Synthetic data generation is important in machine learning because real-world datasets can be difficult to access, share, or reuse due to privacy and legal restrictions. This thesis performs a comparative study to investigate machine learning performance and related evaluation attributes when models are trained using synthetic tabular data. The study compares a simple class-conditional rule-based generator with a generative model based on Conditional Tabular Generative Adversarial Network (CTGAN). The problem addressed is that advanced generative methods are often expected to produce more useful synthetic data, but they may also generate records that are closer to the original training data. To investigate this, a controlled experiment was conducted using the Adult Income dataset. Synthetic datasets were generated using both methods and evaluated through machine learning utility, statistical similarity, and privacy-related similarity. The results show that the rule-based generator achieved stronger overall machine learning utility and higher overall statistical similarity in this experiment, while CTGAN produced more near-duplicates and lower nearest-neighbor distance. These findings suggest that, in this setup, the simpler rule-based method provided a better balance between usefulness and privacy-related similaritythan the selected CTGAN configuration.