More training and more data can have unintended consequences for machine-learning models such as GPT-4