Synthetic data went from "interesting research idea" to "production tool" between 2023 and 2026. Most frontier-lab training pipelines now include large synthetic components, and enterprise teams use it routinely for training, evals, privacy-safe analytics, and rare-class augmentation. But synthetic data also has a sharp failure mode — model collapse — and a surprising number of teams discover it the hard way. This guide is the playbook for using it well.
What changed in 2026
- Frontier-lab pipelines depend on it. Disclosed methodology from leading labs in 2024–2026 confirms substantial synthetic-data components in pretraining and post-training.
- Tabular synthetic data is regulated. EU AI Act and US state laws now treat differential-privacy guarantees as legally meaningful claims; vendors compete on certification.
- The tooling matured. Gretel, MOSTLY AI, Tonic, and open-source frameworks (SDV, NVIDIA NeMo Curator) are stable enough to ship with.
When synthetic data is the right tool
Filling rare-class gaps. Your real data has 10,000 examples of class A and 50 of class B. Generating synthetic examples for class B (with careful diversity) materially helps the model.
Edge-case coverage. Adversarial inputs, error states, rare combinations of features — much faster to generate than to wait for them to occur.
Privacy-safe analytics and dev environments. Synthetic copies of production data with no PII. Engineers debug with realistic distributions; auditors are happier.
Eval set augmentation. Generate variants of test cases to stress-test specific behaviors — paraphrases, translations, structural edits.
Multilingual / locale coverage. Translating and adapting English data to other locales is a tractable synthetic problem.
When it quietly fails
Bulk volume with no curation. Generating 10× more data without quality filtering rarely improves models — the synthetic distribution starts to dominate and pretrained capabilities degrade.
Self-training loops. Training a model on its own outputs without filtering causes mode collapse — diversity narrows, errors compound.
Distribution shift. Synthetic data inherits the generator's biases. A medical synthetic dataset generated by a US-centric model will under-represent the conditions and demographics of other regions.
False privacy guarantees. Synthetic ≠ private. Without differential privacy or careful design, membership inference attacks recover real records.
What "good synthetic data" actually looks like
| Property |
What to do |
| Diversity |
Multi-prompt, multi-seed, multi-model generation |
| Quality filtering |
Reject samples below a quality threshold (judge model or classifier) |
| Real-data anchoring |
Mix synthetic with real at a known ratio (e.g., 1:3 to 1:1) |
| Drift monitoring |
Eval on real validation set after every synthetic injection |
| Provenance tracking |
Log generator, prompt, seed for every sample |
The vendors that matter
Gretel. Strong on tabular synthetic data with privacy guarantees. Used in finance and healthcare for analytics and dev environments.
MOSTLY AI. Tabular and time-series. Differential privacy story is the most rigorous in the category.
Tonic. Database-shaped synthetic data for dev/test environments. Less ML-focused; more "make me a synthetic Postgres".
NVIDIA NeMo Curator. Open-source pipeline for LLM training data — both real and synthetic — with quality filtering and deduplication.
Concrete recipe for LLM training augmentation
- Identify gaps via eval failure analysis. Where does the model lose? Rare intents? Specific languages?
- Generate targeted synthetic examples for those gaps using a strong model and varied prompts.
- Filter aggressively — judge model rejects ≥ 30% in practice.
- Mix with real data (1:1 to 1:3 synthetic:real depending on baseline volume).
- Train, evaluate on held-out real data, repeat.
This loop reliably improves narrow weaknesses without degrading overall quality.
FAQ
Is synthetic data ever as good as real?
For diversity and edge cases, no — real data wins. For specific gaps and privacy substitution, synthetic is genuinely useful.
Will model collapse hit me?
Only if you train on your own output without filtering or real-data anchoring. With the recipe above, it's avoidable.
What about privacy?
Synthetic is not automatically private. Use differential-privacy methods if the original data is sensitive.
How do I measure synthetic data quality?
Train two models — one with, one without — and compare on real held-out evals. The synthetic data either helps or it doesn't.
Where to go next
For related material see AI evals frameworks in 2026, RAG vs fine-tuning in 2026, and AI privacy guide: protect your data.