# Synthetic Data Training data generated by AI systems rather than collected from human activity — AI writing text to train other AI, AI generating images to train image models, AI creating code examples to train coding assistants. The practice emerged from necessity: [[Scaling Laws]] demand exponentially more training data, but high-quality human-generated data is finite, increasingly copyrighted, and expensive to license. ## The circular economy of data Synthetic data creates a circular relationship: AI generates data → data trains AI → AI generates more data. The optimistic framing: synthetic data solves the data-scarcity bottleneck, enables privacy-preserving training (generate medical data without real patient records), and allows targeted improvement of model weaknesses. The pessimistic framing: **model collapse**. Research (Shumailov et al., 2023) has shown that models trained on AI-generated content can degrade over generations — each generation's errors compound, and the model's output distribution narrows, losing the diversity and nuance of human-generated data. Training AI on AI output is potentially like making copies of copies — each generation loses fidelity. The practical reality as of 2026 is somewhere between: carefully curated synthetic data, mixed with human-generated data, appears to work well for specific training objectives. Uncurated synthetic data at scale carries model-collapse risks that are not yet fully understood. ## See Also - [[Scaling Laws]] — why synthetic data is needed - [[Data Moats]] — what synthetic data potentially erodes - [[Foundation Model]] — what synthetic data trains