Synthetic data and data augmentation

In short: When real data is scarce, expensive or private, you augment it or create artificial (synthetic) data, often with other AI models.

In more detail: Both enlarge the training material. Synthetic data has become an important part of training large language models.

In detail

Augmentation for images: flipping, rotating, cropping, brightness, noise, cut-out. The model learns that the class does not change in the process.
For text: paraphrasing, back-translation, replacing words.
For audio: change tempo and pitch, background noise.
Synthetic: a model writes tasks and solutions, dialogues, code. A checker (tests, verifier) filters out good ones.
Simulation: driving scenes, robots, medical images from simulators.
Risks: “model collapse” when models learn only from model outputs, reinforcement of errors, lack of diversity.
Privacy: real persons are harder to find in carefully generated synthetic data, but complete protection is not guaranteed.