“Large language models are prone to projecting stereotypes, thereby exaggerating existing fairness and bias risks, because they are trained on vast amounts of unregulated internet data.”