Synthetic Data is artificially generated information that mirrors the statistical properties of real-world datasets without exposing sensitive personal records.
Synthetic data is generated rather than collected: LLMs author training examples, simulators render labeled scenes, and statistical generators mimic sensitive datasets with privacy guarantees. Pipelines pair generation with rigorous filtering and verification, model-judged quality checks, deduplication, and contamination controls, because generation quality bounds training value.
Synthetic data attacks the field's real bottleneck, labeled, shareable data: it fills rare-case gaps, substitutes for privacy-restricted records, and powers post-training for models large and small. By 2026 most serious model training mixes synthetic and human data deliberately; the craft is in the mixing.
Unfiltered recursive self-training degrades models, but curated synthetic data, verified, diverse, mixed with real data, demonstrably improves them. The risk is quality discipline, not synthesis itself.
Only with explicit guarantees: naive generators can memorize and leak source records. Differential-privacy methods and membership-inference testing turn 'looks anonymous' into measurable protection.
Edge cases too rare to collect, formats needing perfect labels (simulation), privacy-locked domains, and instruction-tuning corpora. It complements real data; full replacement remains rare in production.