The "it" in AI Models Is the Dataset

If the goal is to recreate the training data set then all functional approximations are extensionally equivalent modulo biases introduced by the architecture. What I mean by architectural bias is how missing pieces of the data manifold are imputed, i.e. given some point x (w/o a matching output in the optimization corpus) different algorithms will give different results based on how x is encoded into the interal/latent representation of the data manifold. But even this difference is essentially averaged away by the users b/c the goal is to create something that will please the most number of users so it all eventually converges to the average agreed upon sentiment of a large enough sample of people.