Synthetic data are artificially generated examples, typically through algorithms or AI models, instead of being collected from the real world. They can mimic the statistical properties of an authentic dataset or be created from scratch to cover specific situations, and are used to train, validate, or augment machine learning models.
Their importance has grown for several practical reasons:
- Privacy: they allow working without exposing real personal data, useful in healthcare or finance.
- Scarcity: they cover rare or hard-to-obtain cases, such as infrequent failures or dangerous situations.
- Cost: they are cheaper than collecting and labeling large volumes of real information.
A common example is the use of simulators to train autonomous cars with millions of virtual kilometers. It is worth keeping a nuance in mind: if synthetic data do not reflect reality well, the model will learn misleading patterns. That is why they are usually combined with real data and carefully validated to avoid biases or performance that only works "on paper."