A supplement, not a replacement
In practice, synthetic data is almost always combined with real data rather than used on its own. A typical approach blends a core of real-world examples with a targeted layer of synthetic data added to cover specific gaps, rare scenarios, edge cases, or underrepresented categories, that the real data alone doesn't handle well. The real data anchors the model to genuine patterns in the world, while the synthetic data extends coverage into areas that would otherwise be thin or missing entirely.
How much synthetic data is appropriate varies a great deal by task and domain, and there's no universal ratio that works everywhere. What matters more than any specific proportion is treating synthetic data as a deliberate addition aimed at a known gap, rather than a generic substitute used simply because it's cheaper or faster to produce than real data.