Can Synthetic Data Replace Real Data? Evaluating the Efficacy, Limits, and Label Trustworthiness of Generative Data in Supervised Learning

Synthetic data are increasingly proposed as substitutes for real training records where privacy restricts data sharing. We evaluate this substitution on a medical appointment no-show dataset of 110,526 records with the Train-Synthetic-Test-Real protocol, five classifiers, six synthetic mixing ratios and nine generation runs of CTGAN, TVAE and a Gaussian copula on this and a second dataset (UCI Adult). Full substitution lowered mean AUC-ROC in every run (by 5.72% to 11.46% across classifiers in the primary run), and mixed training beat training on its real part alone in at most six of 40 paired DeLong comparisons per run (Holm correction) while losing in 6 to 30. We propose Synthetic Label Trustworthiness Estimation (SLTE), an empirical label-consistency check that scores each synthetic label against a model trained on real data, relabels low-confidence samples by nearest-neighbour vote and down-weights uncertain ones. SLTE changed 12.88% of the primary synthetic labels, 94.2% of them from minority to majority class, but a label-permutation null reproduced this asymmetry and the same procedure changed 14.20% of held-out real labels, so the changes reflect disagreement with a real-data model rather than generator errors. In the primary run SLTE improved AUC-ROC significantly only at 80% and 100% synthetic data, recovering 57.8% of Gradient Boosting's loss at full substitution, but it lowered recall for most classifiers, and in other runs it lowered AUC-ROC for Decision Tree, Logistic Regression or Naive Bayes. Synthetic data did not replace real data here; SLTE is a quality-control step for when substitution cannot be avoided.