Synthetic data shall be as close as possible, but not too close to actual data. Accuracy and privacy, can both be understood as concepts of (dis)similarity, with the key difference being that the former is measured at an aggregate level and the latter at an individual level.