Glossary
Synthetic data
Quick answer
Data generated by models or simulations rather than collected from real activity.
Useful for filling gaps or augmenting datasets, but often derived from, or evaluated against, real data.
Why Synthetic data matters when licensing company data to AI
Synthetic data is generated by models or simulations rather than collected from real activity. Buyers often ask whether a dataset includes it, because generated records can fill gaps but are usually derived from or tested against real data. The agreement and documentation should state which records are real, which are generated and how each was produced.
- 01Your work records
- 02Personal details removed
- 03You approve the release
- 04Licensed to AI developers
Related terms
How is synthetic data generated?#
Synthetic data is generated by fitting a model to real data or to rules and then sampling new records from it, or by rendering scenes in a simulator. The method depends on the data type: tables, text and images each have their own common approaches.
The Synthetic Data Vault library, for example, offers a statistical synthesizer (GaussianCopula) that learns each column's distribution and the dependencies between columns, and a deep learning synthesizer (CTGAN) that trains a generator against a discriminator that tries to tell real rows from synthetic ones. For text, generative AI models can create training examples themselves: the Self-Instruct method has a language model generate instructions, inputs and outputs, filters out invalid or near-duplicate ones, and fine-tunes the original model on the result.
| Method | How it works | Typical data | Example |
|---|---|---|---|
| Rules and random sampling | Values are drawn from hand-written rules, formats and ranges | Test records, demo databases | Fake customer rows for software testing |
| Statistical models | Learn column distributions and correlations from real data, then sample | Tabular business data | GaussianCopula in the Synthetic Data Vault |
| Deep generative models | A neural network learns the data and generates new rows or media | Tables, images, audio | CTGAN in the Synthetic Data Vault |
| Large language models | A model is prompted to write examples, which are then filtered | Instructions, conversations, labeled text for NLP | Self-Instruct instruction data |
| Simulation and 3D rendering | A physics or graphics engine renders scenes with known labels | Camera and sensor data for perception models | NVIDIA Omniverse Replicator |
Synthetic data vs simulated data: what is the difference?#
Simulated data is one kind of synthetic data: it comes from running a model of the physical world, while other synthetic data comes from a model that learned patterns from existing records. A simulator produces data by computing what would happen in a scene it was given; a generative model produces data by imitating the statistics of data it was trained on.
NVIDIA describes Omniverse Replicator as a framework for producing physically accurate 3D synthetic data, and its documentation uses domain randomization, varying elements such as lighting, materials and camera positions, to narrow the gap between synthetic and real images. Its documentation also cautions that it is hard to know whether the chosen randomizations represent the real domain, so testing on real data is essential.
Is synthetic data reliable?#
Synthetic data is reliable for the patterns its generator captured and unreliable for anything the generator missed. The main disadvantages are lost rare cases, errors inherited from the source data, and privacy that is not guaranteed by default.
Research published in Nature found that training models indiscriminately on model-generated content causes irreversible defects in the resulting models, with the tails of the original distribution, the rare cases, disappearing first. NIST's guidance on de-identification (SP 800-188) treats synthetic datasets built from privacy-preserving models as one de-identification technique, which means the privacy result depends on how the generator was built, not on the label synthetic. Counsel and privacy reviewers should assess the specific case.
- Compare distributions and correlations of the synthetic set with a held-out real sample.
- Check that rare categories and edge cases still appear at realistic rates.
- Evaluate the final model on real data, not only on synthetic data.
- Record which model, prompts or simulator settings produced each batch.
- Test whether any synthetic records reproduce real individuals or near-copies of source rows.
- Keep a share of real records in training rather than replacing them entirely.
Sources for this section
- The Synthetic Data Vault offers a statistical GaussianCopula synthesizer and a GAN-based CTGAN synthesizer for tabular data. Source
- Self-Instruct uses a language model's own generated instructions, filtered for quality, to fine-tune that model. Source
- Omniverse Replicator generates physically accurate 3D synthetic data and uses domain randomization, with real-data testing still essential. Source
- Training indiscriminately on model-generated data causes model collapse, in which rare cases in the original distribution disappear. Source
- NIST SP 800-188 counts synthetic datasets built from privacy-preserving models among de-identification techniques. Source
Related resources
- SolutionWhat is AI evaluation data?
- SolutionHow AI developers source data
- IndustryBPO & contact centers data
- IndustryRecruiting & staffing data
- DataInternal knowledge bases
- QuestionShould companies sell or license their data?
- QuestionCan CRM data be licensed?
- InsightCan you license spreadsheets and financial models to AI companies?
- InsightCan you license training materials and course content to AI companies?
- InsightCan I limit licensing to evaluation only?
- DataWorkflow data
- GlossaryDataset
See if your company qualifies
A short company assessment. No data uploads are needed.