AI Glossary / Synthetic data

🏗️ Models and training

Synthetic data is training material generated by AI itself: generated dialogues, problems with solutions, image descriptions. It's used when real data is scarce, expensive to obtain, or legally and privacy-wise problematic. Well-designed synthetic data genuinely improves models - especially in math and code, where correctness can be verified automatically. Risk appears with mindless "feeding AI with AI output," which can perpetuate errors and impoverish diversity; in extreme form, this is called model collapse.