Real data has real limits
Training a capable AI model usually takes a large volume of labeled examples, and collecting that data from the real world is often slow, expensive, or both. A team building a computer vision system might need thousands of labeled images of a specific object in varied lighting and angles, each one requiring a human to photograph and annotate it. That cost scales directly with how much data you need, which becomes a real bottleneck as models demand more.
Beyond cost, real data has gaps that are hard to fill no matter how much you collect. Rare but important events, like a pedestrian stepping into traffic unpredictably or a piece of industrial equipment failing in an unusual way, are by definition uncommon, so a dataset built purely from recorded reality will underrepresent exactly the cases that matter most for safety and robustness.