The web crawling is curation, as even the most mundane entries into the dataset will still have a significant weight on the model, even if it's small. What separates the brain from any statistics-driven model is how well it swims through an ocean of information to find the relevant details, tossing out what's left, and filling in gaps with intuition. Sure, its messy process, and mistakes get made, but that's in stark contrast to the very trim way that ML models run.