This is patently false. GPT-4 will have more information pertaining to North America than to Africa because there is more training data available for North Amer

This is patently false. GPT-4 will have more information pertaining to North America than to Africa because there is more training data available for North America than Africa. That's not a reflection of reality; that's a reflection of the distribution of computers and wealth. > If by default an LLM refers to engineers as "he" more than "she", it's because engineers are more likely to be male than female. That's a great point to raise, thank you for mentioning it. This really is the crux - machine models crystallize _current_ bias and amplify it. Most programmers in the early days of computers were women, but were displaced over time. Should the training data reflect the bias of the dominance of women in the early days, the dominance of men now, or some other combination? > Most people would prefer an AI whose model of reality matches actual reality Training data is not reality; training data is training data. It will fundamentally lag behind reality, and it will have a bias towards past conditions and precedents.