Synthetic data is training material made by a computer instead of gathered from people or sensors. An AI model might write practice questions and answers, or a simulator might produce millions of robot camera images that never happened in real life.
It helps when real data is scarce, private or expensive to collect, and Microsoft’s Phi-4 report says that model “strategically incorporates synthetic data throughout the training process”. The risk is that flaws or hidden lessons in generated data are passed on to the model that learns from it.