Datasets are not the only mechanism to train AI. You can also use reinforcement learning. This requires you to have a good fitness function. In some domains, that is not a problem. For LLMs, however, we do not have such a function. However, we can use a hybrid approach, where we train a model based on a data set and optimizing for fitness functions that address part of what we want (e.g. avoiding em dashes). In practice, this tends to be tricky, as ML tends to be a bit too good at optimizing for fitness functions, and will often do it in ways you don’t want. This is why if you want to develop a real AI product, you actually need AI engineers who know what they are doing; not prompt engineers who will try and find the magic incantation that makes someone else’s AI do what they want