Data is necessary for the linear relationship between increased model performance as a function of compute and data (Kaplan). Despite this, the availability of large-scale human feedback datasets remains scarce — AI firms scrape and host proprietary data and provide little to no transparency on its origin. Without vast, high-quality data, even models with substantial compute and pre-training will be insufficient for building the next expansion of AI models. Many empirical challenges arise from using a fixed dataset for AI model training, too. Bias, hallucinations, and sometimes critical security flaws arise as a result from the size of modern pre-training datasets being larger than any comprehensible measure that a human could reasonably parse. Generative AI models also pose reflexive training problems: without a clear understanding of data provenance, it is impossible to tell whether data has interacted with AI systems or comes from human sources.