Some tasks lack enough real data to train on, or the real data is sensitive. Synthetic data is artificially generated to resemble the real thing, providing extra examples without exposing private records or waiting for rare cases to occur naturally.
It is useful for covering edge cases and balancing rare categories, but it carries a risk: if the generator's flaws are learned as if they were real, the model inherits them. Synthetic data works best as a supplement to genuine examples, not a full replacement.