because that only captures the races that woke employees at Google feel are important, but it takes actual intellectual honesty to see one's own biases), regardless of whether that results in historical inaccuracies or downright falsehoods. This is a very intentional bias and dishonesty, not something that can just be shrugged off as "it's hard and takes millions of dollars to train".